chore(release): 0.6.0

Android native video renders a picture, and its transport works.

The path shipped once as audio with no picture and was reverted with the
compositing named as the suspect. It was not the compositing: five independent
defects sat between ExoPlayer and the screen, each able to produce that symptom
on its own — the app shell painting over the surface through a CSS rule aimed at
an attribute nothing set, a poster card with no way to lift on a path that
renders no <video>, JS bridges racing the page load and losing permanently, a
SurfaceView that was never detached, and a frontend that told Rust a webview
element was playing when none existed, so every play/pause intent was aimed at
something that was not there.

Native video stays opt-in. Turning it on surfaced a further unverified path —
the background-audio return is written only for the webview element — and
rotation still needs device confirmation.

Minor rather than patch: the player's touch behaviour changes for everyone (the
control bar now auto-hides on touchscreens, and the system bars go away with the
player), not only for those who opt into native video.

.gitea/workflows/build-release.yml

@@ -96,6 +96,12 @@ jobs:
       - name: Install dependencies
         run: bun install
 
+      # The Linux job previously had no version step at all, so a tagged release
+      # built Linux packages from whatever version happened to be committed.
+      - name: Set app version from tag
+        run: ./scripts/set-version.sh "${GITHUB_REF#refs/tags/}"
+        if: startsWith(github.ref, 'refs/tags/v')
+
       - name: Build for Linux
         run: bun run tauri build
         env:
@@ -121,6 +127,62 @@ jobs:
           path: dist/linux/
           retention-days: 30
 
+  build-windows:
+    name: Build Windows
+    runs-on: linux/amd64
+    needs: test
+    # Cross-compiled from Linux via the official Tauri path (MSVC + cargo-xwin),
+    # baked into the builder image. No toolchain installs here — the image has
+    # cargo-xwin, clang/clang-cl, lld, llvm, nsis and the msvc target.
+    container:
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Cache Rust dependencies
+        uses: actions/cache@v3
+        with:
+          path: |
+            ~/.cargo/registry
+            ~/.cargo/git
+            ~/.cache/cargo-xwin
+            src-tauri/target
+          key: ${{ runner.os }}-cargo-windows-${{ hashFiles('**/Cargo.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-cargo-windows-
+
+      - name: Cache Node dependencies
+        uses: actions/cache@v3
+        with:
+          path: |
+            ~/.bun/install/cache
+            node_modules
+          key: ${{ runner.os }}-bun-${{ hashFiles('**/bun.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-bun-
+
+      # The tag is the single source of truth for a release version; the script
+      # stamps every file that carries it (package.json, tauri.conf.json,
+      # Cargo.toml, Cargo.lock). This step used to sed only tauri.conf.json, so
+      # the other three shipped whatever was committed.
+      - name: Set app version from tag
+        run: ./scripts/set-version.sh "${GITHUB_REF#refs/tags/}"
+        if: startsWith(github.ref, 'refs/tags/v')
+
+      - name: Build Windows (NSIS installer + exe)
+        run: OUTPUT_DIR="$PWD/dist/windows" WIN_BUNDLES=nsis ./scripts/build-windows-cross.sh
+
+      - name: List Windows artifacts
+        run: ls -lah dist/windows/
+
+      - name: Upload Windows build artifact
+        uses: actions/upload-artifact@v3
+        with:
+          name: jellytau-windows
+          path: dist/windows/
+          retention-days: 30
+
   build-android:
     name: Build Android
     runs-on: linux/amd64
@@ -159,48 +221,22 @@ jobs:
       - name: Install dependencies
         run: bun install
 
+      # Stamp before `android init`: it derives its generated project (including
+      # the initial versionCode) from tauri.conf.json.
       - name: Set app version from tag
-        run: |
-          # On a tag build, the tag is the single source of truth for the
-          # version name. On non-tag runs keep whatever is in tauri.conf.json.
-          if echo "$GITHUB_REF" | grep -q '^refs/tags/v'; then
-            VERSION="${GITHUB_REF#refs/tags/v}"
-            echo "Setting version to $VERSION"
-            sed -i "s/\"version\": \"[^\"]*\"/\"version\": \"$VERSION\"/" src-tauri/tauri.conf.json
-          fi
-          grep '"version"' src-tauri/tauri.conf.json
+        run: ./scripts/set-version.sh "${GITHUB_REF#refs/tags/}"
+        if: startsWith(github.ref, 'refs/tags/v')
 
       - name: Initialize Android project
         run: bun run tauri android init
 
+      # Re-run after init: tauri.properties only exists now, and its
+      # autogenerated versionCode (0.0.15 -> 15) is both tiny and NOT monotonic
+      # against the 1000 floor already shipped in the field. The script rewrites
+      # it as 1000 + major*10000 + minor*100 + patch. Runs unconditionally so
+      # untagged builds get a sane code too, derived from git describe.
       - name: Pin a monotonic Android versionCode
-        run: |
-          # `tauri android init` autogenerates src-tauri/gen/android/app/tauri.properties
-          # with a versionCode derived from the semver (e.g. 0.0.15 -> 15). That
-          # number is (a) tiny and (b) NOT monotonic across our history: earlier
-          # local/dev builds shipped versionCode 1000 (from a 0.1.0 config), so a
-          # plain 15 would be a *downgrade* and Android would refuse the update.
-          #
-          # Derive an explicit code that is both monotonic in semver order and
-          # always above the 1000 floor already in the field:
-          #   code = 1000 + major*10000 + minor*100 + patch
-          # e.g. 0.0.14 -> 1014, 0.0.15 -> 1015, 0.1.0 -> 1100, 1.0.0 -> 11000.
-          # POSIX sh only (the runner uses dash): no here-strings, no \s in sed.
-          PROPS="src-tauri/gen/android/app/tauri.properties"
-          VERSION=$(grep '"version"' src-tauri/tauri.conf.json | head -1 | sed -E 's/.*"version"[[:space:]]*:[[:space:]]*"([^"]+)".*/\1/')
-          MAJ=$(echo "$VERSION" | cut -d. -f1)
-          MIN=$(echo "$VERSION" | cut -d. -f2)
-          PAT=$(echo "$VERSION" | cut -d. -f3)
-          # Guard against a malformed/missing component so we never emit code 0.
-          : "${MAJ:=0}" "${MIN:=0}" "${PAT:=0}"
-          CODE=$(( 1000 + MAJ*10000 + MIN*100 + PAT ))
-          echo "version=$VERSION -> versionCode=$CODE"
-          if grep -q '^tauri.android.versionCode=' "$PROPS"; then
-            sed -i "s/^tauri.android.versionCode=.*/tauri.android.versionCode=$CODE/" "$PROPS"
-          else
-            echo "tauri.android.versionCode=$CODE" >> "$PROPS"
-          fi
-          cat "$PROPS"
+        run: ./scripts/set-version.sh "${GITHUB_REF#refs/tags/}"
 
       - name: Sync custom Android sources & gradle config
         run: ./scripts/sync-android-sources.sh
@@ -239,7 +275,7 @@ jobs:
   create-release:
     name: Create Release
     runs-on: linux/amd64
-    needs: [build-linux, build-android]
+    needs: [build-linux, build-windows, build-android]
     if: startsWith(github.ref, 'refs/tags/v')
     container:
       image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
@@ -259,6 +295,12 @@ jobs:
           name: jellytau-linux
           path: artifacts/linux/
 
+      - name: Download Windows artifacts
+        uses: actions/download-artifact@v3
+        with:
+          name: jellytau-windows
+          path: artifacts/windows/
+
       - name: Download Android artifacts
         uses: actions/download-artifact@v3
         with:
@@ -277,6 +319,9 @@ jobs:
           echo "- **AppImage** - Run directly on most Linux distributions" >> release_notes.md
           echo "- **DEB** - Install via \`sudo dpkg -i jellytau_*.deb\` (Ubuntu/Debian)" >> release_notes.md
           echo "" >> release_notes.md
+          echo "#### Windows" >> release_notes.md
+          echo "- **Installer (.exe)** - Run \`jellytau_*-setup.exe\` (NSIS). Unsigned — SmartScreen may warn on first run." >> release_notes.md
+          echo "" >> release_notes.md
           echo "#### Android" >> release_notes.md
           echo "- **APK** - Install via \`adb install jellytau-release.apk\` or sideload via file manager" >> release_notes.md
           echo "- **AAB** - Upload to Google Play Console or testing platforms" >> release_notes.md
@@ -358,7 +403,7 @@ jobs:
           fi
           echo "Release id=$RELEASE_ID"
 
-          for f in artifacts/android/* artifacts/linux/*; do
+          for f in artifacts/android/* artifacts/linux/* artifacts/windows/*; do
             [ -f "$f" ] || continue
             echo "⬆️  Uploading $(basename "$f")"
             curl -fsS -X POST \

.gitea/workflows/traceability-check.yml

@@ -42,30 +42,45 @@ jobs:
           echo "📊 Validating requirement traceability..."
           echo ""
 
-          # Parse JSON
+          # Denominators come from docs/requirements.md at run time — NEVER
+          # hardcode them here. This step previously divided by frozen literals
+          # (UR/39, IR/24, DR/48, JA/3, total 114) while the file had grown to
+          # 211 requirements, so it reported 158% coverage and the threshold
+          # below could never trip. See docs/specs/traceability-gate-repair.md.
           TOTAL_TRACES=$(jq '.totalTraces' traces-report.json)
-          UR=$(jq '.byType.UR | length' traces-report.json)
-          IR=$(jq '.byType.IR | length' traces-report.json)
-          DR=$(jq '.byType.DR | length' traces-report.json)
-          JA=$(jq '.byType.JA | length' traces-report.json)
+          COVERED=$(jq '.coverage.covered' traces-report.json)
+          TOTAL_REQS=$(jq '.coverage.total' traces-report.json)
+          COVERAGE=$(jq '.coverage.percent' traces-report.json)
 
-          # Print coverage report
           echo "✅ TRACES Found: $TOTAL_TRACES"
           echo ""
-          echo "📋 Coverage Summary:"
-          echo "   User Requirements (UR):         $UR / 39 ($(( UR * 100 / 39 ))%)"
-          echo "   Integration Requirements (IR): $IR / 24 ($(( IR * 100 / 24 ))%)"
-          echo "   Development Requirements (DR): $DR / 48 ($(( DR * 100 / 48 ))%)"
-          echo "   Jellyfin API Requirements (JA): $JA / 3 ($(( JA * 100 / 3 ))%)"
+          echo "📋 Coverage Summary (traced / defined):"
+          for T in UR IR DR JA; do
+            TRACED=$(jq --arg t "$T" '[.byType[$t][] | select(. != null)] | length' traces-report.json)
+            DEFINED=$(jq --arg t "$T" '.defined[$t]' traces-report.json)
+            echo "   $T: $TRACED / $DEFINED"
+          done
           echo ""
 
-          COVERED=$((UR + IR + DR + JA))
-          TOTAL_REQS=114
-          COVERAGE=$((COVERED * 100 / TOTAL_REQS))
-
           echo "📈 Overall Coverage: $COVERED / $TOTAL_REQS ($COVERAGE%)"
           echo ""
 
+          # Traced IDs that requirements.md does not define (typo, or a deleted
+          # requirement). These do not count toward coverage.
+          ORPHANED=$(jq -c '.coverage.orphaned' traces-report.json)
+          if [ "$ORPHANED" != "[]" ]; then
+            echo "⚠️  Traced but not defined in requirements.md: $ORPHANED"
+            echo ""
+          fi
+
+          # A ratio above 100% means the computation is broken — the exact
+          # condition that hid the stale-denominator bug. Fail loudly.
+          if [ "$COVERAGE" -gt 100 ]; then
+            echo "❌ ERROR: Coverage ($COVERAGE%) exceeds 100% — the gate is miscomputing."
+            echo "   Orphaned IDs: $ORPHANED"
+            exit 1
+          fi
+
           # Check minimum threshold
           MIN_THRESHOLD=50
           if [ "$COVERAGE" -lt "$MIN_THRESHOLD" ]; then

.gitignore

@@ -64,3 +64,9 @@ src-tauri/.cargo/config.toml
 /docs/README.md
 /docs/api-redirect.md
 /docs-site/book/
+
+# Arch packaging build artifacts (vendored cargo cache, makepkg workdir, output package)
+/.cargo-arch/
+/packaging/arch/pkg/
+/packaging/arch/src/
+/packaging/arch/*.pkg.tar.zst

CLAUDE.md

@@ -31,10 +31,34 @@ bun run android:dev              # build + deploy
 bun run android:logs             # logcat
 ```
 
+The **debug** build type carries `applicationIdSuffix ".debug"`, so
+`com.dtourolle.jellytau.debug` ("JellyTau Debug") installs *alongside* a release
+build with its own data dir — never uninstall the release app to test a debug
+one. `./scripts/build-and-deploy.sh release --device --debug` puts an
+R8-minified *release* build in that same slot, signed with the local debug
+keystore, for validating minification without the real key. Only the
+applicationId is suffixed; Kotlin classes stay in the `namespace` package
+`com.dtourolle.jellytau`, so JNI lookups and R8 keep rules are unaffected. See
+[README_ANDROID_BUILD.md](src-tauri/android/README_ANDROID_BUILD.md).
+
 CI runs on **Gitea Actions** (`.gitea/workflows/`), not GitHub. Use the `gh` CLI
 only against the mirror if one exists; the canonical remote is
 `gitea.tourolle.paris`.
 
+> **🔴 CI installs no system tools.** Never add an `apt-get`, `rustup`,
+> `sdkmanager`, mingw/nsis, or any other *toolchain/system-package* install to a
+> CI workflow step. Every build, test, and packaging **tool** must already live
+> in the Docker image the job runs in — the unified builder (`Dockerfile.builder`
+> → `gitea.tourolle.paris/dtourolle/jellytau-builder`) for Android/Linux/Windows,
+> or `Dockerfile.arch` for Arch. If a job needs a tool the image lacks, **add it
+> to the image, rebuild + push it** (`scripts/build-builder-image.sh`), and use
+> it from CI — do not install it at job time. This keeps builds reproducible and
+> fast, and is why the packaging stages are thin `FROM ${BUILDER_IMAGE}` layers.
+>
+> `bun install` (fetching the project's own JS deps per the lockfile) is **not**
+> a violation — that's project dependencies, not a toolchain. The rule is about
+> system tools, not npm/bun/cargo *packages* declared by the project.
+
 ## Before Committing
 
 - Frontend: `bun run check` and `bun run test` must pass.
@@ -169,8 +193,14 @@ and [docs/build-release.md](docs/build-release.md).
   backend expand it. Single-type presentation (`itemType: "Movie"`, "this page
   shows albums") is fine; a *category → set of types* mapping in `src/` is a leak.
   `bun run check:boundary` is the tripwire; the real gate is the spec's layer
-  assignment. See [scoped-search-boundary.md](docs/specs/scoped-search-boundary.md)
-  for the incident this rule came from.
+  assignment. The canonical example lives in Rust:
+  `SearchScope::item_types()` in `repository/types.rs` expands an opaque scope the
+  frontend sends. See [scoped-search-boundary.md](docs/specs/scoped-search-boundary.md)
+  for the incident this rule came from — note the tripwire missed that leak for
+  months because the mapping was assigned to a named const rather than written
+  inline at the query, so **a green `check:boundary` is not proof**; it flags
+  item-type array literals only, not run-time-built sets or `switch`/`||`
+  taxonomy.
 
 ## Writing specs
 
@@ -252,6 +282,23 @@ tagged responses keep the Rust field names as-is (e.g. `new_url`, not `newUrl`).
 
 ## Testing
 
+### 🔴 Bug fixes: failing test FIRST, then the fix
+
+When fixing a bug, **write a test that reproduces it and watch it fail before
+touching the fix.** Red → green, in that order:
+
+1. Write a test that exercises the broken behavior and **run it — it must fail**,
+   proving the test actually catches the bug (a test that passes before the fix
+   proves nothing).
+2. Apply the fix.
+3. Re-run — the test now passes, and so does the rest of the suite.
+
+Never fix first and backfill the test afterward: a test written against
+already-fixed code can pass for the wrong reason and silently fails to guard the
+regression. If the logic is buried in a component, extract the pure part into a
+plain `.ts` module (e.g. `episodeStrip.ts`) so it can be unit-tested — the same
+pattern as `TrackList.logic.test.ts`.
+
 ```bash
 # Rust
 cd src-tauri && cargo test

Dockerfile

@@ -1,4 +1,11 @@
 # Multi-stage build for JellyTau - Tauri Jellyfin client
+#
+# The desktop packaging stages (desktop-linux-build, windows-cross) build FROM
+# the unified registry builder image, which carries every packaging tool. Declared
+# here (before the first FROM) so it's in scope for those stages' FROM lines.
+# Override for local iteration: --build-arg BUILDER_IMAGE=jellytau-builder:latest
+ARG BUILDER_IMAGE=gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+
 FROM ubuntu:24.04 AS builder
 
 ENV DEBIAN_FRONTEND=noninteractive \
@@ -108,6 +115,34 @@ RUN cd src-tauri && cargo fetch && cd .. && \
     bun run tauri android build --apk true && \
     echo "APK build complete!"
 
+# Desktop packaging stages build FROM the unified registry builder image (see the
+# BUILDER_IMAGE ARG at the top), which already carries every packaging tool
+# (rpm/file for Linux, cargo-xwin + nsis + the x86_64-pc-windows-msvc rust
+# target for Windows). ONE source of dependency truth, shared with CI — no
+# per-stage apt/rustup here.
+#
+# NOTE: Windows uses the MSVC target via cargo-xwin, NOT mingw/GNU — the GNU
+# toolchain cannot bundle an NSIS installer from Linux. See
+# scripts/build-windows-cross.sh.
+
+# Linux desktop packaging environment (deb + rpm; Arch is Dockerfile.arch).
+# Thin layer over the builder — the actual build runs at container-run time on
+# the bind-mounted source (see docker-compose.yml / scripts/build-desktop-linux.sh),
+# matching the `dev` service model. Run standalone with:
+#   docker run --rm -v "$PWD:/app" -v "$PWD/dist:/app/dist" <img> \
+#     bash -c "OUTPUT_DIR=/app/dist scripts/build-desktop-linux.sh"
+FROM ${BUILDER_IMAGE} AS desktop-linux-build
+WORKDIR /app
+CMD ["bash", "-c", "OUTPUT_DIR=/app/dist scripts/build-desktop-linux.sh"]
+
+# Windows cross-compile environment (MSVC target via cargo-xwin). Video works via
+# WebView2 and audio via the webview <audio> backend; NSIS installer is produced
+# from Linux by cargo-xwin. Default bundles NSIS; override WIN_BUNDLES=none for
+# exe-only. Build runs at container-run time like above.
+FROM ${BUILDER_IMAGE} AS windows-cross
+WORKDIR /app
+CMD ["bash", "-c", "OUTPUT_DIR=/app/dist WIN_BUNDLES=${WIN_BUNDLES:-nsis} scripts/build-windows-cross.sh"]
+
 # Final output stage
 FROM ubuntu:24.04 AS final
 RUN apt-get update && apt-get install -y --no-install-recommends \

Dockerfile.arch

@@ -0,0 +1,37 @@
+# JellyTau Arch Linux package builder.
+#
+# Tauri has no pacman bundle target, so we build a real .pkg.tar.zst with makepkg
+# from packaging/arch/PKGBUILD. makepkg refuses to run as root, so we create a
+# non-root `builder` user with passwordless sudo (for `makepkg -s` pacman calls).
+#
+#   docker build -f Dockerfile.arch -t jellytau-arch .
+#   docker run --rm -v "$PWD/dist:/out" jellytau-arch
+FROM archlinux:latest
+
+RUN pacman -Syu --noconfirm \
+    base-devel git sudo \
+    rust cargo nodejs \
+    webkit2gtk-4.1 mpv gtk3 libayatana-appindicator \
+    libsoup3 pkgconf openssl \
+    && pacman -Scc --noconfirm
+
+# Bun is not in the official repos; install the upstream binary.
+RUN curl -fsSL https://bun.sh/install | bash && \
+    ln -s /root/.bun/bin/bun /usr/local/bin/bun
+
+# Non-root build user with passwordless sudo for makepkg's dependency step.
+RUN useradd -m builder && \
+    echo 'builder ALL=(ALL) NOPASSWD: ALL' > /etc/sudoers.d/builder && \
+    ln -sf /root/.bun/bin/bun /usr/local/bin/bun
+
+WORKDIR /app
+COPY . .
+RUN chown -R builder:builder /app
+
+USER builder
+ENV OUTPUT_DIR=/out
+RUN mkdir -p /out
+VOLUME ["/out"]
+
+# Default: build the package. Output lands in /out (mount it to collect the pkg).
+CMD ["bash", "-c", "OUTPUT_DIR=/out scripts/build-arch.sh"]

Dockerfile.builder

@@ -1,5 +1,9 @@
 # JellyTau Builder Image
-# Pre-built image with all dependencies for building and testing
+# Pre-built image with all dependencies for building, testing, and packaging:
+#   - Android APK (SDK/NDK), Linux desktop (deb/rpm),
+#   - Windows cross via the official Tauri path: MSVC target + cargo-xwin + NSIS
+# Arch packages build in a separate archlinux image (Dockerfile.arch) since
+# makepkg is Arch-specific.
 # Push to your registry: docker build -f Dockerfile.builder -t gitea.tourolle.paris/dtourolle/jellytau-builder:latest .
 
 FROM ubuntu:24.04
@@ -83,6 +87,50 @@ RUN $ANDROID_HOME/cmdline-tools/latest/bin/sdkmanager --sdk_root=$ANDROID_HOME \
 # Set NDK environment variable
 ENV NDK_HOME=$ANDROID_HOME/ndk/$NDK_VERSION
 
+# Gradle distribution. `tauri android init` regenerates gen/android with a
+# wrapper pointing at services.gradle.org, so every Android job would otherwise
+# download ~130MB of Gradle at build time — slow, and a hard failure when the
+# CDN hiccups ("Unexpected end of file from server"). Ship the distribution in
+# the image instead; scripts/sync-android-sources.sh repoints the regenerated
+# wrapper at this local copy. Keep GRADLE_VERSION in sync with the version
+# Tauri's generated wrapper requests.
+ENV GRADLE_VERSION=8.14.3 \
+    GRADLE_HOME=/opt/gradle/gradle-8.14.3
+RUN mkdir -p /opt/gradle/dist && \
+    wget -q "https://services.gradle.org/distributions/gradle-${GRADLE_VERSION}-bin.zip" \
+        -O "/opt/gradle/dist/gradle-${GRADLE_VERSION}-bin.zip" && \
+    unzip -q "/opt/gradle/dist/gradle-${GRADLE_VERSION}-bin.zip" -d /opt/gradle && \
+    "$GRADLE_HOME/bin/gradle" --version
+ENV PATH="$GRADLE_HOME/bin:$PATH"
+
+# ---------------------------------------------------------------------------
+# Desktop packaging tools — kept in a trailing layer ON PURPOSE so that adding
+# or changing a packaging tool doesn't invalidate the expensive apt/rust/Android
+# layers above (a tool tweak becomes a ~1-2 min rebuild, not ~15). Covers Linux
+# (deb/rpm) and Windows cross (MSVC via cargo-xwin + NSIS).
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    # Linux desktop packaging: rpmbuild for the .rpm bundle (deb needs nothing extra)
+    rpm \
+    file \
+    # Windows cross-compile (official Tauri path: MSVC target via cargo-xwin).
+    # clang provides clang-cl, the MSVC-compatible C compiler cc-rs uses to build
+    # C deps (bundled sqlite, ring, ...); lld = linker; llvm = llvm-lib/ar etc;
+    # nsis = installer generator.
+    clang \
+    lld \
+    llvm \
+    nsis \
+    && rm -rf /var/lib/apt/lists/* \
+    # Ubuntu's clang package ships clang but NOT the clang-cl alias that cc-rs
+    # invokes for MSVC targets. clang-cl is the same binary in MSVC-compat mode,
+    # so provide it as a symlink.
+    && ln -sf /usr/bin/clang /usr/local/bin/clang-cl
+
+# Windows rust target + cargo-xwin (downloads the MSVC CRT/SDK at build time).
+RUN . $HOME/.cargo/env && \
+    rustup target add x86_64-pc-windows-msvc && \
+    cargo install --locked cargo-xwin
+
 WORKDIR /app
 
 ENTRYPOINT ["/bin/bash"]

docker-compose.yml

@@ -33,6 +33,58 @@ services:
     ports:
       - "5172:5172"  # In case you want to run dev server
 
+  # Linux desktop packages - deb + rpm + pacman into ./dist
+  desktop-linux-build:
+    build:
+      context: .
+      dockerfile: Dockerfile
+      target: desktop-linux-build
+      args:
+        # Defaults to the registry builder (Dockerfile's ARG). Point at a locally
+        # built builder with: BUILDER_IMAGE=jellytau-builder:latest docker compose ...
+        BUILDER_IMAGE: ${BUILDER_IMAGE:-gitea.tourolle.paris/dtourolle/jellytau-builder:latest}
+    container_name: jellytau-desktop-linux-build
+    volumes:
+      - .:/app
+      - cargo-cache:/root/.cargo
+      - bun-cache:/root/.bun
+    environment:
+      - RUST_BACKTRACE=1
+      - OUTPUT_DIR=/app/dist
+    command: bash -c "OUTPUT_DIR=/app/dist scripts/build-desktop-linux.sh"
+
+  # Arch Linux package (.pkg.tar.zst via makepkg) into ./dist
+  arch-build:
+    build:
+      context: .
+      dockerfile: Dockerfile.arch
+    container_name: jellytau-arch-build
+    volumes:
+      - ./dist:/out
+    environment:
+      - RUST_BACKTRACE=1
+      - OUTPUT_DIR=/out
+
+  # Windows cross-compile (MSVC via cargo-xwin). Emits NSIS installer + .exe to
+  # ./dist. Override WIN_BUNDLES=none for exe-only.
+  windows-cross:
+    build:
+      context: .
+      dockerfile: Dockerfile
+      target: windows-cross
+      args:
+        BUILDER_IMAGE: ${BUILDER_IMAGE:-gitea.tourolle.paris/dtourolle/jellytau-builder:latest}
+    container_name: jellytau-windows-cross
+    volumes:
+      - .:/app
+      - cargo-cache:/root/.cargo
+      - bun-cache:/root/.bun
+    environment:
+      - RUST_BACKTRACE=1
+      - OUTPUT_DIR=/app/dist
+      - WIN_BUNDLES=${WIN_BUNDLES:-nsis}
+    command: bash -c "OUTPUT_DIR=/app/dist WIN_BUNDLES=${WIN_BUNDLES:-nsis} scripts/build-windows-cross.sh"
+
   # Development container - for interactive development
   dev:
     build:

docs/build-desktop-packages.md

@@ -0,0 +1,91 @@
+# Desktop packaging (Linux, Arch, Windows)
+
+How to produce distributable desktop packages for JellyTau. All three flows can
+run in Docker so no host toolchain setup is required. Outputs land in `./dist`.
+
+## One builder image (shared with CI)
+
+The deb/rpm and Windows-cross flows build on the **unified registry builder**
+([../Dockerfile.builder](../Dockerfile.builder) →
+`gitea.tourolle.paris/dtourolle/jellytau-builder`), the same image CI uses. It
+carries every packaging tool: Android SDK/NDK, `rpm`/`file` (Linux bundler),
+`cargo-xwin` + `lld` + `llvm` + `nsis` + the `x86_64-pc-windows-msvc` rust target
+(Windows). There is **one** dependency source of truth — no per-stage tool
+installs.
+
+The desktop stages in [../Dockerfile](../Dockerfile) are thin `FROM
+${BUILDER_IMAGE}` environments; the actual build runs at container-run time on
+your bind-mounted source (like the `dev` service), so source edits need no image
+rebuild.
+
+**If you changed `Dockerfile.builder`** (e.g. added a tool), rebuild and push it
+first, or the packaging flows use the stale registry image:
+
+```bash
+scripts/build-builder-image.sh           # build + push :latest to the registry
+# ...or iterate locally without pushing:
+docker build -f Dockerfile.builder -t jellytau-builder:latest .
+BUILDER_IMAGE=jellytau-builder:latest bun run docker:build:windows
+```
+
+Arch uses a separate `archlinux` image ([../Dockerfile.arch](../Dockerfile.arch))
+because `makepkg` is Arch-specific — it is not part of the unified builder.
+
+| Target | Format | Docker command | Functional? |
+|--------|--------|----------------|-------------|
+| Debian/Ubuntu, Fedora | `.deb`, `.rpm` | `bun run docker:build:linux` | ✅ yes |
+| Arch Linux | `.pkg.tar.zst` | `bun run docker:build:arch` | ✅ yes |
+| Windows | NSIS installer + `.exe` | `bun run docker:build:windows` | ✅ yes (unsigned) |
+
+## Linux: deb + rpm
+
+Tauri's bundler produces these natively. The build runs on the existing Ubuntu
+builder image ([../Dockerfile](../Dockerfile), `desktop-linux-build` stage):
+
+```bash
+bun run docker:build:linux            # deb + rpm -> ./dist
+# or, on a host with the Tauri Linux deps installed:
+BUNDLES="deb,rpm" scripts/build-desktop-linux.sh
+```
+
+Runtime dependency: the app links libmpv (audio) and WebKitGTK (webview + HTML5
+transcoded video). The deb/rpm declare these.
+
+> Note: `appimage` is also a valid Tauri target if you want a portable bundle —
+> add it to `BUNDLES`.
+
+## Arch Linux: pacman package
+
+**Tauri has no `pacman` bundle target** (as of tauri-cli 2.9.x — valid targets
+are deb/rpm/appimage/msi/nsis/app/dmg). So we ship a hand-written PKGBUILD in
+[../packaging/arch/PKGBUILD](../packaging/arch/PKGBUILD) and build it with
+`makepkg` on an Arch base image ([../Dockerfile.arch](../Dockerfile.arch)):
+
+```bash
+bun run docker:build:arch             # .pkg.tar.zst -> ./dist
+```
+
+The PKGBUILD is AUR-ready: swap its `source=()` for a release tarball/VCS URL to
+publish. Runtime deps: `webkit2gtk-4.1`, `mpv`, `gtk3`, `libayatana-appindicator`.
+
+`makepkg` refuses to run as root, so the Docker stage builds as a non-root
+`builder` user. Because the image `COPY`s the source at build time, the
+`arch-build` compose service does **not** bind-mount the repo — rebuild the image
+to pick up source changes.
+
+## Windows: NSIS installer cross-compiled from Linux
+
+Produces a working (unsigned) NSIS installer + `.exe` via the official Tauri
+cross-compile path — the `x86_64-pc-windows-msvc` target driven by `cargo-xwin`.
+Video plays via WebView2 and audio via the webview `<audio>` backend. See
+[build-windows.md](build-windows.md) for the full explanation.
+
+```bash
+bun run docker:build:windows                     # NSIS installer + .exe -> ./dist
+WIN_BUNDLES=none bun run docker:build:windows     # exe only, skip bundling
+```
+
+The Docker `windows-cross` stage is a thin layer over the builder, which carries
+`cargo-xwin` + `lld` + `llvm` + `nsis` + the `x86_64-pc-windows-msvc` target.
+Cross-compilation is Tauri's "last resort" path (less tested than building on
+Windows); a `windows-latest` CI job is the fallback if it misbehaves.

docs/build-windows.md

@@ -0,0 +1,78 @@
+# Windows build
+
+JellyTau targets Linux and Android primarily, but a working Windows build —
+including an **NSIS installer cross-compiled from Linux** — is produced by the
+Docker tooling. It is not yet a first-class release target (no code signing / CI
+job / SMTC lockscreen), but it runs and plays media.
+
+## How playback works on Windows
+
+- **Video** — renders through the webview HTML5 `<video>` element (hls.js) on
+  *every* platform; on Windows that is WebView2 (Chromium/Edge), which plays HLS +
+  h264 fine. No Windows-specific code.
+- **Audio-only (music)** — the native audio backends are libmpv (Linux) and
+  ExoPlayer (Android); neither exists on Windows. Instead
+  `create_player_backend()` in [../src-tauri/src/lib.rs](../src-tauri/src/lib.rs)
+  uses `WebviewAudioBackend` on non-Linux/non-Android targets: it hands the stream
+  URL to a webview `<audio>` element (see
+  [../src/lib/services/webviewAudio.ts](../src/lib/services/webviewAudio.ts)),
+  which reports state back through the same `player_report_*` round-trip the video
+  path uses. Pure Rust + Tauri events.
+
+## Cross-compiling from Linux (MSVC + cargo-xwin)
+
+We use the [official Tauri cross-compile path](https://v2.tauri.app/distribute/windows-installer/):
+the **MSVC** target (`x86_64-pc-windows-msvc`) driven by
+[`cargo-xwin`](https://github.com/rust-cross/cargo-xwin), which downloads the MSVC
+CRT / Windows SDK headers and links with `lld`. MSVC is the target Tauri
+officially supports for Windows (mingw/GNU is not), and — unlike GNU — it lets the
+Tauri CLI bundle the **NSIS installer from a Linux host**.
+
+> Why not mingw/GNU? The GNU target *does* link a valid `.exe`, but the Tauri CLI
+> gates `--bundles` by the host OS unless it recognizes a real Windows build.
+> `--runner cargo-xwin --target x86_64-pc-windows-msvc` is what flips it into
+> Windows mode and enables the `nsis`/`msi` bundlers on Linux.
+
+The builder image ([../Dockerfile.builder](../Dockerfile.builder)) bakes in the
+whole toolchain: the `x86_64-pc-windows-msvc` rust target, `cargo-xwin`, `lld`,
+`llvm`, and `nsis`.
+
+```bash
+bun run docker:build:windows            # NSIS installer + .exe -> ./dist
+WIN_BUNDLES=none bun run docker:build:windows   # exe only, skip bundling
+```
+
+Or directly on a host that has the toolchain:
+
+```bash
+scripts/build-windows-cross.sh                    # nsis installer + exe
+WIN_BUNDLES=none scripts/build-windows-cross.sh   # exe only
+```
+
+Under the hood the build runs:
+
+```bash
+tauri build --runner cargo-xwin --target x86_64-pc-windows-msvc --bundles nsis
+```
+
+Outputs:
+- `.exe` — `src-tauri/target/x86_64-pc-windows-msvc/release/jellytau.exe`
+- NSIS installer — `.../release/bundle/nsis/*-setup.exe`
+
+(both copied to `./dist` when `OUTPUT_DIR` is set).
+
+## Caveats
+
+- **Cross-compilation is a last resort** per Tauri's own docs — it's less tested
+  than building on Windows. If it misbehaves, a `windows-latest` CI job or a
+  Windows VM building natively (`tauri build --bundles nsis`) is the fallback.
+- **Code signing is not wired up** — the installer is unsigned, so Windows
+  SmartScreen will warn on first run.
+
+## Outstanding for a first-class Windows release
+
+1. Gapless/crossfade + SMTC (lockscreen) — currently no-ops in the webview audio
+   path.
+2. Downloaded (`Local` source) file playback needs `convertFileSrc` on the
+   frontend; streaming works today.
+3. Code signing + a Windows packaging CI job.

docs/defect-windows.md

@@ -0,0 +1,150 @@
+# Defect windows — which bugs were present when
+
+For each fixed defect, the releases it was actually present in. Companion to
+[CHANGELOG.md](../CHANGELOG.md), which says what changed; this says how long each
+fault had been shipping before it did.
+
+**"Present since"** is the first *release* containing the defective code, not the
+first release where a user could hit it — those differ, sometimes by months, and
+the gap is called out where it matters. **"How dated"** records the evidence, so a
+row can be re-checked or disputed:
+
+| Method | Meaning |
+|--------|---------|
+| `pickaxe` | `git log -S<token>` on the defective token — the commit that introduced the exact string, then the earliest tag containing it. Strongest evidence. |
+| `feature` | The defect is inseparable from a feature that landed whole (bad rung in a new algorithm, missing caller in new plumbing), dated to that feature's release. |
+| `absence` | The fix *adds* something that was never there. Dated to when the surrounding code was built, since there is no introducing commit to find. Weakest — treat as "no later than". |
+
+## Present since the first release
+
+Nine defects date to the initial proof of concept (v0.0.1, 2026-06-23) and shipped
+for between two weeks and seven weeks short of two months before anyone hit them.
+That is the dominant pattern here: not regressions, but original assumptions that
+went unexercised until a later feature leaned on them.
+
+| Defect | Present since | Fixed in | Shipped broken for | How dated |
+|---|---|---|---|---|
+| `AudioStreamIndex=0` pinned the video stream as the audio track (DR-140) | v0.0.1 | **v0.4.6** | ~7 weeks | pickaxe |
+| Download URL spelled `videoBitrate`, which Jellyfin does not bind (DR-123) | v0.0.1 | **v0.5.1** | ~7 weeks | pickaxe |
+| `pause_download` / `resume_download` were no-ops (DR-168) | v0.0.1 | **v0.5.3** | ~7.5 weeks | pickaxe |
+| `.part` sidecar named by `with_extension`, so no cleanup path matched it (DR-169) | v0.0.1 | **v0.5.3** | ~7.5 weeks | pickaxe |
+| `Range` sent on every retry regardless of the response (DR-170) | v0.0.1 | **v0.5.3** | ~7.5 weeks | pickaxe |
+| `/Items/Latest` requested with the default `GroupItems=false` | v0.0.1 | **v0.5.1** | ~7 weeks | pickaxe |
+| `SubtitleStreamIndex` omitted from PlaybackInfo, letting the server burn in (DR-176) | v0.0.1 | **v0.5.5** | ~8 weeks | pickaxe |
+| No `PlaySessionId`, and one hardcoded `DeviceId`, on every stream URL (DR-177) | v0.0.1 | **v0.5.5** | ~8 weeks | pickaxe |
+| `download_item` never recorded `media_type`; NULL read as `'audio'` (DR-135) | v0.0.1 | **v0.4.6** | ~7 weeks | pickaxe |
+| `download_album` read its track list from the local cache (DR-173) | v0.0.1 | **v0.5.5** | ~8 weeks | pickaxe |
+| Device profile carried no `MaxAudioChannels` (DR-141) | v0.0.1 | **v0.4.6** | ~7 weeks | absence |
+| Streaming ceiling fixed at 20 Mbps with no way to lower it (UR-074) | v0.0.1 | **v0.5.3** (as a feature) | ~7.5 weeks | pickaxe |
+
+### Why they took so long to surface
+
+Four of these were **latent until a later feature exercised them**, which is why
+the fix lands so far from the cause:
+
+- The `videoBitrate` casing was harmless while every download was `original`. It
+  became visible only once a quality picker existed to select against — and then
+  produced no error, just a full-size file, because Jellyfin discards an unbound
+  query key silently.
+- The unconditional `Range` header was inert for the same reason: `original` is
+  the one rung served with a `Content-Length` and real byte-range support. It
+  started corrupting files in **v0.5.1**, the moment the casing fix made
+  transcoded downloads actually transcode. So the *code* dates to v0.0.1 and the
+  *corruption* to v0.5.1 — a one-release window for the visible symptom.
+- The missing `PlaySessionId` only bites when a stream is re-opened for the same
+  item. Nothing re-opened one until quality switching, transcoded seek and
+  audio-track switching existed.
+- The omitted `SubtitleStreamIndex` only bites on sources whose own default
+  subtitle track is image-based, since that is what forces the server from
+  sidecar to burn-in.
+
+Two were **masked by soft failure**: the asset protocol being disabled (DR-134)
+was hidden by the thumbnail cache falling back to the server copy whenever the
+server was reachable, and `AudioStreamIndex=0` was hidden by servers that
+silently correct an out-of-range index — which is exactly why it was reported as
+"*some* videos have no audio" rather than as a bug in the client.
+
+## Introduced by a feature, fixed later
+
+| Defect | Present since | Fixed in | How dated |
+|---|---|---|---|
+| Native-path resume position never applied (both layers assumed the other seeked) | v0.0.9/v0.0.10 | **v0.5.1** | feature (`PlayerAdapter` contract) |
+| `get_downloaded_items` matched "this library exists" rather than constraining the item to it (DR-167) | v0.0.17 | **v0.5.3** | feature (browsable downloaded library) |
+| `SCOPE_ITEM_TYPES` — the frontend/backend boundary leak (DR-063) | v0.0.17 | **v0.2.1** | pickaxe |
+| `check:boundary` anchored to the query site, blind to a named const (DR-094) | v0.0.17 | **v0.2.1** | feature (tripwire landed with the leak it missed) |
+| Coverage gate divided by hardcoded denominators, reporting 158% (DR-093) | v0.0.1 | **v0.2.1** | pickaxe |
+| Tap deferral raced the WebView's synthesized click (DR-092 → DR-098) | v0.1.5 | **v0.2.7** | feature (the deferral itself) |
+| Transport for webview media decided from `el.paused` in the DOM (DR-097) | v0.0.9/v0.0.10 | **v0.2.7** | feature (`Html5PlayerAdapter`) |
+| `pick_current_episode` rung 3 returned the first *gap*, not the furthest watched | v0.3.0 | **v0.5.1** | feature |
+| `mirror_user_data` mirrored `is_favorite` alone and returned early (DR-155) | v0.4.0 | **v0.5.1** | pickaxe |
+| Stop-report path never fed the sync queue that existed for it (DR-154) | v0.4.6 | **v0.5.1** | feature (queue + drain landed with no producer) |
+| Background-audio base applied in two display-only places (DR-159) | v0.2.9 | **v0.5.3** | pickaxe |
+| Positions reported as 0 before the first tick, and always 0 for webview media (DR-178/179/180) | v0.5.3 | **v0.5.5** | feature (DR-159's tick boundary) |
+
+Three of these are worth separating out, because the defect is not a mistake in
+the code so much as **plumbing that was built and never connected**:
+
+- `repository_get_next_up_episodes` accepted a `series_id` from the day it was
+  written, and no caller passed one until v0.3.0.
+- The sync queue and its drain were built, tested and running in v0.4.6 with
+  neither of its two would-be producers ever called.
+- Both halves of the watched-state backend existed with no caller before v0.5.3.
+
+An automated check cannot see any of these — the code is present, tested and
+reachable in principle. Only tracing a requirement to a *call site* catches it.
+
+## Short windows (one release or less)
+
+| Defect | Present since | Fixed in | Note |
+|---|---|---|---|
+| `experimentalNativeVideo` defaulted on, shipping audio with a blank screen (DR-161 → DR-172) | v0.5.3 | **v0.5.4** | One release. The decode path was fine; the compositing step never ran. |
+| Webview-shaped audio profile insufficient — server ignores a profile's audio codec (DR-149) | v0.4.7 | **v0.4.8** | The v0.4.7 fix for DR-148 was necessary and not sufficient. |
+| Android `versionCode` floor went stale (`minor*100` yielding less than the 5002 already in the field) | v0.5.0 | **v0.5.3** | Caught before a broken APK shipped; no released build was un-installable. |
+| Subtitle sidecar work reverted by a commit assembled from a stale tree | v0.5.5 | **v0.5.5** | Never released broken — both commits are in v0.5.5. |
+
+## Fixed twice / never actually broken
+
+- **Autoplay time reset (v0.0.2).** Two commit objects carry this identical
+  change: `dcf08f30` (merged via Gitea PR #3, tagged v0.0.2) and `fa7cb6e9` (the
+  local original). Both have the same parent `674c8e5c` and the same diff. A merge
+  chain pulled `fa7cb6e9` and its follow-up `1e599627` into master's history
+  during v0.5.5, so `git log v0.5.4..v0.5.5` lists an autoplay fix that changed no
+  file in that release — `nextEpisodeService.ts` is byte-identical across the tag
+  boundary. The fix shipped in **v0.0.2** and has not regressed.
+
+  This is the one case where reading the changelog off `git log` subjects would
+  have produced a false entry, and it is a good argument for the project's
+  practice of deriving release notes from TRACES rather than commit subjects.
+
+## Recurring shapes
+
+Four causes account for most of the table:
+
+1. **An omitted parameter is not a neutral default.** `SubtitleStreamIndex`,
+   `AudioStreamIndex`, `GroupItems` and `MaxAudioChannels` all had a server-side
+   default that was actively wrong, and in three of the four the server's choice
+   was more expensive than the one intended — burn-in forcing a full re-encode
+   being the extreme case.
+2. **Silent binding failures.** `videoBitRate` produced no error, no warning and a
+   plausible-looking file. So did an unbound `Range`, and so did the coverage gate
+   dividing by a stale denominator.
+3. **Two layers each assuming the other acts.** Native resume (adapter recorded
+   the position, backend never seeked), end-of-playback dispatch (two paths, one
+   unreachable), and the surface/attach split in v0.5.0's native video.
+4. **A guard keyed on state that moves.** The tap deferral keyed suppression on a
+   timer handle the callback had already cleared; the HTML5 toggle keyed
+   play-vs-pause on `el.paused`, which flips while buffering.
+
+## Reproducing this
+
+The pickaxe rows can be re-derived directly:
+
+```bash
+git log --oneline --reverse -S'<defective token>' -- src-tauri/src   # introducing commit
+git tag --contains <sha> | sort -V | head -1                          # first release with it
+```
+
+Blaming the lines a fix removed (`git blame` at the fix's parent) is faster to run
+across many commits but was **not** used for the rows above: it reliably lands on
+whichever commit last touched the adjacent lines, which is usually not the commit
+that introduced the defect. It was used only to shortlist candidates.

docs/requirements.md

@@ -37,11 +37,11 @@ For a narrative overview of the system design, see
 | UR-024 | View recently added content on server | Medium | Done |
 | UR-025 | Sync watch history and progress back to Jellyfin | High | Done |
 | UR-026 | Sleep timer for audio and video playback (roller UI, time/track/episode modes) | Low | Done |
-| UR-027 | Audio equalizer for sound customization | Low | Planned |
+| UR-027 | Audio equalizer for sound customization | Low | Done (Linux only) |
 | UR-028 | Navigate to artist/album by tapping names in now playing view | High | Done |
 | UR-029 | Toggle between grid and list view in library | Medium | Done |
 | UR-030 | Quick genre browsing and filtering | Medium | Done |
-| UR-031 | Crossfade between audio tracks | Low | Done (Linux only) |
+| UR-031 | Crossfade between audio tracks | Low | Not implemented (blocked — see DR-034) |
 | UR-032 | Gapless playback for seamless album listening | Medium | Done (Linux only) |
 | UR-033 | Volume normalization to prevent volume jumps between tracks | Low | Done (Linux only) |
 | UR-034 | Rich home screen with hero banners, carousels, and personalized sections | High | Done |
@@ -62,12 +62,30 @@ For a narrative overview of the system design, see
 | UR-049 | Search is scoped by where it was started — inside a library it searches that library, from Home/library-root/search-tab it searches everything — with the scope shown as filter chips under the search bar that preselect from context and can be changed without retyping (see [ux-flows.md §6.1](ux-flows.md)) | High | Implemented |
 | UR-050 | Reorder search result groups (Songs, Albums, Artists, Movies, TV Shows) by drag and drop in settings, so the media a user cares about most appears first (see [ux-flows.md §6.3](ux-flows.md)) | Medium | Implemented |
 | UR-051 | Browse library pages in a consistent layout where card shape signals media type (square music, poster video, thumbnail episode), ordinal content stays listed, and the grid/list preference persists across pages (see [ux-flows.md §5A](ux-flows.md)) | Medium | Partial (implemented; toggle not reachable from settings) |
-| UR-052 | While offline, library pages show only media available on the device by default, with an opt-in toggle that additionally reveals the cached server catalog as greyed-out entries which can be queued for download on the next reconnect | High | Broken (toggle does not gate the listing; see issue #10) |
+| UR-052 | While offline, library pages show only media available on the device by default, with an opt-in toggle that additionally reveals the cached server catalog as greyed-out entries which can be queued for download on the next reconnect | High | Done |
 | UR-053 | Restrict media downloads to unmetered networks via a "WiFi Only" setting: when enabled, queued downloads are held while the device is on cellular or a metered connection (including metered WiFi hotspots) and resume automatically once an unmetered network is available | Medium | Done (pending device verification) |
 | UR-054 | Reach account actions (Settings, Downloads, Display preferences, Sign out) from every authenticated screen via a single account menu anchored to the user's name, identical on desktop and mobile (see [ux-flows.md §1.2](ux-flows.md)) | High | Done |
-| UR-055 | Browse downloaded media as an offline-scoped library — reusing the same library grids, cards, and detail pages as online browsing, showing only libraries/containers with downloaded content — with the transfer-progress list demoted to a secondary "Transfers" view (see [ux-flows.md §7.2](ux-flows.md)) | High | Planned |
-| UR-056 | See how much disk each downloaded item/album/series consumes, in familiar rounded units shown on the card and detail page, with a device total on the Downloaded surface and a reclaim amount stated at the point of removal (see [ux-flows.md §7.3.1](ux-flows.md)) | Medium | Planned |
+| UR-055 | Browse downloaded media as an offline-scoped library — reusing the same library grids, cards, and detail pages as online browsing, showing only libraries/containers with downloaded content — with the transfer-progress list demoted to a secondary "Transfers" view (see [ux-flows.md §7.2](ux-flows.md)) | High | Done |
+| UR-056 | See how much disk each downloaded item/album/series consumes, in familiar rounded units shown on the card and detail page, with a device total on the Downloaded surface and a reclaim amount stated at the point of removal (see [ux-flows.md §7.3.1](ux-flows.md)) | Medium | Done |
 | UR-057 | Settings apply the instant a control is changed — no "Save" button and no save/dirty state — so leaving the page never loses a change; sliders show a live readout while dragging but persist on release (see [ux-flows.md §8.1](ux-flows.md)) | Medium | Done |
+| UR-058 | On the home screen, a tap on a media card opens the item (movie/episode detail page, or the series Episode Focus View for episodes) rather than starting playback; a long-press starts "play now" after a confirm; an episode detail/focus page links back to its parent series and season (see [ux-flows.md §5B.5](ux-flows.md) and [§5B.1](ux-flows.md)) | Medium | Done |
+| UR-059 | Skipping to the next episode records the episode left behind as **fully watched** rather than saving a mid-episode resume point — skipping means "done with this one", not "stopped here" — and Continue Watching hides episodes the viewer has already moved past (a partial position behind that series' next-up episode), so the row only ever offers genuinely unfinished media | Medium | Done |
+| UR-060 | Search results are ordered by how well they match: a name that *starts* with the query outranks one matching mid-word (typing "parks" finds "Parks and Recreation" before "Sparks of Love"), and at equal match quality a container outranks its contents (a series before its episodes). Results are grouped into distinct categories — TV Shows, Episodes, Movies, Songs, Albums, Artists and People — so a show never competes with its own episodes for the same slot, and searching an actor's name reaches their bio | High | Done |
+| UR-061 | Double tapping the video skips within it — right half jumps **forward 30 seconds**, left half jumps **back 10 seconds** — with an on-screen indicator naming the amount. A double tap leaves the play state unchanged — playing jumps and keeps playing, paused jumps and stays paused — because the second tap re-toggles what the first tap toggled (see DR-098); the skip lands relative to the position the player actually reports, and repeated double taps accumulate rather than all skipping from the same spot | Medium | Done |
+| UR-062 | Opening a TV series lands the viewer **where they are in it**, not at season 1: the series page scrolls the current season into view and highlights the current episode, and the hero button opens that episode (labelled `Resume S2E4` / `Play S1E1`). "Current" means the episode in progress, else the server's Next Up for that series, else the first unwatched episode, else the first — resolved by the backend so it also works offline. A season is **never a page of its own**: every route that names a season lands on the series with that season in view, so the episodes of all seasons are always one continuous scrollable list | High | Done |
+| UR-063 | Each video library is **one page**, not three. Browsing (hero, Continue Watching, Next Up, Recently Added, genre rows), the full title grid, and the genre browser are tabs of `/library/tv` and `/library/movies` rather than separate routes with inconsistent names (`/library/tv/shows` vs `/library/movies/all`, `/library/shows/genres` vs `/library/movies/genres`). The old routes redirect so existing links keep working | Medium | Done |
+| UR-064 | Watch history can be **erased**, per series and per season, from the series page. Clearing marks every episode inside unwatched and clears resume positions, so the show returns to "never watched" and reopens on its premiere. It asks for confirmation first (it cannot be undone) and requires a connection to the server, since history cleared only locally would be undone by the next sync | Medium | Done |
+| UR-065 | Search answers from a **locally indexed copy of the library**, so results appear as fast as the device can query rather than at the speed of a round trip to the server, and the same results are found with the server unreachable. A background job keeps the index current — refreshing on a schedule rather than only at app start, dropping media removed from the server, and covering everything the result groups can show (including artists and people). The server is still queried in the background so media added since the last index still turns up, merged in without reordering what is already on screen | High | Implemented |
+| UR-066 | The app's own chrome stays clear of the device's system chrome. On Android the bottom navigation sits above the navigation/gesture bar instead of underneath it, the header clears the status bar, and full-screen video and audio playback keep their controls inside the usable screen — clear of the gesture bar and, in landscape, of the display notch. This must hold across navigation modes (gesture and 3-button) and rotation, not only on the handsets it happened to be tested on | High | Done |
+| UR-067 | Favourited media can be **found again**. A Favourites page lists everything favourited across all libraries, scoped by tabs (All / Movies / Shows / Music); the home screen carries favourite rows for movies, shows and music, hidden when a category is empty; and each library page can be filtered to favourites in place. Without this the like button writes to a store nothing reads | Medium | Done |
+| UR-068 | Anything the app shows can be favourited where it is shown — from a movie, series, episode, album, artist or playlist page, and from any card in a grid or carousel — not only from the player while the item happens to be playing | Medium | Done |
+| UR-069 | Favourite state agrees with the server in both directions. An item favourited in another Jellyfin client shows as favourited here without being touched, and an item favourited here while the server is unreachable reaches the server once it returns — without the user going back to the screen where they marked it | Medium | Done |
+| UR-070 | Playback quality is the viewer's choice: the player offers the bitrates the server can produce for what is playing, and changing one resumes at the same point with the same audio and subtitle tracks. Because the chosen rendition can change at any moment, nothing that streams for playback is treated as a stored copy unless it happens to be byte-identical to the real file | Medium | Proposed |
+| UR-071 | Media the viewer is watching can be **kept**, by a whole-file download that runs in the background independently of playback and at its own quality, so it is unaffected by bitrate changes. Where the streamed bytes already are that file (direct play), they are kept rather than fetched twice. A completed download is then played from disk rather than streamed again | Medium | Proposed |
+| UR-073 | Watched state is something the viewer can **set**, not only something playback records. Any episode, season, series or movie can be marked watched — or unwatched again — from where it is shown, without sitting through it or erasing its history wholesale. Marking a season or series covers the episodes inside it, and works with the server unreachable | Medium | Done |
+| UR-072 | Each page opens where a page should open. Moving to a new screen starts at the top of it, and going Back returns the viewer to the place they left — their position in a long library grid or home screen, not the top of it. A page never inherits the scroll position of the page before it | Medium | Done |
+| UR-075 | Artwork is shown at the shape it was made in. Where a screen presents a set of things side by side — the libraries on the library page and on home — they are laid out as a mosaic: rows of a common height in which each tile is as wide as its own picture, rather than a grid that crops every cover to one box. Favourites are reachable per category from that same mosaic, beside the library they belong to, not only as one undifferentiated list | Medium | Done |
+| UR-074 | Video streaming can be held to a **bandwidth budget the viewer sets**, rather than spent at whatever rate the server would otherwise send. A ceiling chosen once — from the source's own bitrate down to a rung that still plays on a poor connection — governs every video the app opens, live TV included, and survives a restart, so a metered connection is not quietly drained by the next thing played. A single video can be moved to a different ceiling from the player, resuming where it was, without disturbing that default | Medium | Done |
 
 ---
 
@@ -99,7 +117,7 @@ External system integrations and platform-specific implementations.
 | IR-017 | Jellyfin API client for transcoding parameters | API | UR-022 | Planned |
 | IR-018 | libmpv subtitle rendering and selection | Playback | UR-020 | Planned |
 | IR-019 | libmpv audio track selection | Playback | UR-021 | Planned |
-| IR-020 | libmpv/ExoPlayer equalizer integration | Playback | UR-027 | Planned |
+| IR-020 | libmpv/ExoPlayer equalizer integration | Playback | UR-027 | Done (Linux/MPV; Android parity pending) |
 | IR-022 | Jellyfin API client for person/cast data | API | UR-035, UR-036 | Done |
 | IR-023 | Database schema for person/cast caching | Storage | UR-035, UR-036 | Done |
 | IR-024 | Jellyfin API client for home screen data (featured, continue watching) | API | UR-034 | Done |
@@ -108,6 +126,9 @@ External system integrations and platform-specific implementations.
 | IR-027 | Jellyfin `/System/Info/Public` reachability probe used as an offline→online recovery detector | API | UR-043 | Done |
 | IR-028 | Jellyfin/LMS SyncGroups API client (list, create, join, unsync, dissolve sync groups) | API | UR-046 | Done |
 | IR-029 | Android `ConnectivityManager`/`NetworkCapabilities` transport probe with a `NetworkCallback` change subscription, surfaced to the frontend via the `AndroidNetworkType` JS bridge and the `jellytau-network-changed` WebView event (requires `ACCESS_NETWORK_STATE`) | Platform | UR-053 | Done (pending device verification) |
+| IR-030 | Scheduled full-catalog crawl of every library (`Recursive=true`, paged) feeding the local index, driven by a Rust background task and the `ConnectivityMonitor` reconnect signal rather than by the frontend | Storage | UR-065 | Implemented |
+| IR-031 | Android `WindowInsets` bridge: an `OnApplyWindowInsetsListener` on the decor view reports `systemBars() | displayCutout()` in CSS pixels, pushed into the WebView as `jt-inset` CSS custom properties plus a `jellytau-insets-changed` event, and pullable via the `AndroidInsets` JS bridge | Platform | UR-066 | Done (pending device verification) |
+| IR-032 | Whole-file background download of the item being played, reusing the existing resumable download worker and the Range-capable `/Videos/{id}/stream.mp4` endpoint; plus per-platform read-through caching hooks (ExoPlayer `CacheDataSource`, mpv `stream-record`) for direct-play sessions only | Storage | UR-071 | Proposed |
 
 ### 2.2 Jellyfin API Requirements
 
@@ -147,6 +168,9 @@ API endpoints and data contracts required for Jellyfin integration.
 | JA-030 | Get person details and filmography | Persons | UR-036 | Done |
 | JA-031 | Get items by person (actor/director filmography) | Items | UR-036 | Done |
 | JA-032 | Get audio-only stream URL for a video item (selected audio-stream index) | MediaInfo | UR-040 | Done |
+| JA-033 | Query favourite items (`Filters=IsFavorite`, recursive, scoped by item type) | Items | UR-067 | Done |
+| JA-034 | Read `UserData` (favourite, played, resume position) from item responses | UserData | UR-069 | Done |
+| JA-035 | Mark item played (`POST /Users/{userId}/PlayedItems/{itemId}`) | UserData | UR-025 | Done |
 
 ### 2.3 Development Requirements
 
@@ -185,11 +209,11 @@ Internal architecture, components, and application logic.
 | DR-029 | Sleep timer with roller UI, time/track/episode modes, and auto-stop (audio + video players) | Player | UR-026 | Done |
 | DR-049 | Auto-play episode limit (configurable max episodes per session) | Player | UR-023 | Done |
 | DR-050 | Reusable scroll picker (roller) component | UI | UR-026 | Done |
-| DR-030 | Equalizer UI with presets and custom bands | UI | UR-027 | Planned |
+| DR-030 | Equalizer UI with presets and custom bands | UI | UR-027 | Done |
 | DR-031 | Clickable artist/album links in now playing view | UI | UR-028 | Done |
 | DR-032 | List view option for library browsing (albums, artists) | UI | UR-029 | Done |
 | DR-033 | Genre browsing screen with quick filters | UI | UR-030 | Done |
-| DR-034 | Crossfade engine with configurable duration (0-12s) | Player | UR-031 | Done (Linux only) |
+| DR-034 | Crossfade engine with configurable duration (0-12s) | Player | UR-031 | Not implemented (blocked on MPV: single-stream audio chain; `acrossfade` needs 2 inputs — see docs/specs/playback-backend-unification.md) |
 | DR-035 | Gapless playback between sequential tracks | Player | UR-032 | Done (Linux only) |
 | DR-036 | Volume normalization with preset levels (Loud/Normal/Quiet) | Player | UR-033 | Done (Linux only) |
 | DR-037 | Remote session browser and control UI | UI | UR-010 | Done |
@@ -216,10 +240,10 @@ Internal architecture, components, and application logic.
 | DR-060 | Multi-server store and active-account selection: save/get/delete server, save/get user, set/get active user (per-server), active-session resolution | Storage | UR-047 | Partial (store done; server-switcher UI pending) |
 | DR-061 | Episode Focus View: episode hero followed *immediately* by the "More Episodes" strip — a forward-biased window (~3 before / ~6 after) around the current episode, spanning season boundaries in series order, with the current episode present and badged, per-card resume progress and watched state, and click-to-swap focus (no playback) | UI | UR-048 | Done |
 | DR-062 | Detail-page section ordering: continuation content precedes discovery content — Episode Focus View renders hero → episode strip → cast → similar; Series renders hero → seasons/episodes → cast → similar | UI | UR-048 | Done |
-| DR-063 | Search scope resolver mapping the originating route to an `includeItemTypes` set (All / Music / Movies / TV), defaulting to All for Home, `/library`, and the search tab | UI | UR-049 | Implemented |
+| DR-063 | Search scope taxonomy owned by Rust: `SearchScope` (All / Music / Movies / TV) crosses IPC as an opaque enum and `SearchScope::item_types()` expands it to Jellyfin item types, resolved once in `repository_search` before the cache and server paths diverge so online and offline filter identically; `All` expands to *no* filter rather than the union of the other scopes (which would drop People and folders). The frontend maps the originating route to a scope (`resolveSearchScope`, presentation) and never names an item type for search | Backend | UR-049 | Implemented |
 | DR-064 | Scope chip row rendered under the search bar on both the search page and the in-library header search: preselected from context, horizontally scrollable, re-runs the search preserving the query on change | UI | UR-049 | Implemented |
 | DR-065 | Thread `SearchOptions.includeItemTypes` through `library.search()` so the global/header search honours scope (backend online + offline paths already support it) | UI | UR-049 | Implemented |
-| DR-066 | Persisted search result group order with a drag-and-drop settings list, keyboard-accessible reordering, a shipped default (Songs → Albums → Artists → Movies → TV Shows), and empty-group omission | Settings | UR-050 | Implemented |
+| DR-066 | Persisted search result group order with a drag-and-drop settings list, keyboard-accessible reordering, a shipped default (see DR-091 for the current group set and order), and empty-group omission | Settings | UR-050 | Implemented |
 | DR-067 | `SearchResults` renders groups in the user-configured order rather than hardcoded markup order, without altering intra-group ranking | UI | UR-050 | Implemented |
 | DR-068 | Library card shape by media type: 1:1 square for music (circular mask for artists), 2:3 poster for movies/series/seasons, 16:9 for episodes and collection folders | UI | UR-051 | Done |
 | DR-069 | Responsive library grid (2/3/4/5/6 columns across base→xl) with two-line truncated card text and artwork-overlay progress/watched state | UI | UR-051 | Done |
@@ -227,16 +251,118 @@ Internal architecture, components, and application logic.
 | DR-075 | Shared `AccountMenu` component: identity header (user + server), Downloads / Settings / Display entries, divider, Sign out last; anchored to the username/avatar trigger and identical on desktop and mobile | UI | UR-054 | Done |
 | DR-076 | App shell exposes the header (and therefore the account menu) on every authenticated non-immersive route, including `/`, `/search`, and `/downloads`; only `/player/*` and `/login` remain chrome-free | UI | UR-054 | Done |
 | DR-077 | Display section in Settings binding the existing persisted grid/list `viewMode` store, giving the preference a discoverable home | Settings | UR-054, UR-029 | Done |
-| DR-078 | Catalog-visibility gate spanning the "Show all server media" toggle → `set_show_server_catalog` → `INCLUDE_CATALOG_BROWSE` → the synced-catalog UNION branch of offline `get_items`. Visibility resolves to `serverReachable \|\| showServerCatalog`, so offline with the toggle off lists downloaded/local media only | Storage | UR-052, UR-002 | Partial (gate implemented and unit-tested; defeated upstream by DR-079 and by the repository fallback in DR-080) |
-| DR-079 | `isConnected` derives from backend-reported server reachability alone; `navigator.onLine` is advisory and may only trigger a recheck, never force or clear the offline state (a reachable LAN server while the browser reports offline, and an unreachable server on a live link, must both resolve correctly) | Connectivity | UR-052, UR-043 | Broken (`isConnected` ANDs in `navigator.onLine`, so a live link with an unreachable server never enters offline listing) |
-| DR-080 | With the catalog-browse gate off, an empty offline `get_items` result is authoritative "no downloads here" and must be returned as-is; the hybrid repository must not treat it as a cache miss and fall through to the server | Storage | UR-052, UR-013 | Broken (`has_content()` cache-hit test in `HybridRepository::get_items`/`parallel_race` falls through to the server on an intentionally empty result) |
+| DR-078 | Catalog-visibility gate spanning the "Show all server media" toggle → `set_show_server_catalog` → `INCLUDE_CATALOG_BROWSE` → the synced-catalog UNION branch of offline `get_items`. Visibility resolves to `serverReachable \|\| showServerCatalog`, so offline with the toggle off lists downloaded/local media only | Storage | UR-052, UR-002 | Done |
+| DR-079 | `isConnected` derives from backend-reported server reachability alone; `navigator.onLine` is advisory and may only trigger a recheck, never force or clear the offline state (a reachable LAN server while the browser reports offline, and an unreachable server on a live link, must both resolve correctly) | Connectivity | UR-052, UR-043 | Done |
+| DR-080 | With the catalog-browse gate off, an empty offline `get_items` result is authoritative "no downloads here" and must be returned as-is; the hybrid repository must not treat it as a cache miss and fall through to the server | Storage | UR-052, UR-013 | Done |
 | DR-074 | WiFi-only download gate: `NetworkState`/`NetworkType` transport model reported from the platform via `set_network_state`, checked in `pump_download_queue` before starting any pending row (cellular/metered/unknown fail closed, WiFi and Ethernet require `NOT_METERED`); blocked rows stay `pending` and re-pump on network change, with a `waitingForNetwork` event driving the "Waiting for WiFi" notice. Also wires the previously inert Smart Caching / Queue Pre-caching / WiFi Only settings toggles to `CacheConfig` | Downloads | UR-053 | Done (pending device verification) |
-| DR-081 | `/downloads` split into a default **Downloaded** browse view and a secondary **Transfers** activity view, with a view switch and a Transfers badge shown only while transfers are active | UI | UR-055 | Planned |
-| DR-082 | Offline-scoped browse entry point in the repository client: browse downloaded content only (offline repository `get_items`/`get_libraries` — downloaded items plus their containers) independent of server reachability, without merging server catalog | Storage | UR-055 | Planned |
-| DR-083 | Downloaded browse reuses library grids, cards, and detail pages via the offline-scoped source; omits libraries/containers with no downloaded content; badges partially- vs fully-downloaded containers; play uses the local file; remove available at item/album/season/series level | UI | UR-055 | Planned |
-| DR-084 | Transfers view renders only in-flight rows (downloading/queued/paused/failed/waiting-for-WiFi) with Pause/Resume/Cancel/Retry; completed transfers leave the view and appear in Downloaded | UI | UR-055 | Planned |
-| DR-085 | Per-item on-disk size: stat downloaded files, aggregate to album/season/series subtotals and a device total, format in consistent rounded human units; surface size on cards and detail pages, the device total on the Downloaded surface, and a reclaim figure in the remove confirmation | Downloads | UR-056 | Planned |
+| DR-081 | `/downloads` split into a default **Downloaded** browse view and a secondary **Transfers** activity view, with a view switch and a Transfers badge shown only while transfers are active | UI | UR-055 | Done |
+| DR-082 | Offline-scoped browse entry point in the repository client: browse downloaded content only (offline repository `get_items`/`get_libraries` — downloaded items plus their containers) independent of server reachability, without merging server catalog | Storage | UR-055 | Done |
+| DR-083 | Downloaded browse reuses library grids, cards, and detail pages via the offline-scoped source; omits libraries/containers with no downloaded content; badges partially- vs fully-downloaded containers; play uses the local file; remove available at item/album/season/series level | UI | UR-055 | Done |
+| DR-084 | Transfers view renders only in-flight rows (downloading/queued/paused/failed/waiting-for-WiFi) with Pause/Resume/Cancel/Retry; completed transfers leave the view and appear in Downloaded | UI | UR-055 | Done |
+| DR-085 | Per-item on-disk size: stat downloaded files, aggregate to album/season/series subtotals and a device total, format in consistent rounded human units; surface size on cards and detail pages, the device total on the Downloaded surface, and a reclaim figure in the remove confirmation | Downloads | UR-056 | Done |
 | DR-086 | Settings page persists each control on change via per-group writers (`playerSetAudioSettings` / `playerSetVideoSettings` / `updateCacheConfig`) rather than a batch Save action; slider controls persist on `change` (pointer release) not each `input` tick; no Save button, `saving`, or `saveMessage` state | Settings | UR-057 | Done |
+| DR-087 | `MediaCard` gains an `onLongPress` prop with pointer-based long-press detection (~500 ms hold, cancelled on >10 px move so carousel scroll is unaffected, trailing click suppressed); home carousels wire tap→detail/focus routing and long-press→confirm→player; episode taps route to `/library/<seriesId>?episode=<id>`; the bare-episode detail page links to its parent series/season | UI | UR-058 | Done |
+| DR-088 | Skip-to-next-episode marks the outgoing episode played (`markAsPlayed`) instead of reporting a stop position, and arms a one-shot suppression consumed by the player's stop handler so `VideoPlayer`'s post-navigation unmount stop report cannot overwrite the 100% progress with the partial position | UI | UR-059 | Done |
+| DR-089 | Continue Watching suppresses resume entries superseded by Next Up: an in-progress episode whose series has a next-up entry strictly later in series order (season, then episode) is dropped from the Home and TV rows; movies, series without a next-up entry, and items with unknown/mixed episode ordering are always kept | UI | UR-059 | Done |
+| DR-090 | Relevance ranking in Rust (`domain/search_rank.rs`): results sort by match position (prefix → word-start → mid-word substring → no name match) then by media kind (containers before their contents), stably so the backend's own relevance breaks ties. Applied in `repository_search` to both the instant cache result and the merged cache+server union, so the list does not reshuffle when server results land | Backend | UR-060 | Done |
+| DR-091 | Search result groups split TV into separate Shows and Episodes groups and add a People group (default order: Shows → Episodes → Movies → Songs → Albums → Artists → People); a stored `tvShows` order from before the split expands in place to shows+episodes so an upgrading user keeps their arrangement | UI | UR-060 | Done |
+| DR-092 | Video tap gestures resolve in `tapGestures.ts` (pure, unit-tested) rather than inline in `VideoPlayer.svelte`: `registerTap` classifies each tap and the component acts on it immediately — `togglePlayPause` for a first tap, or `seek` (+30 s right / −10 s left) plus a re-toggle for a second tap inside `DOUBLE_TAP_WINDOW_MS` (300 ms). A consumed pair resets the state, and a swipe forgets the tap. The deferral this originally used was removed in DR-098, which also covers suppressing the compatibility `click` the browser synthesizes after a touch tap. `resolveSeekTarget` converts the delta to the absolute position the facade requires, clamped per DR-095 and chained off a still-in-flight `pendingSeekTarget` so back-to-back skips accumulate instead of all resolving against a not-yet-updated position | UI | UR-061 | Done |
+| DR-094 | Frontend boundary tripwire (`scripts/check-frontend-boundary.sh`) detects Jellyfin item-type array literals **anywhere** in `src/` rather than only inline at an `includeItemTypes:` query site, so a category→type mapping cannot evade the check by being assigned to a named const (the evasion that let the `scoped-search` leak pass CI); requires two adjacent type literals so single-type presentation and `item.type ===` inspection stay legal, and caps the allowlist to force taxonomy into Rust instead of accumulating exceptions | Tooling | - | Done |
+| DR-098 | Video tap gestures act **immediately** — no deferral, no timer, and only first/second taps exist. A first tap toggles play/pause; a second tap inside `DOUBLE_TAP_WINDOW_MS` seeks *and* toggles again, so the two toggles cancel and a double tap preserves the play state (playing → jump and keep playing; paused → jump and stay paused). This replaces a design that deferred the first tap behind a 300 ms timer so a second tap could cancel it: the timer cleared its own handle *before* invoking the toggle, which reopened the `tapTimeout !== null` guard in `handleVideoClick` meant to suppress the compatibility `click` Android's WebView synthesizes after a touch — the late click then toggled a second time, producing a pause/unpause loop (long-press was unaffected, which is what identified the tap path). Click suppression no longer depends on the timer: `handleVideoClick` ignores `detail === 0` *and* any click within `TOUCH_CLICK_SUPPRESS_MS` of a touch tap. A swipe undoes the touchstart toggle exactly once (latched on `swipeGestureActive`) so brightness swipes never change play state. Click suppression is shared by **every** click target layered over the video via `isSynthesizedTouchClick`, not just the `<video>`: pausing renders a full-screen play-overlay button, so the synthesized click lands on *that* and an unguarded handler there resumed immediately — pausing appeared impossible while unpausing worked, because unpausing removes the overlay | UI | UR-061 | Done |
+| DR-099 | The video seek bar is usable by touch. Two Android-only defects made dragging or tapping it move the thumb without moving playback. (a) *Gesture hijack*: the container-level gesture layer skips `touchstart` on a control (DR-098) but kept handling `touchmove`, so a seek-bar drag was measured against the **previous** gesture's start point — a huge bogus vertical delta that read as a brightness swipe, dimmed the screen to the 0.3 floor, and fired a spurious play/pause "correction" mid-drag. A gesture is now latched at `touchstart` (`playerGestureActive`) and `touchmove` ignores anything not latched, since re-checking the move target cannot recover a start point that was never recorded. (b) *Commit signal*: the seek was committed **only** from `change`, which Android's WebView does not reliably fire for a touch interaction on a range input — the thumb moved to the tapped position and no seek ever ran. `touchend`/`mouseup` now commit as well; `input` arms a one-shot latch so whichever release signal arrives first commits and the other is a no-op. `seekRelative` shares the same `commitSeek` entry point instead of fabricating a synthetic `change` event | UI | UR-005, UR-061 | Done |
+| DR-097 | Transport authority (play/pause/toggle) lives in Rust for **webview-rendered** media, not just native. The controller tracks the state the HTML5 element reports (`html5_playing`, fed by `report_html5_state`, which now *stores* rather than only re-emitting); `play`/`pause`/`toggle_playback` consult it and drive the element by emitting a `ControlCommand` that `playerEvents.handleControlCommand` executes against the active adapter. A `stopped`/`idle` report clears it so the native backend (MPV/ExoPlayer) regains authority for music. The frontend facade no longer short-circuits transport into the adapter: `adapter.toggle()` previously decided play-vs-pause by reading `el.paused` off the DOM, a value that flips transiently while an element buffers or settles a seek — so two intents ~150 ms apart read *different* values, performed *opposing* actions, and self-sustained a play/pause loop needing no further input (observed on Android with a fully-buffered `readyState=4 networkState=1` element). Same "backend decides, adapter executes the primitive" split as `player_seek_video` | Player | UR-005 | Done |
+| DR-096 | `Html5PlayerAdapter.play()` is resilient to stall recovery: an in-flight attempt is memoised so concurrent callers (UI plus hls.js gap-controller recovery) share one `element.play()` instead of stacking calls, and an `AbortError` ("play() request was interrupted by a call to pause()") is logged at debug rather than pushed to `host.onError`. The browser raises it whenever a pending play promise is superseded by a pause/seek/source change, which hls.js does routinely while nudging past a stall — reporting it surfaced a player error roughly once per second for the whole stall and left the UI stuck showing paused | Player | UR-005 | Done |
+| DR-095 | Seek targets clamp strictly *inside* the media (`clampSeekTarget`, `END_SEEK_MARGIN_SECONDS` = 6 s ≈ one HLS segment) instead of to the exact `duration`. Landing on the duration makes hls.js request the segment whose start time lies past the end of the media (e.g. a 6330.324 s item → segment 1055 starting at 6336.33 s), which Jellyfin never produces; the fetch times out and hls.js' gap-controller stalls at the last buffered position, presenting as "unpausing or skipping bounces straight back to paused". Applied on both seek paths — the relative-skip `resolveSeekTarget` and the seek-bar drag, whose range input `max` is the duration itself — and floored at 0 so media shorter than the margin still seeks to the start | UI | UR-061 | Done |
+| DR-100 | Leaving a video and re-entering it renders the **video** player, never the audio one. Both halves of the `/player/[id]` decision are pure and unit-tested in `playerSurface.ts`. (a) `shouldReuseActivePlayback` excludes video: the "already playing, just show the UI" shortcut (added for expanding the audio mini player) returns *before* a stream URL is fetched, which is fine for audio — the backend owns the stream and the route only mirrors it — but leaves `<VideoPlayer>` with nothing to render. Closing a webview-rendered video deliberately emits no `stopped` state (that would break the autoplay handoff, see DR-047), so the Rust controller still reports that movie/episode as its loaded media and re-entering the same item hit the shortcut. (b) `resolvePlayerSurface` maps video-without-a-stream-URL to `pending` (spinner) instead of falling through to `<AudioPlayer>`, so no future path can put video content in the audio surface. Video now always takes the full load path, which fetches the stream URL and applies the stored resume position | UI | UR-005 | Done |
+| DR-101 | "Where is this viewer in this series" is resolved in **Rust**, not the frontend. `repository_get_series_episodes` performs the season fan-out (`get_items(series_id)` → seasons → `get_items(season_id)`, plus the flat-series fallback for shows whose children are episodes rather than season folders) and returns them in series order — season index ascending, episode index ascending, specials (season 0) after every numbered season. `repository_get_series_current_episode` layers the pure policy `pick_current_episode` over that list: an **in-progress** episode wins (earliest in series order on a tie — it is literally where playback stopped, and Next Up would skip past it), then the server's **Next Up** for that series, then the **first unwatched** episode, then the first. The third rung is the offline path, not dead code: `OfflineRepository::get_next_up_episodes` returns an empty vec, so without it the feature would be online-only. A failing Next Up or resume lookup degrades to empty rather than failing the call. `repository_get_next_up_episodes` had accepted a `series_id` since it was written and **no caller had ever passed one** | Repository | UR-062 | Done |
+| DR-102 | The series detail page anchors on that answer. It calls `repositoryGetSeriesEpisodes` once instead of fanning out over seasons in TypeScript (the fan-out *and* its flat-series fallback were domain knowledge in the presentation layer), groups the returned episodes under season headers by `parentIndexNumber`, and passes the resolved current episode to `SeasonSection` → `EpisodeRow`, which renders a highlight ring and scrolls itself into view. The hero button navigates to `/library/<seriesId>?episode=<currentId>` — the Episode Focus View, where an explicit Play/Resume commits — per ux-flows §5B.5: Play on a *container* is navigation, Play on a *leaf* commits. It previously resolved `$libraryItems[0]`, the first **season** by `SortName`, and navigated to `/player/<seasonId>`, which the player route bounced back to `/library/<seasonId>` — so Play on a series played nothing and landed on the season-1 page | UI | UR-062 | Done |
+| DR-103 | A season is not a destination. `/library/<seasonId>` redirects to `/library/<seriesId>#season-<indexNumber>`, the anchor `SeasonSection` renders, so a season link scrolls the series' continuous episode list rather than opening a page. Every inbound link follows: the episode breadcrumb, `handleItemClick case "season"`, the TV landing page's `case "Season"`, and `DownloadedBrowse`. A season carrying no `seriesId` (deep link into a stale cache) still renders the generic view so the user is never stranded. This removes a surface that had no route of its own — it fell through the detail page's `kind` chain to the generic "Contents" poster grid, contradicting ux-flows §5A.2 (episodes must be a row list), and clicking an episode there opened a bare Episode page, which §5B.1 forbids | UI | UR-062 | Done |
+| DR-104 | The "More Episodes" strip spans the **whole series** in series order, per ux-flows §5B.2's cross-season continuity rule: at the end of a season the window runs on into the next season's first episodes instead of dead-ending. `adjacentEpisodes` previously filtered the pool to `parentIndexNumber === current.parentIndexNumber` and sorted by `indexNumber` alone, so the window could never leave the current season — and, when episodes of several seasons did reach it, sorting by episode number alone interleaved them. Cards crossing a season boundary are labelled `SxEy` rather than a bare episode number so the jump is legible | UI | UR-062 | Done |
+| DR-105 | Video library routes collapse to one per library. `/library/tv` and `/library/movies` render browse / all-titles / genres as in-page tabs driven by `?view=`, omitted for the default `browse` (the convention `searchRouteUrl` already uses for the `all` scope); `resolveLibraryView` is pure and unit-tested. The four legacy routes become redirect-only `+page.ts` loads rather than deletions, because `GenreTags` links to them and users have them in history; `resolveSearchScope` keeps its `/library/shows` branch for the same reason. The "Browse" tile grid at the bottom of both landing pages is removed — it was a second navigation affordance to the same destinations the carousels' "Show all" links already reach | UI | UR-063 | Done |
+| DR-106 | Erasing watch history goes through the repository, not the local cache: `clear_watch_history(item_id)` maps to Jellyfin's `DELETE /Users/{userId}/PlayedItems/{itemId}`, which clears the played flag *and* zeroes the resume position, and which the server applies recursively to a folder — so one call handles a whole series or season. `OfflineRepository` returns `RepoError::Offline` rather than clearing locally, because history diverged only on the device would be silently undone by the next sync; the button disables itself while the server is unreachable. `ClearHistoryButton` is shared by the series hero and each `SeasonSection` header, confirms before acting (there is no undo), and reloads the page on success so the recomputed current episode — the premiere, for a fully cleared series — is what the viewer sees | Repository | UR-064 | Done |
+| DR-107 | Seasons on the series page are collapsible, and **only the current season is expanded** on load — the one holding the episode DR-101 resolved. A show with ten seasons otherwise renders every episode of every season at once, burying the one episode the viewer came for under hundreds of rows. Expansion state is per season and pure (`initialExpandedSeasons` in `seriesNavigation.ts`): the current season, or the first season when there is no current episode, so a never-watched show still opens on season 1 rather than fully collapsed. A `?episode=` deep link expands that episode's season too. Toggling is local and not persisted — it is a reading position, not a preference | UI | UR-062 | Done |
+| DR-108 | The instant (cache) leg of `repository_search` searches the **synced catalog**, not just downloads. `OfflineRepository::search` replaces its `downloaded_items` CTE with the `available_items` CTE `get_items` already uses — the same downloads branches plus a `synced_at IS NOT NULL` branch gated on the same `include_catalog_browse()` flag — so search and browse cannot diverge on what is visible. Online (flag true) search reads the whole index and answers before any HTTP request completes; offline with "Show all server media" off (flag false) it stays downloads-only, unchanged. Requires no frontend change, since the flag is already set correctly for all three states. The `include_item_types` filter is switched from string interpolation to bound parameters, as `SearchOptions` is settable from the frontend and not only from `SearchScope` | Backend | UR-065 | Implemented |
+| DR-109 | Index freshness is a Rust-owned policy, not a frontend startup call. A tokio task ticks every 30 min and runs a full pass when a repository is active, the server is reachable, and `last_catalog_sync` (already persisted to `app_settings`, previously read only for a UI hint) is older than `CATALOG_INDEX_TTL` (6 h); the `ConnectivityMonitor` reconnect signal re-evaluates the same condition immediately. An `AtomicBool` prevents concurrent passes, replacing `offlineCatalog.ts`'s `syncInProgress` — the frontend trigger is removed rather than left alongside, since two triggers with one guard each is how double-crawls happen. `RepositoryManager` gains an active-handle slot so the task has something to run against. Progress is emitted as the kebab-case `catalog-index-event` | Backend | UR-065 | Implemented |
+| DR-110 | Index hygiene. `save_to_cache` switches from `INSERT OR REPLACE INTO items` to `ON CONFLICT(id) DO UPDATE`: REPLACE fires no `AFTER DELETE` trigger unless `recursive_triggers` is on (it is not — only `foreign_keys` and `journal_mode` are set), so `items_ad` never ran, and because `items.id` is a `TEXT PRIMARY KEY` each replacement also took a fresh rowid and appended a second `items_fts` entry — a duplicate index per sync, invisible in results but permanently degrading `MATCH`. The upsert preserves the rowid `items_fts` keys on and fires `items_au`; migration `021_rebuild_items_fts` clears orphans on existing installs. Separately, a post-crawl sweep deletes synced-but-not-downloaded rows a successful library crawl did not return, so media removed from the server stops being searchable; it skips items with completed downloads and skips any library whose crawl errored, because `items.parent_id` is `ON DELETE CASCADE` and a partial crawl would cascade away a whole series | Storage | UR-065 | Implemented |
+| DR-111 | The index covers what the result groups render: `CATALOG_ITEM_TYPES` gains `MusicArtist` and `Playlist`, and migration `022_people_fts` adds a `people_fts` virtual table over the existing `people` table (which had no FTS, and is populated incidentally by item-detail fetches) with the same trigger pattern as `items_fts`. `OfflineRepository::search` UNIONs `people_fts` matches in as `Person` items when the resolved scope admits them — i.e. `SearchScope::All`, which expands to no filter (DR-063). Without this, the Artists and People groups UR-060 mandates can only ever be filled by the server leg | Storage | UR-065, UR-060 | Implemented |
+| DR-112 | Safe-area insets come from **native**, not from `env()` alone. `env(safe-area-inset-*)` is 0px without `viewport-fit=cover` (missing from `app.html`, so every safe-area rule in the app was already a no-op), and even with it Android WebView maps only the *display cutout* — never the status bar or navigation bar. Since `enableEdgeToEdge()` plus `targetSdk 36` make edge-to-edge unconditional, the WebView always spans the system bars, so CSS could not learn about them by any route. `WindowInsetsBridge` reads the real insets and publishes `jt-inset` custom properties; `app.css` folds them with `env()` via `max()` into `--safe-*`, which is the only thing components may pad from. Ownership is exactly one element per edge: the app shell takes top/left/right, and BottomUi takes bottom wherever it renders (`shellReservesBottomInset` hands it back to the shell on routes with no bottom UI) so the padding sits inside BottomUi's surface box and the colour extends behind the gesture bar. The full-screen players inset their control layers only, leaving video and artwork edge-to-edge. The theme's `fitsSystemWindows=true` — which claimed the opposite and was overridden at runtime and ignored at this target SDK — is removed | UI | UR-066 | Done |
+| DR-113 | `MediaItem.user_data` is populated from the server instead of being hardcoded `None`. `JellyfinItem` gains a `UserData` field (`#[serde(alias = "UserData")]` → the existing `UserData` type) and `to_media_item` maps it, so every list and detail response carries favourite/played/resume state. `UserData` is named explicitly in the `Fields=` list rather than relying on Jellyfin's default. Without this no card or detail page can render a favourite it did not itself set, and the mini player's per-track `storageGetPlaybackProgress` fetch is the only way to colour one heart | Repository | UR-069 | Done |
+| DR-114 | Server favourite state is mirrored into the local `user_data` table by `OfflineRepository::save_to_cache` — the single choke point every cached server result passes through — so offline browsing and the offline Favourites page see the same favourites as the server. The upsert carries `pending_sync = 0` and is guarded by `WHERE user_data.pending_sync = 0`, which is the conflict rule: a toggle made offline is never overwritten by a stale server value before it has been pushed | Storage | UR-069 | Done |
+| DR-115 | Cross-library favourites query: a `get_favorites(scope, options)` repository method plus the `repository_get_favorites` command. Online issues `Filters=IsFavorite&Recursive=true` with `IncludeItemTypes` expanded from `SearchScope::item_types()` in Rust (the frontend sends the opaque scope, never a type list — DR-063); offline reads `items ⨝ user_data (is_favorite = 1)` under the same `include_catalog_browse()` gate as browsing; hybrid races cache against server like `get_items` — saving server results through to the cache on a miss, so the favourites page does not re-query the server every visit and the DR-114 mirror is filled on a fresh install — and applies the DR-080 rule that an empty offline result is authoritative when the gate is off. The command falls back to this read when nothing is cached, rather than painting an empty state it will correct a round trip later. A separate method rather than `get_items` because favourites span libraries and `get_items` is `ParentId`-shaped | Repository | UR-067 | Done |
+| DR-116 | `GetItemsOptions.favorites_only` filters an existing library listing in place — online by appending `Filters=IsFavorite`, offline by joining `user_data` into the existing `available_items` CTE so the downloads-only gate still applies. This is what backs the per-library favourites toggle, and composes with the genre and item-type filters already there | Repository | UR-067 | Done |
+| DR-117 | The Favourites page (`/library/favorites`) renders favourites across libraries with All / Movies / Shows / Music scope tabs, reusing `LibraryViewTabs` + `LibraryGrid` + `MediaCard` so card shape still follows the media (§5A.1) and a mixed All tab reads as posters, squares and thumbnails side by side. Each tab sends a `SearchScope` value and nothing else. Reached from the library overview and from "See all" on the home rows | UI | UR-067 | Done |
+| DR-118 | Home carries favourite rows for movies, shows and music, loaded via `repository_get_favorites` per scope and rendered below Recently Added. A row with no items does not render at all, so a fresh install shows no empty favourite rows | UI | UR-067 | Done |
+| DR-119 | `FavoriteButton` is mounted wherever a whole item is shown — movie/series/episode detail heroes, album/artist/playlist headers, and as a `MediaCard` artwork overlay — and a `favorites` store holds in-session optimistic state so un-hearting on one surface updates every other without a refetch. Resolution order is `store override ?? item.userData?.isFavorite ?? false`. On a card the heart is its own button and stops propagation, so hearting never also opens, plays, or triggers the §5B.5 long-press; it is suppressed on server-only (greyed) cards | UI | UR-068 | Done |
+| DR-120 | Favourite toggles made while offline reach the server. A Rust drain, triggered by the `ConnectivityMonitor` offline→online transition, pushes every `user_data` row with `pending_sync = 1` and clears the flag on success, leaving failures pending for the next transition. It lives in Rust rather than the frontend because a frontend drain dies with the component that started it. Both the drain and the hybrid background refresh emit the kebab-case `favorites-changed` event (`{ itemIds }`) so open views update — without it a favourite marked on another client appears only on the *second* visit to a page, since the cache-first read returns local rows and the server refresh is invisible to the frontend. Supersedes the unused `syncService.queueFavorite`, which is deleted rather than left as a second queue | Backend | UR-069 | Done |
+| DR-121 | Player quality selector: Rust reports the bitrates available for the current media source and owns the quality→transcode-parameter mapping (the one `get_video_download_url` already holds — playback calls into it rather than restating it, or the two tables drift). Changing quality re-negotiates the stream URL and resumes at the current position with audio/subtitle selection preserved. On Linux, video re-negotiates *within* HLS: returning `stream.mp4` is the documented cause of transcoded playback never starting. The frontend renders the list and remembers the choice; it does not decide what the choice resolves to | UI | UR-070 | Proposed |
+| DR-122 | The playback path is ephemeral. Streamed bytes are never persisted unless DR-124 rules them keepable, and any in-flight capture is abandoned — partial file deleted, never promoted — the moment the viewer changes quality, because a capture spanning a rendition change is a splice of two encodings rather than a playable file | Playback | UR-070 | Proposed |
+| DR-123 | The download path is independent of playback: a whole-file fetch through the existing download manager at one canonical quality (default `original`, the direct static copy) over the Range-capable `/Videos/{id}/stream.mp4`, unaffected by bitrate changes and completing into an ordinary `downloads` row so offline browsing and `refresh_queue_local_sources` pick it up unchanged. Prerequisite: downloaded video is currently never played locally — `repository_get_video_stream_url` goes straight to the online repo and the player route calls it with no local check, so a completed video download is still streamed. Without that fix nothing in this spec is observable for video | Repository | UR-071 | In Progress |
+| DR-124 | Streamed bytes are kept only where they *are* the download artifact — a direct-play session. Android uses ExoPlayer `SimpleCache`/`CacheDataSource` keyed by item **and** media-source id so renditions cannot collide, sharing the existing smart-cache storage budget rather than opening a second one over the same disk; Linux audio uses mpv `stream-record`, abandoned on seek because it is documented as intended for linear streams and seeking breaks the recording. Transcoded Linux video is **not** captured: HLS segments are not a file, and assembling one needs ffmpeg, which is not a dependency and which CI may not install at job time — DR-123 covers that case instead | Playback | UR-071 | Proposed |
+| DR-125 | A capture is promoted to a completed `downloads` row only when it covers the whole resource; partials stay evictable cache. A new `downloads.source_rendition` column records the negotiated quality/container/codec (`NULL` for the existing paths, which are always `original`) so a captured transcode and a real download are distinguishable rows and an "upgrade to original" remains possible. A quality change never touches a file that already exists — not a permanent download, and not a completed temporary one, both of which stay valid copies of the rendition they hold. It invalidates only an **in-flight** capture or background download of cached media, which is abandoned and restarted at the newly chosen quality, because a capture spanning a rendition change is a splice of two encodings rather than a playable file | Storage | UR-071 | Proposed |
+| DR-126 | Cache eviction only reclaims the *temporary* tier. `evict_lru_async` selected every completed download ordered by `completed_at ASC` with no `download_source` filter, so hitting the 10 GB storage limit deleted the **oldest** download — typically a film saved deliberately for offline — to make room for a newly precached track. It now evicts only `COALESCE(download_source, 'user') = 'auto'` rows; `COALESCE` rather than a bare equality because rows predating migration 012 can be NULL and unknown provenance must be treated as the user's, never as disposable. Freeing less than requested is the correct outcome when only user downloads remain — the caller reports "unable to free enough space" instead of silently deleting them | Storage | UR-071 | Done |
+| DR-127 | A cache entry *is* a download with a shorter life: same `downloads` row and same file handling, distinguished by `download_source = 'auto'` plus an expiry, so there is one storage model rather than a cache and a download library that can disagree. Temporary rows are reclaimed on whichever comes first — the life limit elapsing, or eviction under space pressure (DR-126). Permanent (`'user'`) rows have no expiry. A temporary row can be promoted to permanent by the user choosing to keep it, which only clears the expiry and flips the source; the bytes never move | Storage | UR-071 | Done |
+| DR-128 | Audio-only playback of *downloaded* media reads the local file rather than fetching an audio-only stream. No transcode is involved or wanted: the Linux backend already runs MPV with `video: no`, so handing it the downloaded video file decodes the audio track and ignores the video, and ExoPlayer disables its video renderer equivalently. Transcoding to a separate audio artifact would cost CPU and battery, need an encoder the project does not ship, and produce a second file to keep in step — for no gain over simply not decoding the video | Playback | UR-071 | Done |
+| DR-129 | A stream that stops delivering is recovered, not treated as terminal. Two failure shapes, because the streams differ. (a) *Phantom end* — the background audio-only handoff uses a progressive mp3 transcode over plain HTTP, chunked and therefore length-less, so a dropped connection reaches the player as end-of-input and ExoPlayer reports `STATE_ENDED` indistinguishably from the real end. The item's runtime is the only thing that can tell them apart: an end reported more than a tolerance short of it (comparing the *absolute* position — handoff base plus the player's relative position) is a truncation. Left unhandled, playback parked in `STATE_ENDED` and the next play intent from the lockscreen, notification or a Bluetooth reconnect seeks an ended player to position 0 — the user-visible "the episode randomly restarted". (b) *Recoverable error* — music (`/Audio/{id}/stream?Static=true`) and video (`/Videos/{id}/master.m3u8`) declare their length, so the player detects the truncation itself and raises an error; the frontend's handler stopped playback outright, turning a hiccup into silence. Both resume the current item **in place** (never via `play_item`, which would replace the queue with a single item and lose the album), the error path after a per-attempt backoff. Seekable streams are re-prepared at the URL they already have and seeked; the length-less transcode, which cannot be seeked, has `StartTimeTicks` rewritten into its existing URL so the user's audio-track selection survives and recovery needs no network round-trip. Only `Remote` sources qualify — a local file cannot fail from the network. A shared budget of consecutive attempts at the same position, refilled whenever playback progresses, stops an unreachable server from looping | Playback | UR-040, UR-004 | Done |
+| DR-130 | A backend's position and duration must survive the end of the file they describe. MPV exposes `time-pos`/`duration` as properties of the *loaded* file, so at EOF it unloads and both stop resolving — the accessors reported `0.0`/unknown at exactly the moment end-of-file handling asks where playback reached, and any position-versus-runtime check would have read every natural end as a truncation. The poll thread records the last reading and the accessors fall back to it. Linux resilience is layered on the same principle that the stream, not the player, is what failed: MPV is configured with ffmpeg reconnection (`stream-lavf-o`, `network-timeout`) so ordinary blips never surface, and `EndFile(ERROR)` — previously a bare log, which left playback halted while the UI still showed "playing" — is emitted as a *recoverable* error. Because MpvBackend is constructed before `PlayerController` exists, it cannot decide in-process like the Android JNI callback: the frontend echoes the error into `player_recover_stream`, which keeps the decision in Rust (the same shape as `PlaybackEnded` → `player_on_playback_ended`). Android reports errors it has already declined as *unrecoverable*, so the echo never asks twice | Playback | UR-004, UR-040 | Done |
+| DR-131 | The offline mutation queue is drained. `sync_queue` had producers and no consumer: `PlaybackReporter::queue_for_sync` writes a row for every start/stop/mark-played that cannot reach the server, `sync_mark_processing`/`_completed`/`_failed` were registered commands with no callers, and no Rust task processed the table — so queued watch positions never reached Jellyfin and the offline banner's count only ever grew. A drain hangs off the same `connectivity:reconnected` transition as DR-120 (in Rust, because a drain started by a component dies with it) and replays rows oldest-first, so a stale start cannot move the server's resume position backwards after a later stop. `update_progress` replays as *stopped at N* rather than as progress — replaying a mid-playback report hours later would claim the item is still playing — and payloads are read in both dialects that exist in users' databases (`position_ticks` from Rust, camelCase `positionMs` from the frontend helper). A failed row stays queued for the next reconnect; after `MAX_SYNC_ATTEMPTS` it is `abandoned` and stops counting, because a row nothing can ever push is what turns the queue into a counter that only grows. An *unreachable* server is not counted as an attempt at all — the row goes back to `pending` untouched — so opening the app offline a few times cannot abandon good rows; only a server that answers and refuses spends the budget. The drain also runs once at startup, because a queue built in a previous session would otherwise sit untouched for a whole run whenever the server was reachable the entire time and no offline→online transition ever fired. Requires `MediaRepository::mark_played` (JA-035) — the previous stand-in reported a stop at `i64::MAX` | Backend | UR-025, UR-002 | Done |
+| DR-132 | The pending-sync count is answerable. The offline banner's badge read "N pending sync(s)" and led nowhere, so it was taken for pending *transfers* and looked for on the Downloads page — which lists the `downloads` table and structurally cannot show `sync_queue` rows. The badge becomes a button opening the queue it counts: each row's operation, the item's title (resolved by a `LEFT JOIN items` in `sync_get_pending`, not a per-row frontend fetch), when it was queued, and the error of anything failing, plus a "Sync now" that runs the DR-131 drain on demand. The same list is a Settings section, because a row that keeps failing is still queued when the server is reachable and no banner is on screen. The drain emits `sync-queue-changed` so the badge updates on reconnect instead of lagging by up to one 10s poll | UI | UR-025 | Done |
+| DR-133 | A downloaded file has exactly one on-disk path, and the row that names it is authoritative. `downloads.file_path` starts relative to the storage root, but the worker rewrites it to the absolute path it actually wrote when the transfer completes — so a *completed* row is already rooted. The video player's offline branch rooted it a second time, handing the asset protocol `/data/user/0/app//data/user/0/app/videos/x.mp4`; the webview reported `MEDIA_ERR_SRC_NOT_SUPPORTED` with `NETWORK_NO_SOURCE`, so every downloaded video failed to play while audio — which resolves the same column through Rust's `resolve_local_media_path`, without re-rooting — played fine. The join is absolute-aware (POSIX, Windows drive letters and UNC) so rows written before completion still resolve | Playback | UR-071 | Done |
+| DR-134 | The webview can actually fetch the local files it is handed. `convertFileSrc` rewrites a path to `http://asset.localhost/…` unconditionally, but Tauri only answers that origin when the `protocol-asset` cargo feature is compiled in *and* `app.security.assetProtocol.enable` is set — neither was, so every such URL reached a protocol with no handler and the webview reported `NETWORK_NO_SOURCE`. This silently defeated both offline video (`<video src>`) and the cached-thumbnail path in `imageCache`, which fails soft to the server copy and so hid the breakage whenever the server was reachable. The scope is `$APPDATA/**` — the storage root under which the database, `downloads/` and the thumbnail cache all live — rather than an unrestricted grant, so the webview can read the app's own media and nothing else | Security | UR-071 | Done |
+| DR-140 | An audio track is pinned only when the user picked one. Jellyfin's `MediaStream.Index` is global across every stream in a media source, so index 0 is the *video* stream on virtually all files — yet `AudioStreamIndex=0` was sent as "the first audio track" on the HLS transcode URL, the background audio-only handoff URL, the direct-play fallback URL, and the `PlaybackInfo` negotiation body. A server that honours the request literally then transcodes the video stream into the audio slot and the result plays as a picture with no sound; only servers that silently correct the index hid the bug, which is why it presented as "some videos have no audio". The parameter is now omitted whenever no track has been chosen, so the server resolves the source's `DefaultAudioStreamIndex`; an explicit selection from `player_switch_audio_track` is still carried through unchanged. On the `static=true` direct-play URL it is dropped outright — the original file is served untouched, so the parameter could only mislead | Playback | UR-004, UR-040 | Done |
+| DR-147 | One search input per screen, and the URL is the search's single source of truth. The header bar rendered only under `/library/**` and merely *navigated* to `/search` (DR-063), so a desktop search handed the user to a screen whose input was a different element — the header box cleared itself and vanished, and the page's own box took over mid-word. That page then re-derived its input from `?q=` against `library.searchQuery` on every store write, so the next keystroke re-ran the effect and snapped the text back to the query the header had sent (and a scope chip back to the URL's scope); entering from the bottom-nav Search tab skipped it only because the effect early-returned on an empty query. The bar now renders on `/search` too (`showHeaderSearch`) and is the sole md+ input — the page's own input is `md:hidden` — and on that route it republishes the query into the URL with `replaceState`, so a whole session of typing costs one history entry. The page *consumes* that URL once per distinct value (`seedFromSearchUrl` against a non-reactive `applied` marker) instead of continuously reconciling it, and the scope chips publish through the same URL so the bar and the chips cannot disagree. Landing on `/search` with a seeded query focuses the bar and puts the caret at the end, because the box the user was typing in belonged to the unmounted route | UI | UR-049, UR-054 | Done |
+| DR-142 | An episode has exactly **one** surface, and it is complete. Two divergent renderings existed: `EpisodeFocusView` (reached from Continue Watching, the series episode list, the TV landing page and Downloads — i.e. every real entry point) offered only Play and Favourite, while the bare `/library/<episodeId>` page nobody routed to carried the download button, the series/season breadcrumbs and the cast section. Opening an episode the normal way therefore silently lost the ability to download it. The Focus View is now the single surface and carries the full §5B.2 composition — hero action row `Play / Download / Favourite`, series name and `SxEy` badge as links back to the series and to that season's anchor, then genres → cast → similar shows *below* the episode strip, never above it (DR-062). `/library/<episodeId>` redirects into it (`episodeRedirectTarget`, the same rule seasons follow under DR-103), and an episode with no `seriesId` renders the same component series-less rather than falling back to a second, lesser page. The focused episode is fetched in full rather than reused from the season fan-out, because that is a *list* query and carries neither cast nor genres — the sections would have rendered empty. The strip hides itself when the episode has no siblings, a card that only shows the episode you are already on being noise | UI | UR-048, UR-058 | Done |
+| DR-141 | The device profile states how many channels the audio route can actually voice. `MediaCodecList` answers "can this device *decode* 5.1", which is not the question that decides whether the user hears anything — a phone decodes an AC-3 5.1 track happily and still has two channels to play it out of. With no `MaxAudioChannels` in the profile, Jellyfin was free to direct-play the multichannel track, and the result is device dependent: a failed `AudioSink` configuration (silence) or dialogue folded into surround channels that go nowhere. media3's `AudioCapabilities.maxChannelCount` for the current route is reported over JNI alongside the codec lists, and bounds both the direct-play profile and the transcoding profiles, so the server downmixes rather than shipping channels the sink cannot take. Codecs are never removed from the profile — a device with genuine surround output keeps direct-playing it. A missing or zero reading means "route not yet established", not "no audio", and falls back to stereo, the one capability every sink has | Playback | UR-004 | Done |
+| DR-145 | Video playback starts only once the app actually holds audio focus. Video manages focus by hand (`handleAudioFocus=false`, because ExoPlayer's automatic handling is reserved for the audio path), and the request's three outcomes were all treated as success: `AUDIOFOCUS_REQUEST_DELAYED` — which `setAcceptsDelayedFocusGain(true)` explicitly invites, and which means the system is *withholding our audio* until it calls back — and an outright `REQUEST_FAILED` were logged and then followed by `playWhenReady = true`. The picture rolled with no sound, indistinguishable to the user from a broken stream. Playback is now held when focus is not granted and started from the `AUDIOFOCUS_GAIN` callback; an explicit `play()` re-requests focus rather than resuming into a stream the system is still muting, guarded by a held-focus flag so repeated plays do not leak focus requests. A `LOSS` clears the pending flag, so an unrelated later `GAIN` cannot start playback the user never asked for | Playback | UR-004 | Done |
+| DR-146 | The no-audio-track fallback picks a track the renderer can actually play. When ExoPlayer selected no audio track, the recovery forced group 0 / track 0 unconditionally — but the most likely reason nothing was selected is that this very track cannot be decoded on this device, so the override reinstated the silence it was meant to fix. It now scans the groups for the first `isTrackSupported` track and overrides to that, and clears `setTrackTypeDisabled(TRACK_TYPE_AUDIO)` because audio may equally have been off at the type level, which an override alone does not undo. When no group holds a supported track the condition is logged as an error — the server was expected to transcode — rather than leaving a silent video with no explanation in the log | Playback | UR-004 | Done |
+| DR-148 | The video direct-play profile advertises only what the **webview** can decode. The audio codec list comes from `MediaCodecList`, which describes ExoPlayer — but video does not play through ExoPlayer on either platform: Android force-renders every video in the webview `<video>` element (the interim override in `VideoPlayer.svelte`, because the native SurfaceView sits behind an opaque webview) and Linux always has. Chromium and WebKit decode a far narrower set than the platform does, and the gap is widest on devices whose vendor licenses Dolby: a phone shipping `/vendor/etc/media_codecs_dolby_audio.xml` reports `ac3,eac3`, so Jellyfin direct-played an E-AC-3 track with `static=true` and the webview built a video decoder and no audio decoder at all — full picture, no sound. The defect is triggered by *capability*, not the lack of it, which is why it reproduced on one Motorola while a Fairphone and an Honor tablet played the same file on the same build: a device without the Dolby decoder never claims the codec, so the server transcodes to AAC and it plays. `video_audio_codecs` narrows the platform list to the webview-decodable set (`aac,mp3,opus,vorbis,flac`) for the video direct-play profile *only* — the audio-only profile keeps the full list, since that playback really is the native player's and narrowing it would transcode music that plays perfectly well. A list with nothing decodable still claims `aac` rather than going out empty, because a profile that claims nothing invites the server to give up instead of transcoding. The video codec list is deliberately untouched: HEVC direct-plays through the webview correctly, so the constraint is specific to audio | Playback | UR-004 | Done |
+| DR-149 | The client decides whether its own renderer can decode the audio, rather than trusting the server's negotiation. Advertising a webview-shaped profile (DR-148) turned out to be necessary but not sufficient: Jellyfin 10.11.5 enforces a `DirectPlayProfile`'s `Container` and `VideoCodec` — excluding either returns `SupportsDirectPlay: false` with `TranscodeReasons=ContainerNotSupported` / `VideoCodecNotSupported` — but **ignores its `AudioCodec`**, offering an E-AC-3 track for direct play against a profile listing only `aac,flac,mp3,opus,vorbis`. Neither a `VideoAudio` `CodecProfile` forbidding the codec nor a `MaxAudioChannels: 2` against a 6-channel track changes the answer, so no profile the client can send fixes it and the picture plays silent. The negotiated source's audio is therefore checked locally against what the webview decodes, and an undecodable track forces the existing h264/aac HLS transcode URL regardless of the server saying direct play is fine — `direct_play` and `needs_transcoding` are corrected to match, so the frontend and the reporting path agree with the URL actually used. The track judged is the one the server would serve: the default, or the first when nothing is marked default, since a supported track further down the list is not the one that plays. A source with no audio streams, or a stream whose codec the server did not name, is left alone — forcing a transcode on a guess spends server CPU on files that already play | Playback | UR-004 | Done |
+| DR-150 | Android video renders on the native ExoPlayer surface behind a transparent WebView, behind the `experimentalNativeVideo` opt-in. Rust already reported `use_html5_element: false` on Android, but two frontend overrides discarded it — `createAdapter()` hardcoded `"html5"`, and `VideoPlayer.svelte` forced `useHtml5Element = true` and stopped the native backend `player_play_item` had just started. The flag is a **suppressor, never a promoter**: off forces HTML5 even where Rust says native, so an in-progress spike cannot ship as the default, but it can never select native where Rust reported HTML5 (Linux cannot composite behind WebKitGTK, so promoting there is a black screen). Compositing requires clearing two independent opaque layers, and clearing only one leaves audio over a black picture — the WebView widget background and window drawable from Kotlin (`AndroidVideoSurface.setTransparent`), and the page's `html`/`body` and app-shell background from CSS (`data-native-video`). Transparency is declared in `tauri.android.conf.json` rather than the base config, because a transparent window on Linux has nothing behind it, and is toggled per playback session rather than set once, because a permanently transparent window shows the launcher through the rest of the app | Playback | UR-003, UR-004 | Done (behind `experimentalNativeVideo`, default off) |
+| DR-151 | The player's video SurfaceView actually reaches the view hierarchy. `JellyTauPlayer.setActivity()` had zero callers, so `currentActivity` was always null and `autoAttachSurface()` returned at "Cannot attach surface - no Activity reference". The surface was created and handed to ExoPlayer but never added to the content view, so native video decoded to a surface that was never on screen — independent of any webview transparency. `MainActivity.onCreate` now supplies the reference, which also revives PiP on the video path: `canEnterPip()` gates on `isVideoSurfaceAttached()`, which had been permanently false | Playback | UR-003, UR-041 | Done |
+| DR-152 | Platform playback facilities are reported by Rust, not sniffed from the user agent. `webviewAudio.ts` re-derived "does this platform have a native audio backend" by matching `navigator.userAgent` against `android`/`linux` — a second copy of the `cfg!` gate the backends are compiled under, free to drift from it. `player_get_capabilities` now returns `usesWebviewAudio` and `supportsNativeVideo` from the same cfg gates, and the frontend consumes them; the settings toggle for native video is hidden entirely where the platform cannot support it | Player | UR-003, UR-005 | Done |
+| DR-153 | The git tag is the single source of truth for a release version. The version lived in four files (`package.json`, `tauri.conf.json`, `Cargo.toml`, `Cargo.lock`) that had to be hand-edited in lockstep, and CI's release job rewrote exactly one of them — so a tagged build produced an installer named for the tag wrapped around package metadata naming the previous release, while the Linux job had no version step at all and shipped whatever was committed. `scripts/set-version.sh` writes all four from one argument and is the only thing that does; every release job calls it with the tag. The Android `versionCode` is derived in the same place as `1000 + major*10000 + minor*100 + patch`, which is monotonic in semver order and clears the 1000 floor already installed in the field — a lower code than the installed one makes Android refuse the update. A prerelease suffix is stripped before that arithmetic, which would otherwise abort the script, and a non-tag ref (CI passes `${GITHUB_REF#refs/tags/}` unconditionally) falls back to `git describe` rather than failing a branch build | Build | - | Done |
+| DR-154 | A watch position that cannot reach the server is queued, not dropped. `sync_queue` and its drain (DR-131) were built, tested and running, but the stop-report path never fed them: `HybridRepository::report_playback_stopped` is a bare pass-through to the online repository ("Playback reporting goes directly to server"), and on failure the error surfaced to a frontend `catch` whose own comment read "Server error - could queue, but for now just log". Both producers that *would* have queued it — `PlaybackReporter::queue_for_sync` in Rust and `syncService.queuePlaybackProgress` on the frontend — have no callers on the playback path, so closing a video while the server was unreachable lost the resume point outright even though `user_data.pending_sync` was dutifully set to 1 and nothing ever drains that flag for positions (unlike favourites, DR-120). The command layer now enqueues a `report_playback_stopped` row whenever the push fails, which the existing drain already knows how to parse and replay. The pending row for an item is **superseded in place** rather than appended to: progress is reported every 10s, so a server that stays down would otherwise add a row per tick, all of them obsoleted by the newest — the unbounded queue DR-131 exists to prevent. Only `pending`/`failed` rows are superseded, because an `abandoned` row has been given up on and reviving it would restore that same growing counter. Queueing is best-effort and never fails the command: the local position is already saved, so a failed *queue* write must not be reported as a lost position | Backend | UR-025, UR-002 | Done |
+| DR-155 | A watch position set on another device reaches this one. The resume check reads the local `user_data` row and nothing else, but `mirror_user_data` — the only path by which server `UserData` lands in that table — mirrored `is_favorite` alone, and returned early whenever that field was absent, which is exactly the shape of an ordinary watched episode. So `playback_position_ticks` was write-only from this device's perspective: watch 40 minutes in a browser, open JellyTau, and it resumed from whatever *this* device last saw or offered no resume at all — the same user-visible symptom as DR-150's Android bug, from an unrelated cause, which is why resume read as broadly flaky. The mirror now carries the position alongside the favourite flag under the same `pending_sync = 0` conflict rule, so a local position still waiting to be pushed is never pulled *backwards* by a server that has not yet heard where we got to; `COALESCE(excluded.x, user_data.x)` means a field the server omitted keeps its stored value rather than being nulled, and a row with neither field is still skipped rather than fabricated as zeroes. Mirroring alone was not sufficient: `get_item` — the call the player route makes — returned the cached copy on a hit and never consulted the server, so for an already-cached item the mirror never ran. It now refreshes in the background on a cache hit (`race_with_refresh`, the reusable form of what `get_items` already did inline), which is why browsing a season picked up other devices' state while opening the episode directly did not. The refreshed value lands for the next read, the cache-first race still answering immediately | Backend | UR-025, UR-002 | Done |
+| DR-156 | A page no longer inherits the previous page's scroll position. The shell keeps its scrollers alive across navigation by design — the root layout, the home page and the library layout each own a `flex-1 overflow-y-auto` box that outlives the route inside it, which is what lets `BottomUi` be a flex sibling rather than a measured overlay — but the element therefore never remounts and its `scrollTop` survives the route change. SvelteKit's own scroll restoration could not help: it saves and restores `window` scroll, and in this app the window never scrolls at all, so there was no scroll handling of any kind. The symptom was that opening an item from half-way down a library grid dropped the viewer half-way down the detail page, and returning to the grid landed at the top of it — exactly backwards. `ScrollMemory` (pure, one instance per container, keyed on path + query so a genre-filtered grid keeps its own place) records the offset a route is left at in `beforeNavigate` and decides in `afterNavigate`: `link`/`goto`/`form` reset to the top, `popstate` restores that route's saved offset, and the initial `enter` is left alone. Deciding does not consume the offset, so a route returned to more than once restores each time. Applied via the `scrollContainer` action on all three scrollers | UI | UR-072 | Done |
+| DR-160 | Picture-in-picture works on the path that actually plays video. PiP shrinks the whole *Activity*, so `canEnterPip` demanded a native ExoPlayer `SurfaceView` be attached and rendering — `isPlayingVideo() && getSurfaceView() != null && isVideoSurfaceAttached()`. But the native path sits behind `experimentalNativeVideo`, which defaulted to **off**, so in the shipping configuration video played in the WebView's `<video>` element and all three conditions were false. `enterPip` bailed with "Not entering PiP: no local video playing" every single time: the button was offered (gated only on OS capability) and could not work, however it was pressed. The manager now accepts either surface. The frontend reports the element through `AndroidPictureInPicture.setHtml5VideoState(active, width, height, playing)` — intrinsic size because the PiP window's aspect ratio came from the letterboxed surface's measured bounds, which do not exist here, and play state because `ExoPlayer.isPlaying` is false on this path and the PiP play/pause action would be frozen on "Play" mid-playback. Two behaviours invert when the WebView *is* the video: it must stay visible in PiP rather than be hidden (`hideWebView` is now gated on the native path — hiding it would leave an empty black window), and the play/pause `RemoteAction` has to reach the element, so the receiver dispatches `jellytau-pip-play`/`jellytau-pip-pause` DOM events instead of driving ExoPlayer. `jellytau-pip-entered`/`-exited` let the player strip its own chrome, since controls, title and gradients would otherwise be rendered into a window a couple of inches wide. The `<video>` is deregistered on teardown so PiP is never offered over a video that has gone | UI | UR-041 | Done (pending device verification) |
+| DR-167 | Each downloaded library shows only its own media. Cached items carry no link back to their library — `library_id` and `parent_id` are NULL on every row ([[offline-libraries-never-cached]]) — so `get_downloaded_items` matched the library branch with `EXISTS (SELECT 1 FROM libraries l WHERE l.id = ?)`, which asserts only that the requested library *exists* and never constrains the item to it. Opening any downloaded library therefore listed every downloaded top-level item on the server: films under Music, albums under TV. The sibling query that decides which libraries *appear* already carried the right rule — a `collection_type` ↔ `item_type` mapping — so the two disagreed about the same question. That mapping is now the named constant `LIBRARY_HOLDS_ITEM`, used by both, and a library of unknown collection type still keeps everything rather than being emptied by a rule that cannot classify it. The taxonomy stays in Rust, never the frontend | Downloads | UR-055 | Done |
+| DR-168 | Pause and resume actually stop and restart the bytes. `pause_download` wrote `status = 'paused'` and did nothing else, and no cancellation existed anywhere in the download stack — no token, no flag, no abort — so the streaming task ran on, kept writing, and overwrote the row with `completed`/`failed` when it finished: the row flicked to "paused" and undid itself. `resume_download` had the mirror defect, flipping the row to `pending` without calling `pump_download_queue`; the pump runs when something calls it rather than polling, so a resumed download sat untouched until an unrelated event happened to pump the queue. A per-download stop flag (`download::stop`) is the missing half — a module-level registry because the two sides never meet, the command holding Tauri state and the worker running detached in `async_runtime::spawn`. The worker reads it between chunks and on retry (so a pause is not swallowed by a 45-second backoff), flushes, and returns `Stopped`, which is deliberately **not** retryable and **not** recorded as a failure: the `.part` file is left intact because that is exactly what the resume's Range request continues from. Registering returns a *fresh* flag, or a resumed download would inherit the pause that stopped it and halt instantly. Cancel and `clear_stale_downloads` signal it too, so neither deletes a file still being written | Downloads | UR-055 | Done |
+| DR-169 | Partial files are actually reaped. The worker named its sidecar with `Path::with_extension("part")`, which *replaces* the extension — `movie.mp4` became `movie.part` — while every cleanup path deleted `"{file_path}.part"`, i.e. `movie.mp4.part`. The two never matched, so the partial file of every cancelled or failed download stayed on disk indefinitely, invisible to the disk-usage totals because no `downloads` row pointed at it. `partial_path` appends instead, is the single definition both the writer and the cleaners use, and incidentally removes a collision the old form had, where `movie.mp4` and `movie.mkv` mapped to one `movie.part` | Downloads | UR-055 | Done |
+| DR-173 | Downloading an album queues the **whole** album, and every track it queued is findable offline afterwards. Two independent gaps left an album with a handful of its tracks on the device while the button reported the album as downloaded. First, `download_album` took its track list from `items WHERE album_id = ?` — the local catalog cache. Jellyfin does not return `AlbumId` on every listing endpoint, so tracks cached by one of those endpoints sit in `items` with a NULL `album_id` and are invisible to that query; on the reporter's database three whole albums (18, 12 and 9 tracks) had it NULL on *every* track, so "download album" would have queued nothing for them, and a partially-linked album queued only the linked subset. Second, the frontend then resolved one stream URL per track from its own list and paired it with the returned row ids **by position** — a pairing with no basis, since the ids came back in the backend's `index_number` order over a different set of rows, so a row could be handed another track's URL and any track past the end of the shorter list was never started at all; on Android that loop also stopped wherever the webview was suspended. The same `album_id` is what `OfflineRepository::get_items` joins a track to its album on, so a track that did download stayed invisible under its album offline — the two halves of the same missing link. The operation now belongs to Rust end to end: `HybridRepository::get_album_tracks` asks the **server** what the album contains (cache-first `get_items` is right for browsing and wrong for deciding what to download) and errors offline so the caller falls back to the ungated local catalog, keeping the queue-while-offline flow; `queue_album_tracks` writes the album link onto every track it queues — queuing a track *is* the statement that it belongs to the album, rather than something to hope a listing endpoint recorded — and the stream URLs are resolved here through the existing reconnect resolver, now scoped to the rows just queued so one album cannot start every unrelated pending row. Nothing crosses the IPC boundary but the album id. Re-queuing a broken album heals it: the missing tracks are added and the tracks already on disk get their link. `download_series`/`download_season` still derive their episode lists from the cache the same way and want the same treatment | Downloads | UR-018, UR-055 | Done |
+| DR-170 | Downloads at a chosen bitrate are no longer corrupted by their own retries. Only the `original` preset asks for `Static=true`; every other rung requests a **transcode**, which Jellyfin serves chunked, with no `Content-Length`, and cannot byte-seek — so it ignores `Range` and answers `200` with the whole stream from the beginning rather than `206` with the requested tail. The worker sent the Range header whenever a `.part` existed and appended the body unconditionally, so each retry and each resume concatenated a fresh copy of the entire transcode onto the bytes already on disk: the file grew past its real size and would not play, which is why "downloads for different bitrates" stayed broken after the `videoBitRate` casing fix (DR-adc460f3) corrected the *request*. `resume_offset` makes the response decide — append only on a `206`, otherwise truncate and take the stream from the top — and the total size is computed from that offset rather than from a partial length the server never agreed to | Downloads | UR-071 | Done |
+| DR-172 | Native Android video is opt-in again, because as a default it shipped as **audio with no picture**. DR-161 flipped `experimentalNativeVideo` on so picture-in-picture could shrink a real video surface; on a device that produced sound and a blank screen. The decode path was never the problem — logcat showed ExoPlayer running (`Position update` ticks) and feeding a live `SurfaceView` with an active BufferQueue. The compositing was: the SurfaceView sits *behind* the WebView, and the step that clears the opaque layers above it never took effect, with `WebView transparent = false` logged and `= true` never appearing. So the video rendered correctly the whole time, behind an opaque page. This is exactly the defect the flag existed to contain — `VideoPlayer.scrubRegression.test.ts` had recorded that "the native SurfaceView has never been visible through the webview" — and enabling it by default shipped a verified decode path on top of an unverified display path. Reverting costs nothing that matters: PiP does not depend on it (DR-160 drives PiP from the WebView `<video>`), and working video outranks PiP showing a native surface. The flag stays available in Settings, now described as incomplete rather than as a performance win, and the scrub-regression mocks that were made explicit under DR-161 are kept explicit so those tests state which path they guard rather than inheriting a default that has now moved twice. Fixing the compositing is the prerequisite for trying this default again | UI | UR-003, UR-004, UR-041 | Done |
+| DR-171 | A downloaded video keeps audio the device can actually decode. `original` quality asked for `Static=true`, which hands back the source file byte-for-byte — E-AC-3/AC-3/DTS/TrueHD track included — and video is rendered on both platforms by the webview `<video>` element, which decodes none of them. Streaming already knew this: DR-149 judges the track the server would serve against `WEBVIEW_AUDIO_CODECS` and forces a transcode over Jellyfin's own direct-play offer, because 10.11.5 honours a `DirectPlayProfile`'s container and video codec but ignores its audio codec. The download path never consulted that policy, so the *same film* had sound when streamed and played as picture in silence once downloaded — and offline a download is the only source a video has, so there was no working path left to fall back to. The rule is now one rule: `served_audio_codec` picks the track the server will serve (the default, or the first when none is marked) and both callers judge it, the streaming verdict staying a bool and the download path needing the codec itself so it can say what to re-encode. Only the audio is re-encoded — `allowVideoStreamCopy=true` keeps an h264 source's picture byte-for-byte and no bitrate or resolution cap is added, so `original` still means original quality; a source the webview could not have rendered anyway (HEVC) becomes h264 as a side effect, which is the only form of it that would have played. The decision is per item rather than blanket because the transcode costs the byte-range resumability `Static=true` gives the download worker (see DR-170 for what a chunked, length-less response does to a resume), so a file whose audio already plays keeps the direct copy. An unknown codec — item not fetchable, or the server named none — changes nothing: the policy only ever *adds* a transcode, so it cannot make a working download worse. The codec set judged against is the **webview's**, not the platform's, even though DR-161 made ExoPlayer the Android default: `experimentalNativeVideo` is a user setting, a downloaded file outlives whatever it was set to when the file arrived, and the narrow list is the only one that holds on both sides of it — at the cost of a Dolby-licensed device re-encoding a track its ExoPlayer could have played. `resolve_video_download_url` is the single entrance for all three resolution sites (the frontend's per-item command, the bulk series/season enqueue, and the offline-queued resume), since the pure builder cannot look a codec up and a caller that forgets to is exactly how the silent downloads shipped. **Files already downloaded stay silent** — the bytes on disk are the wrong bytes and only a re-download replaces them | Downloads | UR-071, UR-004 | Done |
+| DR-162 | Video streams are opened against a **bandwidth ceiling the user chose**, instead of a fixed allowance nobody could change. Every video URL carried `MaxStreamingBitrate=20000000`/`VideoBitrate=18000000`, `PlaybackInfo` negotiated at 20 Mbps, and the device profile advertised `999999999` — so on a metered or slow connection the only lever was not watching. `StreamingQuality` is a ladder of ceilings (Original, 20/10/8/4/2/1 Mbps, 720 kbps) in which a step is not a label but a bundle of transcode parameters: the total ceiling, the audio share of it, and the resolution that budget can carry. It lives in Rust because those numbers are Jellyfin encoding vocabulary — the frontend names a variant and reads labels back over `player_get_streaming_qualities`, the same arrangement as the EQ preset curves. The video bitrate is the total *minus* the audio share, so the two together honour the cap rather than overshooting it by the size of the audio track, and `MaxHeight` falls with the ladder so a small budget is not spent on pixels it cannot afford. The cap has to reach the **negotiation**, not only the transcode URL: `max_static_bitrate` in the device profile is what makes the server refuse to direct-play a source fatter than the ceiling, and without it a 30 Mbps remux is handed over untouched and every URL parameter downstream is moot — which is why it is applied at all four places that decide bandwidth (the HLS builder, `PlaybackInfo`, `open_live_stream`, and the background-audio handoff, which takes the lower of the cap and its own 384 kbps). The ceiling is process-wide rather than a field on `OnlineRepository`, mirroring `INCLUDE_CATALOG_BROWSE`: it is a preference about *this device's connection*, it must survive a repository rebuilt on re-login, and every builder plus the negotiation have to agree on it or the cap leaks. Settings owns the durable default and is the only writer to `app_settings` — persisted unlike the rest of `VideoSettings`, because a limit set for a metered connection that silently reverts to uncapped on the next launch spends the user's data with no changed setting to show for it — and it is restored at startup from the async runtime, defaulting to uncapped if the read fails so a database problem degrades to the old behaviour rather than to an arbitrary limit. The in-player menu is the per-video override: a cap is a property of the stream the server is producing, so it cannot be applied to one already in flight — `player_set_stream_quality` re-opens the stream at the new quality and resumes at the current position, reloading a native backend itself and handing HTML5 a URL for the same `reloadSource` primitive the audio-track switch uses, so no strategy branch lives in the UI. It deliberately does not persist. This gives UR-070 its resume-at-the-same-point mechanism; the server-offered per-item rendition list that requirement also asks for remains proposed | Playback | UR-074, UR-070 | Done |
+| DR-174 | Tiles of mixed shapes are laid out **justified** rather than gridded. A CSS grid gives every cell one box, so on a page holding square music covers, 16:9 library backdrops and 2:3 posters at once, everything that is not the chosen shape is cropped to it — the home shortcut strip was explicitly forcing `aspect="video"` on music libraries for exactly this reason, which lined the row up by cutting the covers down. `layoutMosaic` packs tiles into rows of a **shared height** and gives each its own width from its own aspect ratio: it adds tiles to a row until the height needed to fill the container has fallen to the target, closes the row there (so rows land at or below the target, never above), and justifies the row to the container width by absorbing the rounding remainder into its widest tile, where a pixel is least visible. The last row is deliberately *not* justified — with one tile left over, filling the width would inflate it to a banner — so it sits at the target height, left-aligned. Ratios are clamped to a band, which costs a crop on genuine outliers and stops one panorama owning a row or one very tall image shrinking to a sliver. It is a pure module with no DOM: the component supplies only the two things the DOM knows — the measured container width, and the artwork's *decoded* aspect ratio, reported by `CachedImage` so the layout uses the shape an image actually has rather than the one its item type implies. Those measurements are committed in one debounced batch rather than per image, because artwork arrives over several hundred milliseconds and re-packing on each arrival would shuffle the grid under the pointer repeatedly. Labels are drawn *over* the bottom of each tile rather than beneath it: a caption below sits outside the computed box, and one that wraps to two lines would break the row alignment the layout exists to provide | UI | UR-075 | Done |
+| DR-175 | A library knows which favourites category it belongs to, and the frontend does not work it out. The mosaic offers a favourites tile per category beside its library, which needs a collection-type → category answer; deriving it in Svelte would have re-created the exact leak `SearchScope::item_types` was extracted to close (docs/specs/scoped-search-boundary.md) — one table of Jellyfin vocabulary, differing only in which vocabulary. `SearchScope::for_collection_type` maps `movies`/`tvshows`/`music` and returns `None` for everything else, so a Live TV or books library gets no tile at all rather than one opening an unfiltered list; `All` is never derived from a library, being the cross-library entry offered beside them rather than a property of one. `Library::new` stamps the result onto every library at construction — a constructor rather than a struct literal precisely so a derived field cannot be forgotten at one of the four sites — and it rides to the frontend as an optional `favoritesScope`, absent rather than null when there is none. The UI's remaining share is presentation only: what to call the tile, where to put it, and showing a category's tile **once** however many libraries share it, since two movie libraries have one favourites list between them | UI | UR-075, UR-067 | Done |
+| DR-176 | The server is never asked to burn a subtitle into the picture. `PlaybackInfo` omitted `SubtitleStreamIndex`, which does not mean "none" — the server then honours the source's default/forced flag and picks a track itself. On a source whose default subtitle is image-based (PGS/DVD/DVB) that track cannot go out as a sidecar, so the server falls back to `SubtitleMethod=Encode` and composites it into the video. The cost lands on the *video*, not the subtitle: burn-in rules out remuxing, so an HEVC stream the device could have taken untouched is re-encoded frame by frame. Observed on an HEVC + E-AC-3 + PGSSUB episode, where only the audio actually needed transcoding: the server could not sustain the re-encode in real time, the buffer never grew past a single segment, and playback stalled every few seconds — taking seeking with it, since each seek restarted the encoder and cost seconds before the first frame. The fix is to request `SubtitleStreamIndex=-1` explicitly and to advertise every *text* format we can render (`srt`/`subrip`/`ass`/`ssa`/`vtt`) as `External`, so a subtitle can only ever arrive as a sidecar. Nothing is lost, because the app already fetches subtitle tracks itself and draws them over the video (UR-020) — the server's composited copy was always redundant. Image-based tracks are consequently not offered, which is honest rather than a regression: the renderer cannot composite a bitmap, and the previous behaviour paid for them by making the whole stream unwatchable. Both halves of that hold at the layer that can enforce them. The sentinel travels on the stream URL as well as in the negotiation, because the negotiation is not what opens most streams — a quality switch, a transcoded seek and an audio-track switch each rebuild the URL on their own, and an omitted index there lets the server pick the default track back up out of whatever session state it still holds. And "not offered" is enforced where the offer is made: each subtitle stream crosses the boundary carrying the backend's verdict on whether it can arrive as a sidecar, so the picker lists only tracks the app can draw instead of showing an entry that ticks and displays nothing. Only an explicit "no" hides a track, so a stream carrying no verdict behaves as before | Playback | UR-020, UR-004 | Done |
+| DR-177 | Each video transcode this device opens is its own server-side job, and the one it replaces is stopped. Jellyfin keys a transcode job by device **and** play session, and every stream URL the app built carried the same hardcoded `DeviceId` with no `PlaySessionId` at all — so the second stream for an item was indistinguishable from the first. Re-opening a stream is not rare: a mid-playback quality switch (UR-074), a transcoded seek and an audio-track switch all do it, each leaving the previous ffmpeg running. Observed on-device when switching bitrate mid-film: the server served the new playlist, then rejected the new job's segments with `400 hls1/main/0.ts` while the two jobs contended for one transcode path, and playback stalled — reproducible against the server, where a second stream for a live job's item alternates between serving bytes and 400ing per attempt, which is what made it read as flaky rather than broken. `begin_video_play_session` mints a session id per open and reports the one it supersedes; the URL builder stops that job (`DELETE /Videos/ActiveEncodings`, un-retried and best-effort — a slow stop must not delay playback, and the new stream no longer collides either way) before returning. Placing it in the URL builder rather than in each caller means every re-open path is covered by construction. Two client faults made the same incident worse and are fixed with it: the fatal-HLS-error handler added the transcode seek offset to a position that already included it, so past roughly the halfway mark of a film any transient network error cleared the "near end" threshold and was reported as end-of-stream — turning a recoverable stall into a skip to the next item, exactly when a quality switch had just made the offset large; and the HTML5 reload primitive resolved on its own `canplay` timeout, so a reload the server never served reported success, leaving the picker showing a quality that was not playing and the caller with nothing to revert | Playback | UR-074, UR-004 | Done |
+| DR-178 | Every position that leaves the app is read from the controller, not from a backend that may not be playing anything. `PlayerController::position()` forwards to the native backend, which is authoritative for exactly one of the three ways this app renders media. On the **webview** path — the shipping default for video on both platforms — nothing is loaded into that backend at all: the `<video>` element is the player, its ticks were re-emitted to the frontend and then dropped, and the backend answered 0 forever. During a **background-audio handoff** the base that converts the stream's relative timeline to the episode's is applied once at the native tick boundary (DR-159), so before ExoPlayer's first tick nothing has applied it and the reading is 0 there too. Both holes surfaced as the same user-visible bug through different doors: returning to the foreground while the audio-only transcode was still opening handed the frontend `0.0`, and the video reloaded at `StartTimeTicks=0` — the episode restarting from the beginning — while the `Stopped` report that followed wrote that zero to Jellyfin as the resume point. `absolute_position()` answers for all three paths: the maximum of the backend's reading, the last position webview-rendered media reported, and the handoff base. The maximum is exact rather than a heuristic, because at most one term is ever meaningful at a time and the base is a floor the stream cannot physically be behind. `duration()` gains the same fallback for the same reason. The element's reading is cleared wherever it stops being the player — teardown, a handoff taking over, a different item loading — so it can never be attributed to what plays next | Player | UR-005, UR-025, UR-040 | Done (pending device verification) |
+| DR-179 | Jellyfin is told what was played: progress while it plays, and a stop when it ends. A device trace of 35 minutes' playback requested `/Sessions/Playing/Progress` **zero** times and sent 14 `Stopped` reports, every one of them at position 0. Three faults, one subject. *Progress never left the device*: the frontend service writes it to the local DB by design, and nothing on the Rust side reported it for webview-rendered media — so the server learned a position only when the player was closed, and a crash or a swipe-away cost the session. It is now reported from the controller's own position ticks, through the 30s throttler it already owned and shares with the native audio path, which covers all three rendering paths in one place instead of adding a second frequent IPC caller. *Zero-position stops were sent*: Jellyfin stores the reported position as the resume point, so a zero does not merely fail to inform, it instructs the server to forget — and no zero was ever real, each one coming from asking a player that was not rendering the media (see DR-178). They are withheld; one landed 40s after the frontend had correctly reported 15:22 for the same episode, overwriting it. *A finished episode reported nothing at all*: Jellyfin decides "watched" from the stop report and its percentage, and in background audio-only mode nobody sends one — the webview is suspended and its element was torn down at the handoff, while the backend advances to the next episode without a word about the one that ended, so an episode listened to end-to-end on the lockscreen never counted as watched. `on_playback_ended` now reports it stopped at its **runtime** (not the last tick, which can be seconds short or, on a handoff whose ticks stopped early, nowhere near the end) before any advance, since after one the queue's current item is the next episode. Scoped to the audio-only handoff, the case the frontend provably cannot cover, so foreground playback keeps its single existing report; music ending natively remains unreported and wants its own change. The reporting seam is a `PlaybackReportSink` the controller sends to, which also collapses three copies of the spawn-a-task-and-hope block into one and is what let all of this be written as failing tests rather than found on a device a second time | Player | UR-025, UR-005, UR-040 | Done (pending device verification) |
+| DR-180 | A background-audio handoff of a **downloaded** episode starts where the video left off. The handoff prefers a local file over the audio-only stream (DR-128), but the two begin in different places and were treated alike: a stream is built with `StartTimeTicks`, so the server makes the handoff point that stream's zero and the base is the handoff position with no seek — while a file has no such parameter and begins at the episode's own zero, so basing it at the handoff position claimed minutes of audio that were about to play from the beginning. Backgrounding a downloaded episode therefore restarted it while the lockscreen scrubber, dutifully adding the base, showed the position it should have been at. `background_audio_plan` splits the two: a file gets no base and a real seek, a stream keeps the base and no seek (seeking one would skip *past* the content by the handoff position again). The same distinction settles an inbound seek — `seek_absolute` re-opens a *streamed* handoff at the requested position because a chunked length-less transcode cannot honour a seek, which is not true of local media, and `resume_stream_at` refuses a non-remote source outright, so routing a lockscreen scrub of a downloaded episode through it failed the seek rather than performing it | Player | UR-040, UR-071 | Done (pending device verification) |
+| DR-181 | A resumed transcode plays. Every video stream URL carried the resume position as `StartTimeTicks`, which is correct for a progressive response and fatal for an HLS one: Jellyfin builds each segment URI by echoing the **master playlist's** query string into it, and its segment handler opens by rejecting any request carrying `StartTimeTicks > 0` (`ArgumentException` → `400`). One position on the playlist therefore 400s every `hls1/main/N.ts` behind it, so hls.js exhausted its retries and gave up — presenting as an episode that will not resume while the same episode from the beginning is fine, the `> 0` being exactly why the beginning survived. The parameter is also unnecessary there: a playlist spans the whole item and asking for segment N *is* the seek, which the server transcodes from. So it is removed from the URL builder entirely rather than conditionalised — the builder has one caller shape and no way to know whether the response will be segmented — and the position becomes what it always was for HLS, a seek issued once the player has loaded: the seek path reloads at zero and seeks the element, and the resume path lets the player seek itself. The progressive `/Audio/universal` builder used by the background-audio handoff is a different endpoint with no segments and keeps its `StartTimeTicks`, which is why an audio-only handoff resumes correctly and a video one did not | Playback | UR-004, UR-074 | Done |
+| DR-182 | Native video shows a picture. The poster/title card is an opaque `bg-black` overlay drawn over the whole video area while `isMediaReady` is false, and **every** signal that clears it is emitted by the HTML5 `<video>` element — `canplay`, `loadedmetadata`, hls.js `FRAG_BUFFERED`, the `playing` event, and two `readyState` timeouts. The native path renders no such element (`{#if !!useHtml5Element}`), so on Android nothing could ever clear it: ExoPlayer decoded to a live SurfaceView behind a black div for the entire session. That is DR-172's "audio with no picture" report, and it is indistinguishable on screen from the compositing failure DR-172 attributed it to — which is why the flag was reverted rather than fixed. Both the overlay and the native branch date from the original POC commit, so the native path has never been able to reveal itself; the 2026-08-11 device verification predates neither and does not contradict this, since a spike run that never reached a steady state would not have shown it. The backend's own events are the equivalent signals and `nativeSignalRevealsVideo` is the rule for reading them: `state === "playing"` mirrors the element's `playing` event, and a position tick carrying a real position or duration mirrors the `readyState` backstops, covering a first state event that is dropped or arrives before the listener is attached. `buffering`/`paused`/`stopped`/`error` deliberately do not qualify — revealing on `error` would replace the title card with a transparent hole showing the launcher through the app. The rule is a pure module rather than a branch inside the component because the decision that was missing is exactly the part worth guarding, and the component needs a DOM and a mounted player to exercise | UI | UR-003, UR-004, UR-041 | Done |
+| DR-183 | The JavaScript bridges are installed before the page that uses them loads. WebView binds an injected object into JS at **page-load time**: an `addJavascriptInterface` call landing after the page has loaded does not appear to that page. They were installed from `configureWebViewForMedia`, which finds the WebView by walking the view tree 500 ms after `onCreate` — a race against Tauri's own page load, and one that is *permanent* when lost, because the identity guard added for DR-097's stale-proxy bug then declines to re-inject on every later resume pass. The whole set (`AndroidVideoSurface`, `AndroidPictureInPicture`, `AndroidBackgroundAudio`, `AndroidNetworkType`, `AndroidImmersive`, `AndroidInsets`) would simply be absent from `window`, and silently: every call site optional-chains the bridge, so a missing one is a no-op rather than an error. This is a candidate explanation for DR-172's other piece of evidence — `WebView transparent = false` logged, `= true` never appearing, i.e. the enable call never reaching Kotlin at all. `WryActivity.setWebView()` calls the `onWebViewCreate` hook immediately before wry issues the first `loadUrl` (confirmed in wry 0.55's `main_pipe.rs`, where the `setWebView` JNI call precedes `load_url`), so a bridge installed there is bound by the time any page runs. The hook can fire during `super.onCreate()`, before the rest of our own `onCreate`, so only work needing nothing but the WebView moves into it — insets stay in `configureWebViewForMedia`, which runs later and on every resume. The tree-walk path is kept as a fallback, and `enableNativeVideoCompositing` now logs an explicit error when the bridge is missing, so the ambiguity that left DR-172 unresolved cannot recur silently | Android | UR-003, UR-004, UR-040, UR-041 | Done |
+| DR-184 | The video SurfaceView leaves the view hierarchy when the video does. `VideoOverlayManager.detachVideoSurface` had **no callers anywhere in the tree** — the mirror of the DR-151 defect, where `setActivity` had none — so `attachVideoSurface` was one-way: `JellyTauPlayer.clearVideoSurface()` dropped its `surfaceView` reference and cleared ExoPlayer's without removing the view, leaving it parented to the content view for the life of the process, with the next native video adding another SurfaceView beneath it. The stack was invisible while the WebView was opaque, which is why it went unnoticed. Two consequences outlive the leak: `isVideoSurfaceAttached()` gates `PictureInPictureManager.canEnterPip` through `isNativeVideoPath()`, so it reported an attached surface forever after the first native video (saved from offering PiP over nothing only by the `isPlayingVideo()` check beside it), and every abandoned surface held its `OnLayoutChangeListener` on the content view. Detach is called from `clearVideoSurface`, which covers stop, the switch to audio, and the background-audio handoff, and always runs on the main thread because every caller is already inside a `mainHandler.post`. It removes the view from its *own* parent rather than looking the content view up from an Activity reference, so an Activity recreated underneath it cannot strand the view | Android | UR-003, UR-041 | Done |
+| DR-185 | The app shell stops painting over the video surface. `app.css` clears the page's opaque layers for native video through three selectors, and one of them — `html[data-native-video="active"] [data-app-shell]` — was written against an attribute **no component has ever set, in any commit**. The shell is `+layout.svelte`'s root `div`, which paints `--color-background` across the entire viewport; VideoPlayer is `fixed inset-0 z-50` and correctly makes *itself* transparent on the native path, but it stacks *above* the shell, so the WebView still composited the shell's opaque background over the whole screen and the SurfaceView behind it could never be seen. This is the missing half of the compositing DR-172 went looking for: the spec's own layer table lists this layer as "cleared by `data-native-video` → app.css", which was written but never wired, and `html`/`body` being genuinely transparent made the CSS look correct in isolation. The failure is invisible three ways over — the CSS is valid, the selector is plausible, and a rule matching nothing looks exactly like a rule matching something already transparent — while the symptom (black screen, audio fine) is identical to a real compositing failure, which is how it survived DR-150 through DR-172. Fixed by setting the attribute the rule was written for, and guarded by asserting the *relationship* rather than the rule: every attribute the compositing block targets must be set somewhere in the app, so a selector aimed at nothing fails the suite instead of failing silently on a device | UI | UR-003, UR-004, UR-041 | Done |
+| DR-186 | The play overlay comes down when the backend plays. `isPlaying` was assigned once from the `player_play_item` response and thereafter only by the `player://state-changed` listener — a channel the backend never emits, the same dead wire that DR-182's first fix was mistakenly hung on. On the native path the flag therefore froze at whatever the initial response said: with ExoPlayer playing, the UI still believed it was paused, so the `bg-black/30` play-button overlay stayed raised across the whole video area and the transport button kept showing ▶. The video was simultaneously dimmed and covered while it played, which reads as "the overlay never goes away" and is easily mistaken for a second compositing fault. The mirror reads the same `player` store `playerEvents.ts` feeds, which is what the architecture already says is authoritative — the player reports state, the UI consumes it — and is gated to the native path so HTML5 keeps its element-event wiring, which is authoritative there | UI | UR-003, UR-005 | Done |
+| DR-187 | The system bars go away with the player, not only with the fullscreen button. `enterImmersive()` had exactly one caller, `toggleFullscreen()`, so opening the player left the status and navigation bars painted over it until the user pressed a button most never press. On the native path this is worse than cosmetic: the SurfaceView fills the content view, so the bars sit directly on top of the video. The player is a full-screen surface by construction — `fixed inset-0 z-50` over a `MATCH_PARENT` surface — so entry is the right moment. Called synchronously in `onMount` before any `await`, per the native-mode pitfall, and paired with the `exitImmersive()` already unconditional in `onDestroy`, so a player torn down while immersive cannot leave the rest of the app without bars | UI | UR-066, UR-003 | Done |
+| DR-188 | Native Android video is **ready to be the default except for the background-audio handoff**, and the flip therefore waits. The picture defects behind DR-172 are all found, fixed and device-verified — DR-185 (the app shell painted over the surface through a CSS rule targeting an attribute nothing set), DR-182 (nothing could lift the poster card on a path with no `<video>` element), DR-183 (the JS bridges raced the page load, so `setTransparent(true)` could never arrive), DR-184 (the SurfaceView was never detached), plus DR-186 and DR-187, the two UI defects only this path could reveal. On a device logcat now carries `WebView transparent = true` and `Marking media ready` with video on screen, which is the pair DR-172 went looking for and could not find, and skip, seek and rotation were exercised by hand. Turning the default on then surfaced a *different* unverified sub-path: returning from background audio is HTML5-only (DR-190), so on the native path playback simply stays dead. Shipping it would have repeated DR-161 exactly — a verified sub-path made default over an unverified one — so the default stays off and the flip is gated on DR-190 rather than on more confidence | UI | UR-003, UR-004, UR-041 | Blocked by DR-190 |
+| DR-191 | Forcing the WebView overlay to redraw from the Activity, because with the ExoPlayer **SurfaceView** beneath it the overlay's ordinary damage stopped reaching the screen: the page kept mutating — the clock text every second, the control bar's opacity going to 0 — while the display held whatever frame it last presented, over video that animated perfectly. Not a state defect; the live DOM showed the slider advancing 476 → 479 across three seconds behind a screen showing neither. Only **structural** changes got through, which is why the play overlay always appeared to work (an `{#if}` block, added and removed) while the progress bar never did, and why rotation lost the transport UI. A CSS animation cannot help, since opacity animates on the compositor without repainting the layer. **Superseded by DR-192**: this drove `postInvalidateOnAnimation` in a loop, which treats the symptom — the cause is the SurfaceView's separate layer, and removing that removes the need. Kept as the record of how the mechanism was identified | Android | UR-003, UR-004 | Superseded by DR-192 |
+| DR-195 | Play/pause works on the native path, because the frontend stops claiming a webview element is playing when there is none. `html5_playing` is Rust's record of "a webview `<video>` is active and in this state", and `toggle_playback`, `play` and `pause` all route transport to that element whenever it is set. The player route mirrored element state into it **unconditionally** — from `handleReportStart` and, fatally, from `handleReportProgress`, which VideoPlayer calls on a 10-second interval — so on the native path the frontend re-declared every ten seconds that an element was playing when none existed, and every transport intent was emitted into the void. The pause button was dead from the on-screen tap, from the control bar, and from a direct `player_toggle` invocation, while seek and skip kept working because `player_seek_video` decides elsewhere; that asymmetry is the signature. It also explains the flashing, since the control bar and the JRay overlay both key off `isPlaying`, which was being contradicted on every interval tick. DR-193 clearing the flag at load was necessary but insufficient on its own — the interval put it straight back. The mirror now lives in `mirrorElementStateToRust` in VideoPlayer, gated on `useHtml5Element`, which is the only place that knows whether an element renders at all; the route cannot tell the two paths apart, which is precisely how it came to lie. Confirmed on device by ADB: surface tap and control bar each pause (position frozen across repeated samples, transport label flipped) and resume | Playback | UR-005, UR-003 | Done |
+| DR-194 | The previous frame flashing on rotation. It reads as a TextureView artefact — the view retains its last frame, so between a rotation and `fitSurfaceToScreen()` landing that frame sits at the old size — and two fixes were built on that reading: revealing after two `postOnAnimation` hops, then revealing on `onSurfaceTextureUpdated`, which required owning the `SurfaceTextureListener` and handing ExoPlayer the Surface directly rather than via `setVideoTextureView`. **Neither stopped the flash.** The mechanism is the *window's* rotation animation: Android cross-fades a **screenshot of the old orientation**, that screenshot contains the old video frame at the old size, and no TextureView bookkeeping can reach it — nor can the app pre-empt the screenshot, since `onConfigurationChanged` fires after it is taken. The only lever is to stop the animation: `ROTATION_ANIMATION_JUMPCUT`. That was accepted and silently ignored at first, and the platform said why out loud — `VRI[MainActivity]: setLayoutParams: not fullscreen` — because the attribute is honoured only for a fullscreen window. `FLAG_FULLSCREEN` (deprecated for hiding system bars, which immersive mode does instead, but still what marks the window fullscreen for this decision) is therefore set alongside it, scoped to while native compositing is active so the rest of the app keeps its normal animation. The frame-arrival reveal is kept: it replaced a fixed-timeout guess with a real signal, and its timeout is required rather than defensive, since a resize while paused means no new frame is ever coming. **The flash is not confirmed fixed on device** — the forced-rotation harness (`settings put system user_rotation`) proved unreliable, and `screenrecord` fixes its canvas at start so a rotation inside a recording never changes frame dimensions, which defeated two attempts at measuring it | Android | UR-003, UR-066 | Needs device verification |
+| DR-193 | Play/pause reaches the player that is actually rendering. `toggle_playback`, `play` and `pause` all route to the webview element when `is_html5_active()`, which is `html5_playing.is_some()` — a flag written **only** by the element's own state reports and cleared only when it reports "stopped"/"idle" (or on a background-audio handoff). An element that went away without that final report, or webview-rendered music earlier in the same process, therefore left the flag set, and on Android's native video path every transport intent was emitted as a `ControlCommand` at an element that no longer existed: the pause button did nothing, from the on-screen tap and from the control bar alike, while seek and skip kept working because `player_seek_video` decides elsewhere. Whether it happened at all depended on what had played before, which is exactly what made it read as flaky rather than broken. `load_and_play` — the native load path, and the one the HTML5 video path deliberately avoids via `set_current_item` — now clears the flag, because loading into the native backend *is* the statement that native renders this item. Nothing is lost on the webview path: an element re-establishes its own authority the moment it reports again, so this is the existing "element is gone" semantics applied where it can be known directly rather than inferred from a report that may never arrive | Playback | UR-005, UR-003 | Done |
+| DR-192 | Native video presents through a **TextureView**, not a SurfaceView. A SurfaceView renders on its own layer *outside* the app window and punches a transparent region through it; everything drawn above that hole — for us the entire Svelte UI in a transparent WebView — depends on that composition path, and Android's own graphics documentation states that "overlays do not currently work correctly with SurfaceView or TextureView". The consequences were four symptoms of one cause (DR-191): a frozen progress bar, controls that would not fade, rotation losing the transport UI, and overlays that lingered after the DOM removed them. A TextureView is an ordinary view whose frames are drawn as a texture in the window's normal rendering pass, so there is no second layer and no transparent region, and the WebView above composites like it would over any other view — which is why media3 offers `surface_type="texture_view"` and why it is the standard remedy for ExoPlayer overlay problems. The trade is accepted rather than hidden: TextureView costs more power and memory than SurfaceView and adds a frame of latency, but hardware decode through MediaCodec is untouched, so the reason native video exists survives it. `setVideoTextureView` installs ExoPlayer's own `SurfaceTextureListener`, so the old `SurfaceHolder.Callback` wiring is deleted rather than ported — adding a listener of ours would displace it and the video would never appear. PiP needs no change, since a TextureView is a View and the aspect-ratio probe reads its measured bounds | Android | UR-003, UR-004, UR-041 | Done |
+| DR-190 | The background-audio handoff can return to the native path. Everything that restores playback on the way back is written around the WebView `<video>`: `applyPendingForegroundSeek` returns early on `!videoElement`, the HLS re-init `$effect` returns early on `!useHtml5Element`, and `pendingForegroundSeek`/`pendingForegroundPlay` — which own the post-handoff position and play/pause — are consumed only by `handleCanPlay` and `markMediaReady`, an element event and a path that reaches the same guard. On the native path there is no element, so `exitBackgroundAudioHandoff` completes, clears `handoffState`, blanks and reassigns `currentStreamUrl` to force an effect that will not run, and nothing ever restarts ExoPlayer: the user returns from the lockscreen to a dead player. This never showed while the path was opt-in and its picture was invisible anyway. The return needs the native equivalent of the element reload — re-issue the item to the backend, seek to the position `player_exit_background_audio` reports, then honour `wasPlaying` — routed through the adapter rather than the element, so both paths restore through one contract | Playback | UR-040, UR-003 | Proposed |
+| DR-161 | Native video is the default, so picture-in-picture has a real surface. DR-160 makes PiP work on the HTML5 path, but that path can only ever shrink the *UI* into the PiP window; showing the video itself needs the SurfaceView behind the WebView, which is what `experimentalNativeVideo` gates. The flag now defaults to on when the user has never chosen, with an explicit stored choice still winning in both directions so anyone who turned it off keeps it off. This is a deliberate acceptance of risk: the flag existed because the native path was an unfinished spike, and `VideoPlayer.scrubRegression.test.ts` documents its history — a native init that flipped to HTML5 mid-lifecycle and left seeks going down one path while ExoPlayer played on another. Those tests pin the **flag-off** interim override (native response overridden to HTML5, backend stopped once), which the default no longer selects, so they now mock the flag off rather than inherit it: they still guard that path, but they no longer describe what ships. The native scrub/seek path is consequently not covered by the suite and needs device verification | UI | UR-041, UR-003 | Needs device verification |
+| DR-159 | The background-audio handoff stops leaking its relative timeline. The handoff plays the episode as a *relative* stream — the audio-only URL is built with `StartTimeTicks` = the position the screen was locked at, so ExoPlayer's zero is the handoff point — and `background_audio_base` holds the offset that turns one back into a real position. The base was a **display-only** correction, applied in exactly two places (the lockscreen scrubber and the internal truncation maths) while every other consumer worked in the relative timeline treating the number as absolute. Each crossing threw away exactly `base` seconds, which is why the jump-back distance varied with where the screen was locked and read as random. Three crossings were live: progress reporting to Jellyfin sent the relative position every 30s, so the server was told `real − base` — and since DR-155 now mirrors the server's position back and refreshes on a cache hit, that regressed value returned as the resume point (lock at 40 min, listen to 90, reopen at 50); lockscreen seeks went out absolute and came back relative, against a chunked length-less transcode that cannot honour a seek at all, so a clamped seek landed at stream zero; and media3's own `seekToDefaultPosition`/`seekBack`/`seekForward` bypassed the `ForwardingPlayer` wrapper entirely, reaching the real ExoPlayer — `Util.handlePlayButtonAction` seeking an ended player to the relative zero being the same mechanism as DR-129's truncation bug through a different door. The fix converts **once, at the boundary**: `JellyTauPlayer`'s position tick adds the base (and shifts the duration with it, since the stream's own length is only what remains) before either `nativeOnPositionUpdate` or the lockscreen sees it, so position updates, progress reports, the frontend and the truncation check all speak the episode's timeline and none needs to know a handoff happened. The base is consequently *removed* from `claim_stream_resume`, `truncated_stream_resume_position` and `player_exit_background_audio`, where adding it now double-counts, and the lockscreen's `positionOffsetMs` addition goes with it (the field remains, read-only, as the tick's input). Inbound seeks go the other way: `seek_absolute` is the new boundary for every outside seek, re-opening the stream at the requested position via `resume_stream_at` when a handoff is active — which is what `onSeekTo` had claimed for months in a comment describing code that did not exist — and an ordinary seek otherwise. `seekToDefaultPosition` is swallowed rather than forwarded, since Rust already owns what "play after the stream ended" means and the `play()` that follows reaches it. Exit reads the position *before* clearing either base, or a tick landing in between hands back a relative one | Player | UR-040, UR-005, UR-025 | Done (pending device verification) |
+| DR-158 | A watched toggle, on the episode row, the season header, the series and movie hero, and the Episode Focus View. Both halves of the backend already existed and neither had a caller: `mark_played` (`POST /PlayedItems`) was reachable only from the sync drain replaying rows the *reporter* had queued, and `clear_watch_history` (`DELETE /PlayedItems`) only from the destructive "erase this series' history" button — so the sole way to mark something watched was to play it. Jellyfin applies both recursively over a season or series, so the container case needs no client-side fan-out *online*. Offline it does: `storage_set_watched` writes the item **and its descendants** (drawn from `items` by `parent_id`/`album_id`/`season_id`/`series_id`, so an uncached id selects nothing and the statement no-ops instead of raising a foreign-key error), because otherwise marking a season watched with no server would tick the season and leave every episode inside it unwatched. It is deliberately separate from `storage_mark_played`, which stays the single-item "this finished playing" path that increments `play_count`. Un-marking clears the resume position as well as the flag, matching the server. `QueuedOp::MarkUnplayed` gives the queue the missing direction — pushing as `clear_watch_history` — so the toggle works offline both ways rather than only one; without it un-marking would have been the half that needed a connection. The button is an everyday toggle, so unlike `ClearHistoryButton` it does not confirm, and it holds an optimistic state because the caller's `watched` prop only catches up after a reload (a season means a round trip, during which the button would otherwise appear to ignore the tap) | UI | UR-073 | Done |
+| DR-157 | Full-screen video on Android actually goes full screen. `toggleFullscreen` called `document.documentElement.requestFullscreen()` and nothing else, which inside an Android WebView does not touch the Activity window — it expands the element within a viewport that already spans the whole screen, because `enableEdgeToEdge()` is called in `onCreate` and SDK 36 ignores the opt-out. So the control did nothing visible while the status bar and navigation/gesture bar stayed painted over the video, and (unlike DR-112's chrome-clearance work, which is about *reserving* space for the bars) here the bars should not be there at all. `ImmersiveModeBridge` hides them via `WindowInsetsControllerCompat` with `BEHAVIOR_SHOW_TRANSIENT_BARS_BY_SWIPE`, so an edge swipe brings them back transiently over the video instead of resizing the window mid-playback, and the system's own gestures stay reachable. Exposed as the `AndroidImmersive` bridge and posted to the main thread, since `@JavascriptInterface` methods arrive on a WebView binder thread. `requestFullscreen()` is kept for the platforms where it does work, but its rejection is caught rather than allowed to abort the immersive call. Restoring is wired to three paths, not one: leaving fullscreen, Escape (which previously called `document.exitFullscreen()` directly, bypassing the flag and the bars), and `onDestroy` — the bars belong to the Activity, so a player torn down while immersive would strand every screen behind it without them. The `--jt-inset-*` properties need no special handling: hiding the bars fires the decor view's inset listener with zeroes and `WindowInsetsBridge` republishes them | UI | UR-066 | Done |
+| DR-143 | Flipping the offline downloaded-only gate actually re-queries the listing. The gate (DR-078) is a process-wide flag in Rust consulted only *while a query runs*, but no library surface re-queried when its inputs changed: `useServerReachabilityReload` fires only on the offline → **online** transition, and `GenericMediaListPage`, `GenericGenreBrowser` and the favourites page never even called its `checkServerReachability`. So going offline left the full server catalog on screen under a now-closed gate, and toggling "Show all server media" only greyed cards — `MediaCard.isServerOnly` is a pure frontend derivation that updates instantly — without adding or removing a single row. The filter therefore read as "shows everything until I filter, then greys some of it" while the backend gate was correct and simply never exercised. `catalogFilterVersion` is the refetch signal: `pushCatalogVisibility` now awaits `set_show_server_catalog` and bumps the version only **after** the backend accepts the new flag, since a reload racing the push would re-query under the old gate and undo itself. A failed push clears `lastIncludeCatalog` instead of latching it, so the next identical transition is retried rather than skipped as a no-op and left permanently disagreeing with the backend. `useOfflineFilterReload` subscribes pages to that signal, skipping the value they already loaded under; it is wired into both generic list components and the movies/music/tv/favourites landing pages and the `/library/[id]` detail page | UI | UR-052 | Done |
+| DR-135 | A download's media type comes from the item, not a default. `download_item` — the path a media card uses to queue an item while offline — never records `media_type`, and the reconnect resolver read that NULL as `'audio'`, so a **movie** queued from a card had its URL resolved by `get_audio_stream_url`. The file that landed on disk was an audio-only transcode, which is why a "downloaded" film could never play offline no matter how the path or protocol was fixed. The resolver now falls back to the item's own `item_type` (`VIDEO_ITEM_TYPES` in Rust, so the frontend never learns which types are video) and only defaults to audio when the item is not cached locally. An explicit `media_type` on the row still wins | Downloads | UR-071, UR-052 | Done |
+| DR-136 | Rows already downloaded under the audio default are repaired, not just prevented. They are identifiable after the fact — no `media_type`, but a video item — so on reconnect they are reset to `pending` with their audio URL cleared and re-resolved by DR-135's corrected logic, overwriting the audio file in place. Without this the fix is invisible to anyone who had already queued a film: the row still reads "downloaded" and still fails to play. Rows carrying an explicit `media_type` and genuine audio downloads are left untouched | Downloads | UR-071 | Done |
+| DR-137 | Local media is served to the player over a loopback HTTP server, not the asset protocol. Tauri's `asset` protocol answers a request carrying no `Range` header by reading the whole file into memory, and only advertises `Accept-Ranges: bytes` from *inside* its range branch — so the first request never learns ranges exist and a multi-gigabyte body is attempted instead. Chromium abandoned it with `PIPELINE_ERROR_READ` after ~31s, which reached the user as "downloaded video does not play offline". Real HTTP on `127.0.0.1` is chosen over a custom URI scheme deliberately: range support becomes a property of the transport rather than depending on whether a platform's webview forwards `Range` to a custom scheme. No response ever exceeds a 4 MiB chunk and bodies stream from the file handle, so memory is bounded regardless of file size. Because **loopback is shared between apps on Android**, the server binds `127.0.0.1` only and every URL carries a random per-session token; paths are additionally confined to the app data directory, so a leaked URL cannot read outside it. This is stage 1 of making the server the single media origin — remote passthrough and download-while-watching are deliberately out of scope here | Playback | UR-071 | Done |
+| DR-138 | Loopback is exempted from Android's cleartext ban, and nothing else is. Release builds set `usesCleartextTraffic="false"`, so the webview's request to the local media server (DR-137) was rejected by network security policy before any I/O — `<video>` failed in the same millisecond as `loadstart`, with `NETWORK_NO_SOURCE` and no server-side log at all, which is why it looked identical to a missing file. A `network-security-config` resource permits cleartext for `127.0.0.1` only and keeps `base-config cleartextTrafficPermitted="false"`, so a remote server must still be HTTPS; this is deliberately not a blanket opt-in. The manifest attribute is ignored once the config is present, so the config is the single authority. `sync-android-sources.sh` also had to learn to copy `res/xml`, which it skipped — the manifest references the resource, so a missed copy fails the resource link rather than degrading quietly | Security | UR-071 | Done |
+| DR-093 | Traceability coverage gate derives its requirement denominators from `requirements.md` at run time rather than hardcoded literals: `countDefinedRequirements` counts an ID only where it leads a markdown table row (ignoring the "Traces To" column and prose) and deduplicates IDs listed both in the definition tables and in the §3 traceability matrix; `computeCoverage` reports the *intersection* of traced and defined IDs so an ID traced in code but absent from `requirements.md` is surfaced as `orphaned` instead of inflating the ratio past 100%. UT/IT test identifiers are excluded as a separate taxonomy. CI and `bun run traces:coverage` share this computation and fail on both a sub-threshold and an impossible >100% result | Tooling | - | Done |
 
 ---
 
@@ -248,9 +374,9 @@ Internal architecture, components, and application logic.
 |----------|-------------------------|-------------------------|
 | UR-001 | IR-001, IR-002 | - |
 | UR-002 | IR-013 | DR-003, DR-012, DR-013, DR-014 |
-| UR-003 | IR-003, IR-004, IR-011 | DR-002, DR-004, DR-010 |
-| UR-004 | IR-003, IR-004, IR-008, IR-011 | DR-002, DR-004, DR-006 |
-| UR-005 | - | DR-001, DR-005, DR-009 |
+| UR-003 | IR-003, IR-004, IR-011 | DR-002, DR-004, DR-010, DR-182, DR-183, DR-184, DR-185, DR-186, DR-187, DR-188, DR-190, DR-191, DR-192, DR-193, DR-194, DR-195 |
+| UR-004 | IR-003, IR-004, IR-008, IR-011 | DR-002, DR-004, DR-006, DR-129, DR-171, DR-176, DR-177, DR-181, DR-182, DR-183, DR-185, DR-188 |
+| UR-005 | - | DR-001, DR-005, DR-009, DR-178, DR-179, DR-186, DR-193, DR-195 |
 | UR-006 | IR-005, IR-006, IR-007, IR-008 | - |
 | UR-007 | IR-010 | DR-007, DR-008, DR-016 |
 | UR-008 | IR-010 | DR-007, DR-011 |
@@ -263,14 +389,14 @@ Internal architecture, components, and application logic.
 | UR-015 | - | DR-005, DR-020 |
 | UR-016 | - | - |
 | UR-017 | - | DR-014, DR-021 |
-| UR-018 | IR-013 | DR-015, DR-018 |
+| UR-018 | IR-013 | DR-015, DR-018, DR-173 |
 | UR-019 | IR-015 | DR-022 |
-| UR-020 | IR-016, IR-018 | DR-023 |
+| UR-020 | IR-016, IR-018 | DR-023, DR-176 |
 | UR-021 | IR-016, IR-019 | DR-024 |
 | UR-022 | IR-017 | DR-025 |
 | UR-023 | IR-010 | DR-026, DR-047, DR-048, DR-049 |
 | UR-024 | IR-010 | DR-027 |
-| UR-025 | IR-015 | DR-028 |
+| UR-025 | IR-015 | DR-028, DR-131, DR-132, DR-178, DR-179 |
 | UR-026 | - | DR-029, DR-048, DR-050 |
 | UR-027 | IR-020 | DR-030 |
 | UR-028 | - | DR-031 |
@@ -285,24 +411,41 @@ Internal architecture, components, and application logic.
 | UR-037 | IR-010 | DR-042 |
 | UR-038 | IR-010 | DR-043 |
 | UR-039 | - | DR-045, DR-046 |
-| UR-040 | IR-025 | DR-051, DR-052 |
-| UR-041 | IR-026 | DR-053 |
+| UR-040 | IR-025 | DR-051, DR-052, DR-129, DR-130, DR-159, DR-178, DR-179, DR-180, DR-183, DR-190 |
+| UR-041 | IR-026 | DR-053, DR-160, DR-161, DR-172, DR-182, DR-183, DR-184, DR-185, DR-188 |
 | UR-042 | IR-009, IR-014 | DR-054 |
 | UR-043 | IR-027 | DR-055 |
 | UR-044 | - | DR-056 |
 | UR-045 | - | DR-057 |
 | UR-046 | IR-028 | DR-058 |
 | UR-047 | IR-013 | DR-060 |
-| UR-048 | - | DR-061, DR-062 |
-| UR-049 | IR-010 | DR-063, DR-064, DR-065 |
+| UR-048 | - | DR-061, DR-062, DR-142 |
+| UR-049 | IR-010 | DR-063, DR-064, DR-065, DR-147 |
 | UR-050 | - | DR-066, DR-067 |
 | UR-051 | - | DR-068, DR-069, DR-070 |
-| UR-052 | IR-027 | DR-078, DR-079, DR-080 |
+| UR-052 | IR-027 | DR-078, DR-079, DR-080, DR-143 |
 | UR-053 | IR-029 | DR-074 |
-| UR-054 | - | DR-075, DR-076, DR-077 |
-| UR-055 | - | DR-081, DR-082, DR-083, DR-084 |
+| UR-054 | - | DR-075, DR-076, DR-077, DR-147 |
+| UR-055 | - | DR-081, DR-082, DR-083, DR-084, DR-167, DR-168, DR-169, DR-173 |
 | UR-056 | - | DR-085 |
 | UR-057 | - | DR-086 |
+| UR-058 | - | DR-087, DR-142 |
+| UR-060 | - | DR-090, DR-091, DR-111 |
+| UR-061 | - | DR-092 |
+| UR-062 | - | DR-101, DR-102, DR-103, DR-104, DR-107 |
+| UR-063 | - | DR-105 |
+| UR-064 | - | DR-106 |
+| UR-065 | IR-030 | DR-108, DR-109, DR-110, DR-111 |
+| UR-066 | IR-031 | DR-112, DR-157, DR-187, DR-194 |
+| UR-067 | - | DR-115, DR-116, DR-117, DR-118 |
+| UR-068 | - | DR-119 |
+| UR-069 | - | DR-113, DR-114, DR-120 |
+| UR-070 | - | DR-121, DR-122 |
+| UR-071 | IR-032 | DR-123, DR-124, DR-125, DR-126, DR-127, DR-128, DR-133, DR-134, DR-135, DR-136, DR-137, DR-138, DR-170, DR-171, DR-180 |
+| UR-072 | - | DR-156 |
+| UR-073 | - | DR-158 |
+| UR-074 | - | DR-162, DR-177, DR-181 |
+| UR-075 | - | DR-174, DR-175 |
 
 ---
 
@@ -373,11 +516,127 @@ Internal architecture, components, and application logic.
 | UT-059 | Audio-only stream URL builder for a video item (selected audio-stream index) | JA-032, DR-052 | Pending |
 | UT-060 | Background-audio handoff state machine (background→audio, foreground→video; no dual audio) | DR-052 | Pending |
 | UT-061 | Background-audio Tauri command param naming (camelCase) | DR-052 | Pending |
+| UT-062 | `setBackgroundAudioEnabled` reports whether the native bridge was actually reached (missing bridge, stale proxy, throwing method) so a dead bridge cannot look armed | UR-040, IR-025, DR-051 | Done |
 | UT-067 | Offline `get_items` gates the synced-catalog UNION on the catalog-browse flag (downloads only when off, full catalog when on) | DR-078 | Done |
-| UT-068 | Catalog visibility resolves to `serverReachable \|\| showServerCatalog`, and is pushed to the backend on every change of either input | DR-078, DR-079 | Pending |
-| UT-069 | `isConnected` follows backend reachability alone: false when the server is unreachable on a live link, true for a reachable server while `navigator.onLine` is false | DR-079 | Pending |
-| UT-070 | Hybrid `get_items` returns an empty offline result as-is when the catalog-browse gate is off, without querying the server | DR-080 | Pending |
+| UT-068 | Catalog visibility resolves to `serverReachable \|\| showServerCatalog`, and is pushed to the backend on every change of either input | DR-078, DR-079 | Done |
+| UT-069 | `isConnected` follows backend reachability alone: false when the server is unreachable on a live link, true for a reachable server while `navigator.onLine` is false | DR-079 | Done |
+| UT-070 | Hybrid `get_items` returns an empty offline result as-is when the catalog-browse gate is off, without querying the server | DR-080 | Done |
 | UT-066 | WiFi-only download gate: cellular and metered WiFi blocked, unmetered WiFi/Ethernet allowed, unknown/none fail closed, desktop default ungated; plus the frontend network reporter (transport reporting, change subscription, teardown, fail-open queries) | DR-074 | Done |
+| UT-071 | Byte-size formatter: zero/negative/non-finite → "0 B"; decimal unit thresholds; 2–3 significant-figure banding; trailing-zero trimming; largest-unit cap | DR-085 | Done |
+| UT-072 | Downloaded-only browse returns a downloaded leaf and its container, filtered to the requested album parent; a non-downloaded sibling is omitted | DR-082, DR-083 | Done |
+| UT-073 | An empty downloaded-only browse is authoritative — no rows, no error — regardless of the catalog-browse flag | DR-082 | Done |
+| UT-074 | Only libraries with downloaded content are listed; an empty one is omitted | DR-082 | Done |
+| UT-075 | Disk usage reports a leaf's own size, a container's summed descendants, and reconciles the device total with the sum of leaves | DR-085 | Done |
+| UT-076 | Downloaded library browse lists album containers, not their individual tracks; drilling into the album returns the tracks | DR-082, DR-083 | Done |
+| UT-077 | Downloaded TV library browse lists the series, not seasons/episodes; drilling returns the season then the episode | DR-082, DR-083 | Done |
+| UT-078 | A downloaded leaf with no cached container (e.g. a movie) still surfaces at the library level | DR-082, DR-083 | Done |
+| UT-079 | Each EQ preset returns a 10-band gain curve within range; Flat is all zeros; Bass Boost lifts lows and leaves highs flat | DR-030 | Done |
+| UT-080 | `with_equalizer_normalised` clamps out-of-range gains and forces the band vector to exactly 10 entries (pad short, truncate long) | DR-030 | Done |
+| UT-081 | Old persisted AudioSettings JSON without EQ fields loads as disabled + flat | DR-030 | Done |
+| UT-082 | EQ fields serialize as camelCase (`equalizerEnabled`/`equalizerBands`) and round-trip | DR-030 | Done |
+| UT-083 | EQ filter entries are empty when disabled or when the curve is flat (clears the `af` filter) | IR-020 | Done |
+| UT-084 | Enabled EQ builds one peaking `equalizer` per non-zero band at the right frequency and gain inside a single `lavfi` chain | IR-020 | Done |
+| UT-085 | A first tap resolves to `togglePlayPause` immediately — no deferral and no timer | DR-092, DR-098 | Done |
+| UT-086 | A second tap inside the window seeks (+30 s right half, −10 s left half) with the matching feedback side **and** re-toggles play/pause, so the two toggles cancel and the play state is unchanged by a double tap | DR-092, DR-098 | Done |
+| UT-087 | A tap after the window, and the tap following a consumed pair, are each fresh first taps that toggle (there is no third-tap case); repeated double taps keep seeking; `cancel()` makes the next tap a first tap so an interpreted swipe cannot seek | DR-092, DR-098 | Done |
+| UT-088 | `resolveSeekTarget` applies the delta to the reported position, clamps into `[0, duration - END_SEEK_MARGIN_SECONDS]`, chains off an in-flight pending target so rapid skips accumulate, and ignores that target once the player reports past it | DR-092, DR-095 | Done |
+| UT-089 | A touch drag on the video seek bar seeks to the dragged position, never toggles play/pause, and never alters brightness — the container gesture layer stays out of a control drag entirely | DR-098, DR-099 | Done |
+| UT-090 | The seek bar commits its seek on `touchend` even when the engine never fires `change`, and commits exactly once when both signals arrive | DR-099 | Done |
+| UT-091 | Transport intents (play/pause/toggle) reach the backend even while a video adapter is registered, and never call the adapter's own `play`/`pause`/`toggle` — the webview must not decide play-vs-pause from the DOM | DR-097 | Done |
+| UT-092 | `shouldReuseActivePlayback` reuses backend playback for an already-loaded audio track but never for video, and never when an explicit start position or a next-episode restart was requested | DR-100 | Done |
+| UT-093 | `resolvePlayerSurface` returns `video` only with a stream URL, `pending` for video whose stream URL is still missing (never `audio`), and `audio` for audio content | DR-100 | Done |
+| UT-094 | `parseNativeInsets` accepts the bridge's JSON or a decoded object, and coerces missing/negative/non-finite edges to 0 rather than emitting `NaNpx` (which would invalidate the whole padding declaration) | DR-112 | Done |
+| UT-095 | `safeAreaCssVars`/`applySafeAreaInsets` emit px-suffixed `jt-inset` custom properties for all four edges | DR-112 | Done |
+| UT-096 | `readNativeInsets` returns null with no bridge and survives a stale WebView proxy (missing or throwing `get`) instead of throwing out of layout init | IR-031, DR-112 | Done |
+| UT-097 | `initSafeArea` primes the document on start, re-applies on `jellytau-insets-changed` (rotation, nav-mode switch), unsubscribes on teardown, and writes nothing without a bridge so `env()` still wins on iOS/desktop | IR-031, DR-112 | Done |
+| UT-098 | `shellReservesBottomInset` gives the bottom inset to BottomUi wherever one renders and to the app shell only on routes without one, so the gesture bar is never ignored nor double-padded | DR-112 | Done |
+| UT-099 | A Jellyfin item payload carrying `UserData.IsFavorite` maps to `MediaItem.user_data.is_favorite` | DR-113, JA-034 | Done |
+| UT-100 | `OnlineRepository::get_favorites` builds `Filters=IsFavorite` + `Recursive=true` + the scope's `IncludeItemTypes`, and omits the type filter entirely for `SearchScope::All` | DR-115, JA-033 | Done |
+| UT-101 | `OfflineRepository::get_favorites` returns only `is_favorite = 1` rows, honours the scope type filter, and stays downloads-only when the catalog-browse gate is off | DR-115 | Done |
+| UT-102 | The `save_to_cache` favourite mirror does not overwrite a row with `pending_sync = 1` | DR-114 | Done |
+| UT-103 | The reconnect drain pushes pending favourites, clears `pending_sync`, and leaves failed rows pending | DR-120 | Done |
+| UT-104 | `get_items` with `favorites_only` filters online (endpoint) and offline (SQL) | DR-116 | Done |
+| UT-105 | `favorites` store precedence: override beats `userData.isFavorite` beats `false` | DR-119 | Done |
+| UT-106 | Un-favouriting removes an item from a favourites listing view | DR-117, DR-119 | Done |
+| UT-107 | The hybrid background refresh emits `favorites-changed` only for ids whose favourite state actually flipped | DR-120 | Done |
+| UT-109 | Search covers synced-but-not-downloaded items when catalog browse is on, and stays downloads-only when off | DR-108 | Done |
+| UT-110 | Search item-type filter is bound, not interpolated: a quote-bearing type neither errors nor widens results | DR-108 | Done |
+| UT-111 | FTS prefix queries quote each token, so apostrophes/hyphens/slashes are data; empty or punctuation-only input returns no rows rather than erroring | DR-108 | Done |
+| UT-112 | Repeated catalog passes leave one `items_fts` entry per item, not one per pass | DR-110 | Done |
+| UT-113 | The stale-catalog sweep removes vanished synced rows, keeps downloaded ones, keeps uncrawled types, and stays scoped to one server | DR-110 | Done |
+| UT-114 | Cached people are reachable from unscoped search and excluded from scoped search | DR-111 | Done |
+| UT-115 | Re-index staleness policy: never-indexed and unparseable timestamps are due, fresh ones are not, future ones are not | DR-109 | Done |
+| UT-116 | `resolve_local_media_path` returns a completed download's file, and `None` for an in-progress download, a row whose file has been deleted, or an unknown item | DR-123 | Done |
+| UT-118 | `resolveVideoSource` prefers a downloaded file, never marks a local file as needing transcoding, and falls back to streaming for a blank path | DR-123 | Done |
+| UT-119 | The audio-only handoff picks a downloaded file over the audio-only stream URL, preserving the Jellyfin id for progress sync | DR-128 | Done |
+| UT-120 | Expiry reclaim takes only expired temporary entries: derived from `completed_at`+TTL, honouring an `expires_at` override, never a user download, and disabled by a zero TTL | DR-127 | Done |
+| UT-108 | LRU eviction reclaims only `'auto'` downloads and never a user's own, even when the user's is the oldest | DR-126 | Done |
+| UT-117 | A background audio-only stream cut short resumes where it died instead of ending the episode; a real end still advances; the absolute position is compared against the runtime; retries at a stuck position give up. A recoverable error resumes music and video too, with growing backoff, leaving the rest of the queue intact and the seekable stream's URL untouched; local and DirectUrl sources are excluded | DR-129 | Done |
+| UT-124 | `downloadedFilePath` leaves a completed download's absolute path alone (POSIX and Windows) and only roots one that is still relative | DR-133 | Done |
+| UT-125 | A NULL `media_type` resolves from the item type — Movie and Episode as video, a track as audio — an uncached item still defaults to audio, and an explicit `media_type` overrides the item | DR-135 | Done |
+| UT-126 | Requeueing takes only video rows downloaded under the audio default, clearing their URL, and leaves correctly-typed video rows and real audio downloads alone | DR-136 | Done |
+| UT-127 | The media server bounds and confines every response: a range-less request yields one chunk rather than the whole file, no range exceeds the chunk cap, explicit/open-ended/suffix ranges resolve correctly, a range past the end is unsatisfiable rather than clamped, a malformed header falls back to the first chunk, path traversal and unrelated absolute paths are refused, a wrong or absent token is rejected, and content type comes from the extension then the magic bytes | DR-137 | Done |
+| UT-121 | An EOF reads as the last observed timestamp, not zero: live readings win while the file is loaded, a not-yet-established duration is not recorded as a real zero, a seek updates the position before the next poll, and loading a new file clears the previous one's | DR-130 | Done |
+| UT-122 | The sync-queue drain pushes queued playback reports oldest-first, defers failures for the next reconnect, abandons a row after `MAX_SYNC_ATTEMPTS`, ignores other users' rows, and parses both payload dialects | DR-131 | Done |
+| UT-123 | Pending-sync rows describe themselves: every queueable operation has a label, an unknown one still renders, the item title falls back to its id, and rows list oldest-first | DR-132 | Done |
+| UT-130 | Video and background-audio stream URLs omit `AudioStreamIndex` when no track was chosen, and carry the exact index when one was | DR-140 | Done |
+| UT-131 | The Episode Focus View hero offers a download control | DR-142 | Done |
+| UT-132 | The series name links to the series and the `SxEy` badge to that season's anchor | DR-142 | Done |
+| UT-133 | Cast renders below the "More Episodes" strip, never above it | DR-062, DR-142 | Done |
+| UT-134 | The episode strip is hidden when the episode has no siblings | DR-142 | Done |
+| UT-135 | An episode with no `seriesId` still renders the Focus View, with title, Play and download | DR-142 | Done |
+| UT-136 | `episodeRedirectTarget` sends a bare episode page into its series' Focus View, and returns null with no series | DR-142 | Done |
+| UT-137 | Going offline with the toggle off pushes the closed gate and bumps `catalogFilterVersion` | DR-143 | Done |
+| UT-138 | The version bumps only after `set_show_server_catalog` resolves, never before | DR-143 | Done |
+| UT-139 | A failed visibility push is retried on the next identical transition rather than latched | DR-143 | Done |
+| UT-140 | `useOfflineFilterReload` skips the value a page already loaded under and reloads on each later change | DR-143 | Done |
+| UT-141 | The advertised channel cap: an unknown or zero reading falls back to stereo, a real route keeps its channels, an absurd driver reading is capped at 7.1, and mono is taken at its word | DR-141 | Done |
+| UT-148 | Forcing a transcode from the client: an undecodable default track forces one, a decodable track does not, the default track decides rather than the first, the first decides when nothing is marked default, and neither an audio-less source nor an unnamed codec is second-guessed | DR-149 | Done |
+| UT-149 | `createAdapter` returns the native adapter only when Rust reports native AND `experimentalNativeVideo` is on; the flag off forces HTML5 even when Rust says native, and the flag on never promotes a platform Rust reported as HTML5 | DR-150 | Done |
+| UT-150 | `set-version.sh` stamps all four manifests without touching dependency versions, and the Android versionCode is monotonic across an upgrade sequence, clears the 1000 floor, and survives a prerelease suffix | DR-153 | Done |
+| UT-151 | An unreportable stop lands in the queue and is pushed by the existing drain; re-queueing the same item supersedes the earlier position rather than adding a row, distinct items keep their own positions, and an abandoned row is not revived by a later report | DR-154 | Done |
+| UT-152 | Caching a server result mirrors its watch position locally — including for an item carrying a position but no favourite flag — without inventing a row for an item the server reported no user data for, and without pulling a still-unsynced local position backwards | DR-155 | Done |
+| UT-162 | Each downloaded library lists only its own media: the music library shows the album and neither the film nor the series, the movie library only the film, the TV library only the series | DR-163 | Done |
+| UT-163 | `partial_path` appends rather than replacing the extension, so it matches what the cleanup paths delete, keeps two sources for one title apart, and still produces a sidecar for an extension-less target | DR-165 | Done |
+| UT-170 | `queue_album_tracks` queues a row for every track of the album — including tracks the cache holds without an `album_id` and tracks it has never seen at all — links each one to its album so offline browsing can find it, returns the row ids in track order, and is idempotent: re-queuing fills the gaps without duplicating rows or resetting a completed track. `cached_album_tracks` (the offline fallback) finds tracks by either album link and does not sweep in another album's | DR-173 | Done |
+| UT-171 | `resolve_pending_download_urls` restricted to a set of row ids resolves only those rows and leaves other pending rows untouched, and an empty id set resolves nothing rather than sweeping everything | DR-173 | Done |
+| UT-172 | `album_file_names` gives every track of an album its own file: a title repeated within the album (deluxe edition, two discs) is disambiguated by track number and item id instead of the second download overwriting the first, an unambiguous title keeps its own name, and path separators in a title are sanitised so a track cannot escape the album directory | DR-173 | Done |
+| UT-164 | `resume_offset` appends only when the server answered `206`; a `200` after a Range request restarts the file, because that body is the whole stream | DR-166 | Done |
+| UT-165 | A registered download starts unflagged, `signal` sets the flag its worker reads, signalling an unregistered id reports not-in-flight, `clear` forgets it, and re-registering drops a previous stop so a resumed download does not halt instantly | DR-164 | Done |
+| UT-166 | `original` quality re-encodes audio the webview cannot decode (E-AC-3/AC-3/DTS/TrueHD) to AAC without capping bitrate or resolution, keeps the `Static=true` direct copy for audio that plays here (AAC/MP3/Opus/Vorbis/FLAC) and for an unknown codec, leaves the explicit quality presets untouched, and picks the served track by the same default-or-first rule the streaming verdict uses | DR-171 | Done |
+| UT-155 | A seek during a background-audio handoff re-opens the stream at the requested absolute position (`StartTimeTicks`) and rebases the handoff to it, while a seek outside a handoff stays an ordinary seek and invents no base | DR-159 | Done |
+| UT-154 | `mark_unplayed` parses to `QueuedOp::MarkUnplayed` and is rejected without an item id, and a queued un-mark drains to the server as `clear_watch_history` | DR-158 | Done |
+| UT-156 | A capped step reaches the transcode URL as all four of its parts (total ceiling, the video/audio split summing to the cap, and a `MaxHeight`), the uncapped default keeps the historical 20/18 Mbps allowance and constrains no resolution, and the background-audio handoff takes the lower of the cap and its own 384 kbps | DR-162 | Done |
+| UT-157 | The quality ladder is internally consistent — video + audio equals the cap at every step, audio never consumes the budget, only `Original` is uncapped — descends in bitrate, resolution and audio share together, and round-trips through the serde token it is persisted as | DR-162 | Done |
+| UT-158 | Justified rows fill the container width exactly and never overflow it, every tile in a row shares one height, and each tile's width follows its own aspect ratio — a 16:9 tile coming out more than twice the width of a 2:3 tile at the same height | DR-174 | Done |
+| UT-159 | The awkward cases of the packing: a short last row is left at the target height rather than stretched across the container, a last row that would overflow is brought down, an extreme ratio is clamped instead of taking a row to itself, a missing or nonsensical ratio falls back to square instead of collapsing the tile, an unmeasured container renders nothing rather than 1px tiles, and every tile is placed exactly once in order | DR-174 | Done |
+| UT-160 | The default row height suits its container: it grows with the width, stays inside its bounds, and at phone width still fits two 16:9 tiles side by side | DR-174 | Done |
+| UT-161 | A collection type maps to its favourites scope (`movies`/`tvshows`/`music`), every other kind — Live TV, channels, box sets, books, unknown — maps to none rather than to `All`, and a constructed library carries the scope across the wire as `favoritesScope`, omitted entirely when it has none | DR-175 | Done |
+| UT-167 | The mosaic's composition: the cross-library favourites entry leads, each library is followed by its own category tile pointing at that category's tab, a category shared by two libraries still yields one tile, a library kind favourites do not carve up yields none, a scope the page offers no tab for is ignored, and every tile is uniquely keyed | DR-174, DR-175 | Done |
+| UT-168 | Subtitles are negotiated as sidecars, never burned in: the requested `SubtitleStreamIndex` is the explicit "none" sentinel (`-1`) rather than omitted, every text format we can render (`srt`/`subrip`/`ass`/`ssa`/`vtt`) is advertised as `External`, and the burn-in verdict is by format — text never forces it, image formats (PGSSUB, dvdsub) always do, case-insensitively. The same sentinel rides the stream URL itself, so a stream re-opened without a fresh negotiation cannot inherit a subtitle. And the verdict reaches the picker: a subtitle stream carries `supportsExternalDelivery` — set only for subtitles, `false` for a bitmap format and for one the server left unnamed — which drops the tracks the app could never draw from the menu, the `<track>` children and the native play request alike, without even fetching their URLs, while a stream carrying no verdict at all is still offered | DR-176 | Done |
+| UT-173 | Every video stream URL carries a `PlaySessionId`, each open mints a fresh one, and the open reports the session it superseded so that job can be stopped | DR-177 | Done |
+| UT-174 | A fatal HLS network error is read against the *absolute* position: mid-film — including after a quality switch, where the seek offset carries the whole resume position — it is retried rather than reported as the end of the stream, the last tenth of a known runtime is treated as the end, an unknown runtime retries, and retries stop once the budget is spent | DR-177 | Done |
+| UT-175 | A stream reload that never becomes playable is reported as a failure instead of resolving as success, so the caller can revert its selection rather than leave the UI claiming a stream that is not playing | DR-177 | Done |
+| UT-176 | A handoff's position is floored at its base: with no tick yet landed the exit position is the point the screen was locked at rather than 0, and once ticks are flowing (the base already applied natively) it is not added twice | DR-178 | Done |
+| UT-177 | Webview-rendered media's reported position and duration are the controller's, and are dropped the moment that element stops being the player — on teardown, and when a handoff takes over | DR-178 | Done |
+| UT-178 | A stop report at position 0 is withheld rather than sent (it would clear the resume point), while a real position is still reported from either rendering path — the element's on the webview path, the backend's on the native one | DR-179 | Done |
+| UT-179 | An audio-only episode that ends naturally is reported stopped at its runtime, so Jellyfin marks it played; a truncated stream, which is about to be re-opened, reports nothing | DR-179 | Done |
+| UT-180 | Position ticks report progress to the server, throttled to one report per item per window rather than one per tick | DR-179 | Done |
+| UT-181 | The handoff plan matches its source: a downloaded file takes no base and a seek, a stream takes the base and no seek, and a handoff at 0:00 takes neither; a downloaded handoff's absolute seek stays an ordinary seek instead of a stream rebuild | DR-180 | Done |
+| UT-153 | Scroll handling per navigation kind: a forward move always lands at the top even when the previous page was scrolled and even when the target was visited before, Back restores that route's own saved offset (and the top when it has none), offsets are kept per route rather than shared, a repeated Back still restores, and the initial load leaves the container alone | DR-156 | Done |
+| UT-142 | The audio codecs offered for video direct play: a Dolby device's real `MediaCodecList` output drops `ac3`/`eac3`, AMR and raw PCM are dropped too, a fully-supported list is passed through untouched, a list with nothing decodable still claims `aac`, and stray spacing or casing does not decide whether the user gets sound | DR-148 | Done |
+| UT-143 | Subtitle URLs resolve to plain strings before they reach the markup (never a Promise), unresolvable tracks are dropped, a stale selection collapses to "Off", and a server-default track is never auto-selected | UR-020, DR-023 | Done |
+| UT-144 | VideoPlayer actually renders `<track kind="subtitles">` children carrying `data-stream-index`, with no `default` attribute and no async `getSubtitleUrl()` bound to `src` | UR-020, DR-023 | Done |
+| UT-145 | The frontend's subtitle payload survives the IPC hop: a camelCase `PlayItemRequest` carrying `subtitles` deserializes, `create_media_item` lands them on `MediaItem.subtitles` in the order sent, and a request without the field still defaults to empty | UR-020, IR-016 | Done |
+| UT-146 | The subtitle JSON serialized across the JNI boundary uses the keys `JellyTauPlayer.load()` reads — `url`, `language`, `label` and `mime_type`, never `mimeType` | UR-020, IR-016, JA-008 | Done |
+| UT-147 | The native subtitle payload and the track-selection index come from the same resolved list: the wire shape keeps `mime_type` and stream order, `playerPlayItem` actually sends it, and the index is a position in the sent list (so a track whose URL failed to resolve cannot shift the others) rather than the menu's row number | UR-020, IR-016 | Done |
+| UT-182 | An HLS video URL never carries `StartTimeTicks` — with a position supplied or not — while the master playlist, codec, media source and chosen audio track still ride on it | DR-181 | Done |
+| UT-183 | A reloaded stream is resumed by seeking the element to the absolute position with the transcode offset cleared to zero — never by carrying the position as an offset base, which since DR-181 would display the position while playing the item from its start — and a reload to 0:00 waits for no seek | DR-181 | Done |
+| UT-184 | The native reveal rule fires on `state === "playing"` and on a position tick carrying a position or a duration, and on nothing else — not `buffering`, `paused`, `stopped`, `ended` or `error`, not an empty tick, and not a negative position | DR-182 | Done |
+| UT-189 | On the native path the player never calls `player_report_state` — driven through the real 10-second progress interval under fake timers, which is the call site that mattered; asserting on a freshly mounted player passes with the guard deleted and guards nothing | DR-195 | Done |
+| UT-187 | On the native path the play overlay follows the backend: it clears when the backend resumes after a pause and is raised again when the backend pauses, and the system bars are hidden on player entry rather than only by the fullscreen button | DR-186, DR-187 | Done |
+| UT-186 | Every attribute the native-video compositing block in app.css targets is set somewhere in the app — `[data-app-shell]` in particular — so a selector aimed at nothing fails the suite instead of failing silently on a device | DR-185 | Done |
+| UT-185 | Mounted on the native path (backend reports native, opt-in flag on, no `<video>` element rendered and the backend not stopped), VideoPlayer keeps the poster card up until the backend reports something, drops it on a playing state or a position tick with a duration, and keeps it up through `error` and `stopped` | DR-182 | Done |
 
 ### Integration Tests
 
@@ -396,8 +655,8 @@ Internal architecture, components, and application logic.
 | IT-011 | Resume playback from server position | IR-015, UR-019 | Pending |
 | IT-012 | Equalizer bands via libmpv | IR-020, UR-027 | Pending |
 | IT-013 | Background-audio handoff on Android: background/lock continues audio via native service and stops video decode; foreground resumes video at position | IR-025, UR-040 | Pending |
-| IT-016 | Offline library listing end-to-end: with the server unreachable, a library page lists only downloaded media with the toggle off, and additionally reveals greyed-out cached catalog entries with the toggle on | UR-052, DR-078, DR-079, DR-080 | Pending |
-| IT-017 | A download queued from a greyed-out offline catalog entry persists and is resolved and started on reconnect | UR-052, UR-011 | Pending |
+| IT-016 | Offline library listing end-to-end: with the server unreachable, a library page lists only downloaded media with the toggle off, and additionally reveals greyed-out cached catalog entries with the toggle on | UR-052, DR-078, DR-079, DR-080 | Done |
+| IT-017 | A download queued from a greyed-out offline catalog entry persists and is resolved and started on reconnect | UR-052, UR-011 | Done |
 
 ---
 
@@ -467,22 +726,36 @@ The `PlayerBackend` trait defines optional audio settings methods with default e
 | Basic playback | ✅ | ✅ | Parity |
 | Volume control | ✅ | ✅ | Parity |
 | Seek | ✅ | ✅ | Parity |
-| Crossfade | ✅ | ❌ | Gap |
-| Gapless playback | ✅ | ❌ | Gap |
-| Volume normalization | ✅ | ❌ | Gap |
+| Crossfade | ❌ | ❌ | Not implemented (blocked on MPV) |
+| Gapless playback | ✅ | ⚠️ | Implemented, pending on-device verification |
+| Volume normalization | ✅ | ⚠️ | Implemented (LoudnessEnhancer — gain stage, approximate vs MPV's dynaudnorm), pending on-device verification |
+| Equalizer (10-band) | ✅ | ⚠️ | Implemented (resampled onto device bands), pending on-device verification |
 | Position updates | 250ms | On-demand | Inconsistent |
 
-**Future Fix**:
-1. Implement `set_audio_settings()` in `ExoPlayerBackend`
-2. Add Kotlin-side ExoPlayer configuration for crossfade (using `ConcatenatingMediaSource` or `DefaultMediaSourceFactory`)
-3. Implement gapless via ExoPlayer's built-in gapless support
-4. Add volume normalization via ExoPlayer's `LoudnessEnhancer` or audio processor
-5. Standardize position update frequency across platforms
+**Status** (see docs/specs/android-audio-settings-parity.md):
+1. ✅ `set_audio_settings()` implemented in `ExoPlayerBackend` (JSON over JNI)
+2. ✅ Gapless via ExoPlayer's `pauseAtEndOfMediaItems`
+3. ✅ Volume normalization via `LoudnessEnhancer`
+4. ✅ Equalizer via `android.media.audiofx.Equalizer`, canonical 10 bands
+   resampled onto the device's band centres
+5. ⬜ **Not yet verified on a physical device** — the EQ/normalization effects
+   depend on device-specific `AudioEffect` availability and band layouts
+6. ⬜ Flip the trait's `set_audio_settings` default from `Ok(())` to
+   `Err(not_implemented())` so a backend that omits it fails loudly instead of
+   silently reporting success. Deferred until (5) confirms the Android path works
+7. ⬜ Standardize position update frequency across platforms
+
+Crossfade is deliberately absent: it is unimplemented on every platform and
+architecturally blocked on MPV, so building it on Android alone would invert the
+parity gap. (The previously suggested `ConcatenatingMediaSource` is also
+deprecated in current Media3.)
 
 **Impact**:
 - Medium - Android users lack audio enhancement features advertised in requirements
 - User experience differs between platforms
-- UR-031 (Crossfade), UR-032 (Gapless), UR-033 (Normalization) only work on Linux
+- UR-032 (Gapless), UR-033 (Normalization) and UR-027 (Equalizer) are now
+  implemented on Android as well as Linux, pending on-device verification
+- UR-031 (Crossfade) works nowhere — see DR-034
 
 **Traces To**: IR-004, UR-031, UR-032, UR-033, DR-034, DR-035, DR-036
 

docs/specs/android-audio-settings-parity.md

@@ -0,0 +1,209 @@
+# Spec: Android audio settings parity (EQ, normalization, gapless)
+
+**Status:** Proposed
+**Requirements:** UR-031, UR-032, UR-033, UR-027 → DR-034, DR-035, DR-036, DR-030; IR-004
+**UX spec:** n/a — no UI change; Settings › Audio already renders these controls
+**Supersedes / revises:** closes the audio half of the parity gap recorded in [playback-backend-unification.md](playback-backend-unification.md)
+
+## Summary
+
+Implement `set_audio_settings` / `audio_settings` on `ExoPlayerBackend` so the
+equalizer, volume normalization, and gapless playback settings actually take
+effect on Android. Today the Settings › Audio panel renders these controls on
+Android and they silently do nothing — `ExoPlayerBackend` is the only backend
+that does not override the trait's no-op defaults.
+
+Crossfade is explicitly **not** included; see Out of scope.
+
+## Motivation
+
+`PlayerBackend` declares `set_audio_settings` with a default `Ok(())` body.
+`MpvBackend`, `NullBackend`, and `WebviewAudioBackend` all override it;
+`ExoPlayerBackend` does not. The settings are persisted, pushed to the backend on
+every track load, and displayed in the UI — and then dropped on the floor.
+
+This is the single most user-visible platform divergence in the app: a user who
+sets a "Rock" EQ preset on Android sees the sliders move and hears no change.
+
+The backend-unification investigation ruled out fixing this by swapping engines
+(video cannot be unified; see the sibling spec), so the fix is to implement the
+trait methods where they are missing.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Band count, centre frequencies, gain range, preset→curve map | Rust (existing) | Already domain-owned in `settings.rs` per [audio-equalizer.md](audio-equalizer.md). Android must consume the same `AudioSettings`, not define its own bands. Duplicating the band layout in Kotlin would be a taxonomy leak of exactly the kind `check:boundary` guards against. |
+| Mapping `AudioSettings` → Android audio-effect parameters | Rust → JNI boundary | Platform playback detail, the direct analogue of `build_af_filter` in `mpv_backend.rs`. Belongs with the other `set_audio_settings` code. |
+| Attaching/detaching `Equalizer` and `LoudnessEnhancer` to the ExoPlayer audio session | Kotlin (`JellyTauPlayer.kt`) | Android platform API mechanics; needs the live `audioSessionId`, which only the Kotlin layer holds. |
+| Normalization preset (Loud/Normal/Quiet) → target gain | Rust (existing) | `VolumeLevel` is domain vocabulary; the same preset must mean the same loudness on every platform. |
+| Rendering sliders / preset chips | Frontend (existing) | Pure presentation; unchanged by this spec. |
+
+Borderline row: attaching the effects could arguably be driven entirely from
+Rust via JNI property calls. It goes to Kotlin because `AudioEffect` construction
+requires the audio session id and must be re-attached when ExoPlayer rebuilds its
+audio sink — lifecycle state that lives in `JellyTauPlayer.kt`. Rust still owns
+*what* the values are; Kotlin owns *when* the effect objects exist.
+
+## Design
+
+### Rust — `ExoPlayerBackend` (`src-tauri/src/player/android/mod.rs`)
+
+Override the two defaulted methods, mirroring the shape of the existing
+`set_audio_track` JNI call:
+
+```rust
+fn set_audio_settings(&mut self, settings: &AudioSettings) -> Result<(), PlayerError> {
+    let s = settings.clone().with_crossfade_clamped().with_equalizer_normalised();
+    // Serialize as JSON — the same pattern load() already uses for subtitles,
+    // avoiding a 6-arg JNI signature that has to change every time a field lands.
+    let json = serde_json::to_string(&s).map_err(|e| PlayerError { message: e.to_string() })?;
+    // Kotlin: fun setAudioSettings(json: String)
+    self.call_player_method_string("setAudioSettings", &json)?;
+    self.shared_state.lock_safe().audio_settings = s;
+    Ok(())
+}
+
+fn audio_settings(&self) -> AudioSettings {
+    self.shared_state.lock_safe().audio_settings.clone()
+}
+```
+
+`ExoPlayerState` gains an `audio_settings: AudioSettings` field. Note
+`ExoPlayerBackend` currently holds no such state — `position`/`state`/`volume` are
+all pushed in by JNI callbacks — so this is the first *pull*-side field. That is
+correct: audio settings are commanded downward, never reported upward.
+
+### Kotlin — `JellyTauPlayer.kt`
+
+```kotlin
+fun setAudioSettings(json: String) {
+    val s = JSONObject(json)
+    applyEqualizer(s.getBoolean("equalizerEnabled"), s.getJSONArray("equalizerBands"))
+    applyNormalization(s.getBoolean("normalizeVolume"), s.getString("volumeLevel"))
+    exoPlayer.pauseAtEndOfMediaItems = !s.getBoolean("gaplessPlayback")
+}
+```
+
+Three independent mechanisms:
+
+- **Gapless** — nearly free. ExoPlayer is gapless by default for compatible
+  formats; honouring the setting means *disabling* it when the user turns it off,
+  via `pauseAtEndOfMediaItems`. Note this only applies within a loaded playlist;
+  our queue loads one item at a time, so verify behaviour before claiming DR-035
+  on Android (see Testing).
+- **Equalizer** — `android.media.audiofx.Equalizer` bound to
+  `exoPlayer.audioSessionId`. Android's EQ exposes a device-dependent band count
+  (commonly 5) at fixed centre frequencies, which will **not** match our 10-band
+  ISO layout. Rust owns the canonical 10 bands; Kotlin resamples them onto the
+  device's bands by nearest-centre-frequency interpolation. Gains are in
+  millibels (`setBandLevel` takes mB, we store dB → ×100), clamped to the
+  device's reported `getBandLevelRange()`.
+- **Normalization** — `android.media.audiofx.LoudnessEnhancer`, also bound to the
+  audio session, `setTargetGain(mB)` derived from `VolumeLevel`. This is a gain
+  booster, not a true EBU R128 normalizer like MPV's `dynaudnorm`; parity is
+  approximate and should be documented as such rather than overclaimed.
+
+Lifecycle: build the effects lazily on first use, release them in `release()`,
+and re-attach on `onAudioSessionIdChanged` — ExoPlayer can rebuild its audio sink
+(e.g. on a format change), which invalidates effects bound to the old session.
+
+### Make the silent-failure mode impossible
+
+The trait's default is the root cause of this whole class of bug:
+
+```rust
+// backend.rs:85 — reports success while doing nothing
+fn set_audio_settings(&mut self, _settings: &AudioSettings) -> Result<(), PlayerError> {
+    Ok(())
+}
+```
+
+Android inherits this, so every EQ/normalization change on Android returns `Ok`
+and silently does nothing — the UI ships and has no effect, with no error anywhere.
+
+Once `ExoPlayerBackend` implements the methods, **change the trait default to
+`Err(PlayerError::not_implemented())`**, matching how `set_audio_track` /
+`set_subtitle_track` already behave. Any future backend that forgets to implement
+audio settings then fails loudly instead of lying.
+
+Check the call sites before flipping it: `NullBackend` overrides both methods, so
+the graceful-degradation path is unaffected, but confirm nothing treats a
+`set_audio_settings` error as fatal to playback.
+
+### Re-application on track load
+
+`PlayerController` already re-pushes `AudioSettings` per track on the platforms
+that implement it; the Android path inherits that for free once the trait methods
+exist. No controller change.
+
+### 🔴 Threading note
+
+`setAudioSettings` is invoked from Rust on whatever thread the command lands on.
+`AudioEffect` construction must not happen on the ExoPlayer application thread
+from inside a player callback — that is the re-entrancy hazard CLAUDE.md warns
+about, and the same shape as the `AutoplayDecision` deadlock. Post the work to
+the player's handler rather than doing it inline in a listener.
+
+## Out of scope
+
+- **Crossfade (UR-031 / DR-034).** Not implemented on *any* platform today, and
+  architecturally blocked on MPV (single-stream audio chain; `acrossfade` needs
+  two inputs). Implementing it on Android alone would invert the parity gap. It
+  needs its own spec and probably two player instances.
+- True EBU R128 normalization. `LoudnessEnhancer` is a gain stage; matching
+  `dynaudnorm` exactly is out of reach without a custom `AudioProcessor`.
+- Windows audio settings — see [windows-native-audio-backend.md](windows-native-audio-backend.md).
+
+## Acceptance criteria
+
+- [ ] `ExoPlayerBackend` overrides `set_audio_settings` and `audio_settings`.
+- [ ] EQ preset change on Android audibly changes playback; setting persists across track changes and app restart.
+- [ ] Normalization toggle audibly changes level; the three presets are ordered Loud > Normal > Quiet.
+- [ ] Disabling gapless produces a gap between consecutive tracks; enabling it does not.
+- [ ] Effects are released on `release()` and survive an audio-session rebuild.
+- [ ] `requirements.md` parity matrix updated: EQ and normalization ✅ Android.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] `bun run check:boundary` passes.
+- [ ] New requirement-implementing code carries `// TRACES:` comments.
+- [ ] `bindings.ts` regenerated if Rust types changed.
+
+## Testing
+
+**Rust** (`cargo test`): `set_audio_settings` stores the sanitized settings and
+`audio_settings()` returns them — assert clamping/normalisation is applied
+(crossfade clamped to 12s, band vector normalised to `EQ_BANDS.len()`). The JNI
+call itself is not unit-testable; extract the JSON serialization into a pure
+function and test that its shape matches what the Kotlin parser expects. That
+serialization contract is the part most likely to silently break.
+
+**Kotlin**: the band-resampling function (10 canonical bands → N device bands) is
+pure arithmetic — extract it and unit-test it, including the degenerate cases of
+a 5-band device and a device reporting 10 bands.
+
+**Manual, on device** (these are the ones that actually prove it):
+1. Set Bass Boost, play a track, confirm audible change.
+2. Toggle normalization mid-track; confirm level change without a playback stall.
+3. Queue two gapless-encoded tracks, toggle the setting, confirm the gap appears/disappears.
+4. Force a format change (44.1kHz → 48kHz track) and confirm the EQ still applies afterwards — this exercises the session-rebuild re-attach.
+
+## TRACES
+
+- `ExoPlayerBackend::set_audio_settings` → `// TRACES: UR-027, UR-032, UR-033 | DR-030, DR-035, DR-036`
+- Kotlin `setAudioSettings` / `applyEqualizer` / `applyNormalization` → same IDs
+- Band-resampling helper + its tests → `DR-030 | UT-xxx`
+
+## Notes for the implementer
+
+- Read [audio-equalizer.md](audio-equalizer.md) first — it defines the canonical
+  band layout and the preset→curve rule this spec consumes. Do not redefine bands
+  in Kotlin.
+- Android source edits go in `src-tauri/android/src` (canonical tree), then run
+  `scripts/sync-android-sources.sh`. Never edit the `gen/` tree.
+- There is a **stale duplicate** `JellyTauPlayer.kt` (285 lines) at
+  `src-tauri/android/app/src/main/java/com/dtourolle/jellytau/player/` alongside
+  the real 1103-line file at `src-tauri/android/src/main/java/...`. Edit the
+  latter. Consider deleting the former as a separate change.
+- A parallel Claude session may be active — `git diff` before "repairing"
+  unexpected changes.

docs/specs/android-native-video-spike.md

@@ -0,0 +1,306 @@
+# Spec: Android native video — transparent-webview spike
+
+**Status:** Spike succeeded (2026-08-11); shipped behind `experimentalNativeVideo`,
+default off. Flipping that default shipped **audio with no picture** and was
+reverted (DR-172). Three defects behind that have since been fixed — DR-182
+(nothing on the native path could lift the poster overlay), DR-183 (the JS
+bridges raced the page load), DR-184 (the SurfaceView was never detached).
+Branch `fix/android-native-video-visible`. **The default stays off until the
+device criteria below are green.**
+
+**The spike's central question is answered: yes.** A `SurfaceView` *can* be
+composited behind a transparent Tauri WebView on Android. Nothing upstream
+blocked it and nothing upstream demonstrated it — this is, as far as the issue
+trackers show, the first working instance. The remaining flag is about test
+coverage and the unverified cases below, not about viability.
+**Requirements:** IR-004, UR-003, UR-004, UR-041 → DR-001, DR-004, DR-150, DR-151, DR-152
+**Note:** the original draft cited DR-023/DR-024 here. Those are the *subtitle*
+and *audio-track selection UI* requirements — unrelated to this work. The IDs
+actually implemented are DR-150 (native rendering behind the flag), DR-151 (the
+severed SurfaceView attach chain) and DR-152 (capabilities reported by Rust).
+**UX spec:** n/a — no intended visual change; the video surface must land exactly where the `<video>` element is today
+**Supersedes / revises:** acts on finding 2 of [playback-backend-unification.md](playback-backend-unification.md)
+
+## Summary
+
+Test whether ExoPlayer's existing `SurfaceView` video path can be composited
+behind a transparent Tauri WebView on Android. If it works, Android regains
+hardware video decoding (MediaCodec) and libass-quality ASS/SSA subtitles, both
+of which the current webview path lacks. If it does not, we document why and
+delete the dead code.
+
+This is a **spike**, not a feature commitment. The deliverable is a yes/no answer
+with evidence, plus either a working path behind a flag or a removal.
+
+## Motivation
+
+`createAdapter()` hardcodes `const effectiveKind = "html5"` and does
+`void backendKind`, discarding the `use_html5_element` value Rust computes in
+`get_player_status`. As a result:
+
+- `NativePlayerAdapter` is dead code.
+- `JellyTauPlayer.kt`'s `getOrCreateSurfaceView()` — which already calls
+  `setZOrderMediaOverlay(false)` and wires `setVideoSurfaceHolder` — is
+  unreachable.
+- Android video decodes in the WebView instead of via MediaCodec, despite
+  `CodecDetector.kt` going to the trouble of reporting hardware codec
+  capabilities back to Rust for DeviceProfile generation.
+
+The code comment in `nativeAdapter.ts:11-14` justifies this by citing
+tauri#10152 as an upstream blocker. **That justification is stale.**
+
+### Why the blocker no longer holds
+
+- tauri#10152 is open but **dead since 2024-07-01**, and it is a *feature
+  request* ("Support transparent webviews on mobile"), not a bug report about
+  compositing.
+- The capability shipped in tauri commit `27d01834` (2024-09-02) — a clippy
+  cleanup that moved `transparent()` out of the desktop-gated impl block, fencing
+  only the tao call behind `#[cfg(desktop)]`. Because it landed as unrelated
+  cleanup, nobody closed the issue.
+- The black/white-screen reports (tauri#8381, tauri#9408) were a real but
+  *different* bug: a broken JNI signature for `setBackgroundColor`, fixed in
+  **wry 0.39.4** (PR #1237). We ship wry 0.55.x.
+- Current wry calls `setBackgroundColor(0)` unconditionally on Android when
+  transparency is requested.
+
+### The honest caveat
+
+**Nobody has demonstrated SurfaceView-behind-WebView on Tauri Android.** A search
+of both `tauri-apps/tauri` and `tauri-apps/wry` issues for `surfaceview` returns
+zero results, and the one native-video Tauri plugin
+(`YeonV/tauri-plugin-videoplayer`) sidesteps compositing by launching a separate
+fullscreen Activity. Nothing upstream blocks this; nothing upstream proves it.
+Hence: spike, not feature.
+
+Note this is the *Android* question only. The equivalent Linux compositing
+problem is maintainer-declared unfixable and is **not** in scope — see the
+unification spec.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Which video backend this platform uses | Rust (existing) | `get_player_status` already computes `use_html5_element`. The frontend must *consume* it, not decide it. Restoring that is the point of the spike. |
+| Surface creation, z-ordering, `setVideoSurfaceHolder` lifecycle | Kotlin | Android platform mechanics; already written in `JellyTauPlayer.kt`. |
+| Seek/audio-track *strategy* | Rust (existing) | Already returned by `player_seek_video` / `player_switch_audio_track`; `NativePlayerAdapter` executes the chosen primitive. Unchanged — this is exactly what the `PlayerAdapter` contract was built for. |
+| Positioning the surface under the video viewport | Frontend | Pure presentation/layout. **This is the risk area** — see Design. |
+
+## Design
+
+### Phase 1 — prove compositing (no app changes)
+
+Before touching the adapter factory, verify the primitive works at all:
+
+1. Set `"transparent": true` in `tauri.conf.json` for the Android build, plus
+   `html, body { background: transparent; }`.
+2. Confirm the WebView is genuinely transparent (a native view behind it is
+   visible) and that the app does not regress to a black/white screen.
+
+If this fails, stop — everything downstream is moot, and the finding is that
+Tauri Android transparency is still broken in practice despite the shipped fix.
+
+### Phase 2 — un-hardcode the factory
+
+```ts
+// src/lib/player/adapters/index.ts
+export function createAdapter({ backendKind, host, bridge }: CreateAdapterArgs): PlayerAdapter {
+  return backendKind === "native"
+    ? new NativePlayerAdapter(host)
+    : new Html5PlayerAdapter(host, bridge);
+}
+```
+
+`backendKind` comes from `get_player_status` (`VideoBackend::Native` on Android).
+Gate behind a setting — `experimentalNativeVideo`, default **off** — so a broken
+spike cannot ship as a regression. Rust already owns this decision; the flag only
+suppresses it.
+
+**Also in scope: remove the user-agent sniffing in
+`src/lib/services/webviewAudio.ts:30-41`.** It re-derives which audio backend the
+platform has from `navigator.userAgent` ("matching the Rust cfg gate", per its own
+comment) — the frontend deciding a backend fact it should be told. Same root cause
+as the hardcode above, same fix: consume the value Rust already computes. Fold it
+in here rather than leaving a second, subtler copy of the bug behind. If
+`get_player_status` does not currently expose enough to cover the audio case, add
+the field — that is backend work, and correct.
+
+### Implementation findings (2026-08-11)
+
+Two blockers existed that this spec did not anticipate. Both were in code the
+spec assumed was merely *unreachable*; it was also *broken*.
+
+**1. The Kotlin attach chain was severed.** `JellyTauPlayer.setActivity()` had
+**zero callers** anywhere in the tree. `currentActivity` was therefore always
+null, so `autoAttachSurface()` logged "Cannot attach surface - no Activity
+reference" and returned. The `SurfaceView` was created and wired to ExoPlayer but
+never added to the view hierarchy — video would have decoded to a surface that
+was never on screen, *regardless* of webview transparency. Fixed by calling
+`JellyTauPlayer.setActivity(this)` from `MainActivity.onCreate`.
+
+Note the knock-on: `PictureInPictureManager.canEnterPip()` gates on
+`VideoOverlayManager.isVideoSurfaceAttached()`, which was permanently false. PiP
+on the video path was dead for the same reason.
+
+**2. `createAdapter()` was not the real gate.** It is never called by production
+code — `VideoPlayer.svelte` constructs `Html5PlayerAdapter` directly. The actual
+override was `VideoPlayer.svelte`'s INTERIM block, which read Rust's
+`useHtml5Element`, forced it to `true`, and called `playerStop()` to kill the
+native backend `player_play_item` had just started. Both sites are now fixed;
+`VideoPlayer.svelte` routes through `createAdapter()` so there is one gate.
+
+**Transparency needs two independent layers cleared,** not one. The spec's
+Phase 1 named only `html, body`. Clearing just the page leaves the WebView
+widget's own background opaque, which is a black screen with audio — the exact
+symptom the INTERIM comment described as "native surface not visible". Both are
+now toggled together by `$lib/utils/videoSurface.ts`:
+
+| Layer | Cleared by | Reachable from |
+|-------|-----------|----------------|
+| WebView widget background + window drawable | `AndroidVideoSurface.setTransparent()` (MainActivity) | Kotlin only |
+| `html`/`body` + app-shell `--color-background` | `data-native-video` attribute → app.css | CSS only |
+
+Transparency is scoped to `tauri.android.conf.json` rather than the base config:
+a transparent window on Linux is a regression, since nothing renders behind it.
+It is also toggled per-session rather than set once — a permanently transparent
+window shows the launcher through the rest of the app.
+
+### Phase 3 — surface positioning
+
+The hard part, and where this most likely fails. The webview's `<video>` element
+occupies a laid-out box; the `SurfaceView` must be positioned to match it, and
+kept matched through scroll, rotation, and mini-player transitions.
+
+Approach: the video view reports its `getBoundingClientRect()` to Rust, which
+forwards the rect to Kotlin to position the `SurfaceView`. This is the same
+"faking it" technique the ecosystem uses on desktop — acceptable here *only if*
+the video is effectively fullscreen on Android, which it is in the player route.
+
+**Explicit failure criterion**: if the surface cannot be kept aligned during
+rotation or the mini-player transition without visible artefacts, the spike fails
+and we keep HTML5. Do not ship a janky native path for a codec win.
+
+**Update: no rect plumbing was needed.** The premise — that the surface must be
+positioned to match a laid-out `<video>` box — does not hold on the player route,
+where video is fullscreen. `VideoOverlayManager` adds the SurfaceView at index 0
+of `android.R.id.content` with `MATCH_PARENT`, and `fitSurfaceToScreen()`
+(`JellyTauPlayer.kt`) already letterboxes/pillarboxes to the real video aspect
+ratio and re-centres via a `Gravity.CENTER` `FrameLayout.LayoutParams`. Rotation
+is handled by an `OnLayoutChangeListener` that re-fits on any bounds change. The
+frontend's native branch is a bare `flex-1` box, so there is no rect to report
+and nothing to keep in sync.
+
+Fullscreen playback is confirmed working on device. But this reasoning rests
+entirely on the fullscreen assumption, so **the mini-player transition is the
+known gap** — it is the one case where the surface is *not* fullscreen, and
+therefore the one case where the "no rect plumbing needed" conclusion could
+still turn out to be wrong. If artefacts appear there, the fix is the rect
+reporting this section originally proposed, scoped to that transition alone.
+
+### A trap for the next implementer
+
+There is a **stale duplicate player** at
+`src-tauri/android/app/src/main/java/com/dtourolle/jellytau/player/JellyTauPlayer.kt`
+(only commit: `cfddc1e` "First working POC"). No `sourceSets` entry points at it,
+so it is not compiled — but edits made there silently do nothing. The canonical
+tree is `src-tauri/android/src`, synced into `gen/` by
+`scripts/sync-android-sources.sh`.
+
+### What we gain if it works
+
+- **Hardware decode via MediaCodec** — `CodecDetector.kt` already reports
+  capabilities; the DeviceProfile would finally match what actually plays.
+- **ASS/SSA subtitles** are *not* automatic. ExoPlayer cannot render them; that
+  would require libmpv, which is a separate and much larger decision (see the
+  unification spec's engine comparison). Scope this spike to hardware decode
+  only, and do not claim subtitle improvements from it.
+
+## Out of scope
+
+- Linux native video. Maintainer-declared unfixable on WebKitGTK/Wayland.
+- Replacing ExoPlayer with libmpv on Android.
+- Windows native video.
+- Removing the HTML5 path. It stays as the default and the fallback.
+
+## Acceptance criteria
+
+The spike is **complete** when one of these is true:
+
+**Success path**
+- [x] Transparent WebView confirmed working on a physical device (reported by the maintainer; the config that enables it is now committed in `tauri.android.conf.json`).
+- [x] `experimentalNativeVideo` off → behaviour byte-identical to today. Guarded by `adapterSelection.test.ts`, which asserts the flag-off case forces HTML5 even when Rust reports native.
+- [x] `webviewAudio.ts` no longer inspects `navigator.userAgent`; the platform's audio backend is read from Rust (`player_get_capabilities` → `usesWebviewAudio`).
+- [x] `experimentalNativeVideo` on → video plays via ExoPlayer, correctly positioned, on a physical device (2026-08-11). The surface reaches the hierarchy and is visible through the transparent WebView — the whole point of the spike.
+- [x] The poster/title card comes down on the native path. It never could: every `markMediaReady()` call site is a `<video>` element event and the native branch renders no element, so an opaque `bg-black` overlay covered the ExoPlayer surface for the whole session. See DR-182; guarded by `mediaReady.test.ts` (UT-184) and `VideoPlayer.nativeReveal.test.ts` (UT-185), the latter written failing first.
+- [x] The `AndroidVideoSurface` bridge is installed before the page that calls it loads, via `WryActivity.onWebViewCreate` instead of a 500 ms tree walk, and a missing bridge now logs an error instead of no-oping. See DR-183.
+- [x] The SurfaceView is detached when video stops, instead of accumulating one leaked view per native video. See DR-184.
+- [ ] Seek, audio-track switch and subtitle selection exercised through `NativePlayerAdapter`. Playback is confirmed; these individual controls are not yet each verified on the native path.
+- [ ] No artefacts on rotation, background/foreground, or **mini-player transition** — the last is the one case the fullscreen assumption does not cover, so it is the likeliest place to find a problem.
+- [ ] `adb shell dumpsys media.metrics` (or logcat) confirms a hardware decoder is in use. Plausible but unmeasured — do not claim the MediaCodec win until this is read.
+- [ ] Measured battery/thermal or CPU improvement over the HTML5 path on the same clip.
+
+**Failure path**
+- [ ] The blocking behaviour is documented in this spec with evidence.
+- [ ] `NativePlayerAdapter` and the unreachable `SurfaceView` code are deleted, or explicitly retained with a *correct* comment.
+- [ ] `nativeAdapter.ts:11-14` no longer cites tauri#10152.
+
+Either way:
+- [x] `bun run check` (0 errors), `bun run test` (997 passed), `bun run check:boundary` pass.
+- [x] `cargo fmt` / `cargo clippy` clean (no new warnings); `cargo test` passes (603 lib + 7 doc).
+
+### Why the 2026-08-11 verification and DR-172 do not contradict each other
+
+The spike was reported working on device; the same path then shipped as audio
+with no picture. Both are consistent with DR-182: the poster overlay is drawn
+only while `isMediaReady` is false, and the native path has no way to set it, so
+what the surface shows depends entirely on **whether that overlay is on screen**
+— not on whether compositing works. Any run that reached the player through a
+path leaving `isMediaReady` already true (a handoff return, a re-render, a
+session that had previously played on the HTML5 path) shows video; a cold start
+into the native path never does. That is also why DR-172 read the symptom as a
+compositing failure: on screen the two are identical, and the one piece of
+evidence separating them — `WebView transparent = true` never being logged —
+points at DR-183 rather than at the compositing itself.
+
+**This reasoning is not yet device-confirmed.** It explains the reports and is
+backed by the code, but the criteria above are what settle it.
+
+> Note: this environment has no host WebKitGTK dev packages, no Android SDK and
+> no `bun`, so all of the above were run inside the CI builder image
+> (`gitea.tourolle.paris/dtourolle/jellytau-builder:latest`). On Fedora the bind
+> mount needs `:z` for SELinux, and `scripts/build-android.sh` hardcodes
+> `ANDROID_HOME="$HOME/Android/Sdk"`, so the image's SDK at `/opt/android-sdk`
+> must be symlinked there rather than passed by env var.
+
+## Testing
+
+Adapter-selection logic is pure and testable without a device: assert
+`createAdapter` returns `NativePlayerAdapter` for `backendKind: "native"` with
+the flag on, and `Html5PlayerAdapter` in every other combination — including that
+the flag off forces HTML5 even when Rust says native. That last case is the
+regression guard.
+
+Everything else is manual on-device; there is no meaningful way to unit-test
+surface compositing. Test on at least two devices — compositing behaviour varies
+by OEM and Android version.
+
+Per CLAUDE.md, if the spike turns into a bug fix (e.g. seek breaks under the
+native adapter), write the failing test first.
+
+## TRACES
+
+- `createAdapter` → `// TRACES: UR-003, UR-004 | DR-004, DR-150 | UT-149`
+- Adapter-selection tests → `UT-xxx`
+- No new requirement IDs; this spike either satisfies existing IR-004 expectations or documents why it cannot.
+
+## Notes for the implementer
+
+- **Do not skip Phase 1.** If transparency does not work, phases 2 and 3 are
+  wasted effort.
+- `VideoPlayer.svelte` has a documented hazard: no lifecycle calls after an
+  `await` in `onMount` — it flips to HTML5 mode and breaks Android seek. The
+  adapter swap touches exactly this code path.
+- tauri-specta tagged responses keep Rust field names (`new_url`, not `newUrl`).
+- Android source edits go in `src-tauri/android/src`, then run
+  `scripts/sync-android-sources.sh`.
+- A parallel Claude session may be active — `git diff` first.

docs/specs/audio-equalizer.md

@@ -0,0 +1,176 @@
+# Spec: Audio equalizer
+
+**Status:** Accepted
+**Requirements:** UR-027 → DR-030 (EQ UI), IR-020 (MPV EQ integration).
+**UX spec:** n/a (extends the Settings › Audio section, ux-flows §8.1 instant-apply).
+**Supersedes / revises:** —
+**Revised by:** [android-audio-settings-parity.md](android-audio-settings-parity.md) — lifts the "Android is a no-op" limitation below.
+
+## Summary
+
+Add a graphic audio equalizer to playback. Users pick a preset (Flat, Rock,
+Pop, Jazz, Classical, Bass Boost, Treble Boost, Vocal) or set custom per-band
+gains, from a new block in Settings › Audio. On Linux the gains apply live via
+MPV's audio-filter chain; the settings persist and re-apply on the next track
+and at startup, exactly like crossfade/gapless/normalize do today. Android is a
+no-op for now (documented parity gap, same as those three features).
+
+## Motivation
+
+UR-027 is one of the few still-unbuilt audio features. The audio-settings
+pipeline it needs already exists — `AudioSettings` + `set_audio_settings` on the
+`PlayerBackend` trait, the `player_set_audio_settings` command, and the Settings
+› Audio UI with instant-apply. Crossfade, gapless, and volume normalization all
+ride that pipeline. The equalizer is the same shape: N more fields on
+`AudioSettings`, an `af` filter on the MPV backend, one more block in the
+settings panel. No new command, no new state machine.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| EQ band count, centre frequencies, gain range/clamping | Rust | Domain of the audio engine; the bands must match what the MPV filter expects. Changing the DSP must not require a frontend change. |
+| Preset name → per-band gain curve | Rust | A preset *is* a domain gain curve, not a label. It changes with the audio engine's band layout, never with the UI. Placing it in the frontend would be the scoped-search taxonomy mistake again (values that look like config but are domain data). |
+| Translating gains → MPV `af` filter string | Rust | Platform playback detail; lives with the other `set_audio_settings` filter code in `mpv_backend.rs`. |
+| Persisting the chosen settings, re-pushing on load | Rust/existing | Same path crossfade/etc. already use; the controller re-applies `AudioSettings` per track. |
+| Rendering band sliders, the preset chips, live readouts | Frontend | Pure presentation; changes only if the settings UI is redesigned. |
+| Which preset chip is highlighted; instant-apply on change | Frontend | Presentation/input handling (UR-057), the same as the normalize preset picker. |
+
+Tie-breaker note: the preset→curve map is the one tempting boundary leak. It goes
+in Rust because a preset is a set of band gains defined *by the band layout*,
+which is an engine property. The frontend only ever names a preset and renders
+the resulting gains; it never defines them.
+
+## Design
+
+### `AudioSettings` (Rust, `settings.rs`)
+
+Add two fields (both `#[serde(rename_all = "camelCase")]` via the existing
+struct attribute):
+
+```rust
+/// Equalizer enabled. When false, no `af` EQ filter is applied.
+pub equalizer_enabled: bool,
+/// Per-band gains in dB, one per FIXED band (see EQ_BANDS). Length is
+/// validated/normalised to EQ_BANDS.len(); clamped to [-12, +12] dB.
+pub equalizer_bands: Vec<f32>,
+```
+
+Fixed 10-band ISO layout (domain constant in `settings.rs`):
+
+```rust
+pub const EQ_BANDS: [f32; 10] =
+    [31.0, 62.0, 125.0, 250.0, 500.0, 1000.0, 2000.0, 4000.0, 8000.0, 16000.0];
+pub const EQ_GAIN_MIN: f32 = -12.0;
+pub const EQ_GAIN_MAX: f32 = 12.0;
+```
+
+- `Default`: `equalizer_enabled: false`, `equalizer_bands: vec![0.0; 10]` (flat).
+- New `with_equalizer_normalised(self)` clamps each gain to `[EQ_GAIN_MIN,
+  EQ_GAIN_MAX]` and pads/truncates the vec to 10 bands. Applied in the command
+  alongside `with_crossfade_clamped` (add that call too — it's currently missing).
+- Backward compat: both fields `#[serde(default)]` so old persisted JSON loads.
+
+### Presets (Rust, `settings.rs`)
+
+```rust
+#[derive(specta::Type, Serialize, Deserialize, Clone, Copy, PartialEq)]
+#[serde(rename_all = "camelCase")]
+pub enum EqPreset { Flat, Rock, Pop, Jazz, Classical, BassBoost, TrebleBoost, Vocal }
+
+impl EqPreset {
+    /// The 10-band gain curve (dB) for this preset.
+    pub fn gains(&self) -> [f32; 10] { /* table */ }
+}
+```
+
+Preset selection is a *frontend* convenience: tapping a chip sets
+`equalizer_bands = preset.gains()` and pushes settings. The curve tables live in
+Rust; the frontend reads them via a tiny `player_get_eq_presets` command
+returning `Vec<(EqPreset, Vec<f32>)>` (or a map), so the frontend never encodes
+the numbers. (If exposing the whole table is awkward through specta, expose
+`player_eq_preset_gains(preset) -> Vec<f32>` instead — pick at implement time.)
+
+### MPV application (Rust, `mpv_backend.rs::set_audio_settings`)
+
+Build an `equalizer` / `anequalizer` filter from the bands and set the `af`
+property. When `equalizer_enabled` is false or all gains are 0, clear the EQ
+filter (leave any other `af` entries intact). Use `af add`/`af remove` or a
+rebuilt `af` string; keep it isolated so it doesn't stomp a future crossfade
+filter. Errors map to `PlayerError` like the gapless code.
+
+### No new persistence table
+
+`AudioSettings` is already round-tripped by the frontend settings store and
+re-pushed via `player_set_audio_settings` on change and on load. The two new
+fields ride along. `NullBackend`/Android inherit the trait default (no-op).
+
+### Wire summary
+
+- Command names unchanged: `player_set_audio_settings`,
+  `player_get_audio_settings` (now carry the EQ fields).
+- New (optional) read-only command for preset curves — kebab n/a (it's a
+  command): `player_get_eq_presets` (or `player_eq_preset_gains`).
+- Regenerate `bindings.ts` from the Rust types; never hand-edit.
+
+## Out of scope
+
+- Android/ExoPlayer EQ (parity gap tracked with crossfade/gapless/normalize).
+  **Now specified in [android-audio-settings-parity.md](android-audio-settings-parity.md)**,
+  which implements `set_audio_settings` on `ExoPlayerBackend`. The canonical band
+  layout and preset→curve map defined here remain authoritative; the Android side
+  resamples those bands onto the device equalizer rather than defining its own.
+- Per-track or per-library EQ profiles — one global profile only.
+- Automatic loudness/room correction; only manual bands + presets.
+- Changing the crossfade/normalize TODOs in `set_audio_settings` beyond wiring
+  the missing `with_crossfade_clamped` call.
+
+## Acceptance criteria
+
+- [ ] Settings › Audio has an Equalizer block: enable toggle, preset chips, 10
+      band sliders with live dB readouts, instant-apply (no Save button).
+- [ ] Choosing a preset sets the bands from the Rust-defined curve; editing a
+      band switches the highlighted preset to "Custom" (frontend-only label).
+- [ ] Gains clamp to [-12, +12] dB; the band vector always normalises to 10.
+- [ ] On Linux, enabling EQ audibly changes output and persists across tracks
+      and app restart; disabling clears the filter without affecting other audio.
+- [ ] Old persisted settings (no EQ fields) load without error, defaulting flat.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] `bun run check:boundary` passes (no preset curve numbers in the frontend).
+- [ ] New requirement-implementing code carries `// TRACES:` comments.
+- [ ] `bindings.ts` regenerated.
+
+## Testing
+
+- Rust (`settings.rs`): default is flat + disabled; `with_equalizer_normalised`
+  clamps out-of-range gains and pads/truncates band length; serialization
+  round-trips the camelCase fields; backward-compat load of pre-EQ JSON; each
+  preset returns a 10-length curve; Flat is all zeros.
+- Rust IPC param naming for any new command (camelCase rule per CLAUDE.md).
+- Frontend (`settings` page or an extracted helper): selecting a preset sets the
+  expected band array; editing a band flips the label to Custom; enable toggle
+  gates the sliders. Keep DSP untested on the frontend (it's Rust's).
+
+## TRACES
+
+- `AudioSettings` EQ fields + normalise + presets: `UR-027 | DR-030` (+ unit tests)
+- MPV EQ filter application: `UR-027 | IR-020`
+- Settings EQ UI block: `UR-027 | DR-030`
+- Preset-curve command: `UR-027 | DR-030`
+
+## Notes for the implementer
+
+- A parallel Claude session is active in this repo (it has touched
+  `tauri.conf.json`, `Dockerfile`, `package.json`, home components, and added
+  build scripts, and the Rust build is currently broken by its
+  `tauri.conf.json` bundle-target change). `git diff` before "repairing"
+  anything you didn't write; keep EQ changes isolated to `settings.rs`,
+  `mpv_backend.rs`, `backend.rs` (trait default already covers it),
+  `commands/player/settings.rs`, and the settings page.
+- Mirror the volume-normalization block in the settings page for the toggle +
+  preset-picker pattern; mirror the gapless code in `set_audio_settings` for the
+  MPV property handling.
+- Confirm the exact MPV filter name available in the linked libmpv
+  (`equalizer` vs `anequalizer`/`superequalizer`) before committing the filter
+  string; gate cleanly if unavailable.

docs/specs/boundary-tripwire-hardening.md

@@ -0,0 +1,215 @@
+# Spec: Harden the frontend boundary tripwire
+
+**Status:** Implemented
+**Requirements:** DR-094
+**UX spec:** n/a — developer tooling.
+**Supersedes / revises:** revises the detection rule in
+[scripts/check-frontend-boundary.sh](../../scripts/check-frontend-boundary.sh);
+the boundary *policy* in [scoped-search-boundary.md](scoped-search-boundary.md)
+is unchanged.
+
+## Summary
+
+`bun run check:boundary` passes on a tree that contains the exact leak it was
+built to catch. It matches a multi-type array only when written **inline at the
+query site**, so assigning the same array to a named const evades it entirely —
+which is how [searchScope.ts](../../src/lib/utils/searchScope.ts) has kept a
+category→item-type mapping through every green CI run. This spec broadens the
+match to item-type array literals anywhere in `src/`, and resolves the handful
+of legitimate hits that broadening surfaces.
+
+## Motivation
+
+The current pattern is anchored to the `includeItemTypes:` key:
+
+```sh
+PATTERN='includeItemTypes:[[:space:]]*\[[^]]*,[^]]*\]'
+```
+
+The live leak is not written that way:
+
+```ts
+// src/lib/utils/searchScope.ts:29 — invisible to the tripwire
+const SCOPE_ITEM_TYPES = { music: ["MusicAlbum", "MusicArtist", "Audio", "Playlist"], … };
+```
+
+The taxonomy and the query are one indirection apart, and the grep only sees the
+query. The script's own header is admirably honest that it is "a TRIPWIRE, NOT A
+PROOF" — but the gap here is not a subtle judgment call it was designed to
+defer to human review. It is the *crudest form* of the violation, one `const`
+away from the shape it does match, in the very file the founding incident was
+written about.
+
+Broadening the pattern to any item-type array literal finds it, with a
+manageable number of other hits (measured, not estimated):
+
+| Site | Verdict |
+|------|---------|
+| `searchScope.ts:30,32` | 🔴 The leak. Removed by [scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md). |
+| `PersonDetailView.svelte:30` | Already allowlisted, with a recorded reason. |
+| `DownloadedBrowse.svelte:95` | Borderline — `["MusicAlbum","Series","Season","BoxSet"].includes(item.type)` as an "is this a container?" predicate. |
+| `GenericMediaListPage.svelte:298` | Borderline — `["MusicAlbum","MusicArtist","Audio","Playlist"].includes(config.itemType)` as a music-styling predicate. |
+| 6 hits in `*.test.ts` | Excluded; tests legitimately name types. |
+
+Four non-test sites total. This is a tractable change, not a boil-the-ocean one.
+
+## Layer assignment
+
+Tooling only — no application logic, nothing crosses IPC. The two borderline
+*application* sites do get a layer decision, below.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Detecting item-type array literals in `src/` | Build tooling (`scripts/`) | Static analysis of repo source; belongs beside the existing check. |
+| "Is this item a container?" (`DownloadedBrowse`) | **Rust** (recommended) | Containers-vs-leaves is Jellyfin structure, and the set grows when Jellyfin adds a container type — the litmus test's "yes". Prefer a `MediaItem.isContainer` boolean from the backend over a type-set predicate in a component. |
+| "Is this music content?" (`GenericMediaListPage`) | **Frontend, allowlisted** | Selects a grid *style*. It reads `config.itemType`, a value the page already declares about itself, and changes only if the UI is redesigned — the litmus test's "no". Single-type presentation is explicitly not the target of the rule. |
+
+`DownloadedBrowse` defaults to Rust per the checklist's borderline rule; see
+Out of scope for why the migration itself is deferred rather than bundled.
+
+## Design
+
+### 1. Broaden the pattern
+
+Replace the key-anchored pattern with one matching an array literal of two or
+more known Jellyfin item types, wherever it appears:
+
+```sh
+# Two or more adjacent item-type string literals inside a bracket.
+TYPES='Movie|Series|Episode|Audio|MusicAlbum|MusicArtist|MusicVideo|Season|BoxSet|Playlist|Book|AudioBook|Video|Person|Folder|CollectionFolder|TvChannel|LiveTvChannel'
+PATTERN="\[[[:space:]]*\"($TYPES)\"[[:space:]]*,[[:space:]]*\"($TYPES)\""
+```
+
+Properties worth stating, because each is a deliberate trade:
+
+- **Not anchored to any key**, so a named const, a function return, a `Record`
+  value, or an inline query all match equally.
+- **Requires two adjacent type literals**, preserving the existing and correct
+  carve-out that single-type presentation (`itemType: "Movie"`) is legitimate.
+- **Requires string literals**, so `item.type === "Audio"` (display inspection)
+  still does not match.
+- **Explicit type list**, not `[A-Z][a-z]+`, so arbitrary string arrays
+  (`["High","Low"]`, `["Songs","Albums"]`) do not produce noise.
+
+Keep `grep -rInE`, the `*.test.*` exclusion, and the allowlist mechanism as they
+are — all three work.
+
+### 2. Resolve the surfaced sites
+
+- `PersonDetailView.svelte` — already allowlisted; entry unchanged.
+- `GenericMediaListPage.svelte` — **add to the allowlist** with the reason from
+  the layer table (grid styling over a self-declared `itemType`).
+- `DownloadedBrowse.svelte` — **add to the allowlist with a `TODO` naming the
+  preferred fix** (backend `isContainer`). An allowlist entry that records a
+  known-borderline decision is honest; silently broadening the pattern to miss
+  it would not be.
+- `searchScope.ts` — **not allowlisted.** It is the leak, and
+  [scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md)
+  deletes it.
+
+### 3. 🔴 Sequencing
+
+**This spec must land after Stage 1 of
+[scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md).**
+Hardening the tripwire first turns `master` red on a violation with no fix
+available, and the only ways out are reverting the hardening or allowlisting the
+leak — the second of which is exactly how a boundary rule dies.
+
+### 4. Keep the allowlist honest
+
+The script already warns that a growing allowlist means the boundary is eroding.
+This change takes it from 1 entry to 3, which is close to that line. Add a hard
+cap so drift is caught mechanically rather than by whoever notices:
+
+```sh
+MAX_ALLOWLIST=4
+if [ "${#ALLOWLIST[@]}" -gt "$MAX_ALLOWLIST" ]; then
+  echo "❌ Allowlist has ${#ALLOWLIST[@]} entries (max $MAX_ALLOWLIST)."
+  echo "   Push taxonomy into Rust instead of appending here."
+  exit 1
+fi
+```
+
+The cap is deliberately just above the current count: the next exception forces
+a conversation instead of a one-line append.
+
+### 5. Restate the limits
+
+The header's "tripwire, not a proof" caveat stays and gets sharper. The broadened
+pattern still cannot see:
+
+- a type set built at run time (`[...musicTypes, "Playlist"]`),
+- types split across variables (`const A = "Audio"; [A, B]`),
+- taxonomy expressed as a `switch` or chained `||` rather than an array.
+
+The spec-review checklist remains the real gate. This raises the floor; it does
+not close the class.
+
+## Out of scope
+
+- **Migrating `DownloadedBrowse` to a backend `isContainer` flag.** It touches
+  `MediaItem`, `bindings.ts`, and the offline path — its own spec. Allowlisted
+  with a TODO here so it is recorded, not forgotten.
+- The scoped-search fix itself — [scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md).
+- Detecting the run-time-construction cases listed above.
+- Extending the check to Rust or Kotlin (the rule is about `src/`).
+- Changing the boundary *policy* in CLAUDE.md.
+
+## Acceptance criteria
+
+- [ ] With `searchScope.ts` reverted to its leaking form, `bun run check:boundary`
+      **fails** and names `src/lib/utils/searchScope.ts`. This is the criterion
+      that proves the fix — verify it explicitly before landing.
+- [ ] On the post-fix tree, `bun run check:boundary` passes.
+- [ ] A newly introduced `const X = ["Movie", "Series"]` in any non-test `src/`
+      file fails the check (regression test for the const-indirection evasion).
+- [ ] `itemType: "Movie"` and `item.type === "Audio"` do **not** trip the check.
+- [ ] `["High", "Low"]` and other non-item-type arrays do **not** trip it.
+- [ ] Test files are still excluded (the 6 known test hits stay silent).
+- [ ] The allowlist has exactly 3 entries, each with a written reason; a 5th
+      entry fails the check via `MAX_ALLOWLIST`.
+- [ ] The script header still states it is a tripwire, not a proof, and names the
+      evasions it cannot see.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `bun run test:all` passes.
+
+## Testing
+
+The script is bash and has no test harness. Verify by construction — each is a
+temporary edit, run, revert:
+
+1. Reintroduce the `SCOPE_ITEM_TYPES` const → **must fail**.
+2. Add `const T = ["Movie","Series"]` to a scratch `.svelte` file → **must fail**.
+3. Add the same to a `.test.ts` file → **must pass** (exclusion holds).
+4. Add `itemType: "Movie"` → **must pass**.
+5. Add a 5th allowlist entry → **must fail** on the cap.
+
+Record the five results in the PR description. A grep-based gate that has never
+been observed failing is indistinguishable from one that cannot fail — which is
+the precise condition this whole spec exists to correct.
+
+## TRACES
+
+Allocate in `requirements.md`:
+
+- **DR-094** — "Frontend boundary tripwire detects Jellyfin item-type array
+  literals anywhere in `src/` (not only inline at an `includeItemTypes:` query
+  site), so a category→type mapping cannot evade the check via a named const;
+  allowlist is capped to force taxonomy into Rust rather than accumulating
+  exceptions." Category: Tooling. Status: Done on merge.
+
+Shell scripts carry no `TRACES:` comment convention in this repo; reference
+DR-094 in the script header comment instead.
+
+## Notes for the implementer
+
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes (CLAUDE.md §Gotchas).
+- **Land after Stage 1 of the scoped-search fix** — see §3. This is the one
+  ordering constraint that will break `master` if ignored.
+- Test the regex against the current tree *before* committing:
+  `grep -rInE "$PATTERN" src/ | grep -v '\.test\.'` should return exactly the
+  four sites in the Motivation table.
+- The `TYPES` list will need occasional extension as Jellyfin adds types.
+  That is acceptable for a tripwire — an unlisted type produces a false
+  negative, never a false positive, so the check degrades safely.

docs/specs/build-provenance.md

@@ -0,0 +1,250 @@
+# Spec: Build provenance (git describe + build profile)
+
+**Status:** Proposed
+**Requirements:** new DR-093 (build provenance surfaced in-app and in logs); no UR — this is a diagnostic capability, not a user feature
+**UX spec:** n/a — adds an About block to Settings; no new flow
+**Supersedes / revises:** —
+
+## Summary
+
+Make every build say exactly what it is. Today a running JellyTau reports no
+version at all — not in the UI, not in the logs — and the only version string in
+the tree is the hand-maintained `0.2.0` duplicated across three files.
+
+This adds a `build.rs`-generated provenance string (`git describe` + short SHA +
+dirty flag + debug/release profile), exposes it over IPC, and renders it in a new
+Settings › About block. It also removes one of the three hand-bumped version
+files.
+
+## Motivation
+
+The concrete problem: when a user reports "the equalizer does nothing on my
+device" — which is a live risk for v0.2.0, whose Android audio settings are not
+yet device-verified — there is currently no way to tell which build they are
+running. Tag? Master? A local debug build from three weeks ago? The bug report
+cannot distinguish them.
+
+Two smaller irritations this also fixes:
+
+- **Debug builds masquerade as releases.** `0.2.0` is `0.2.0` whether it came
+  from a tagged release or `bun run tauri dev`.
+- **Three files carry the version.** `package.json`, `src-tauri/Cargo.toml` and
+  `src-tauri/tauri.conf.json` must be bumped in lockstep; the release checklist
+  exists partly to stop them drifting.
+
+### What this deliberately does *not* do
+
+**The canonical version stays hand-bumped in `Cargo.toml`.** Cargo requires a
+literal semver string at manifest-parse time and cannot derive it from git. The
+same is true of `tauri.conf.json`. Attempting to source the *release* version
+from a tag trades a scripted, reviewable bump for a fragile build-time
+dependency that breaks in exactly the environment we care most about (CI, in
+Docker, from a shallow clone).
+
+So: **the release version is authored; the build provenance is derived.** They
+answer different questions — "what release is this?" versus "what commit is this
+binary actually built from?" — and only the second benefits from git.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Capturing git describe / SHA / dirty state at compile time | Rust (`build.rs`) | Only the Rust build has a compile step that can shell out to git and bake the result into the binary. A frontend equivalent would report the *dev server's* state, not the shipped binary's. |
+| Degrading to a sentinel when git is unavailable | Rust (`build.rs`) | Build-environment concern. Must never fail the build — CI runs in Docker from a shallow clone. |
+| Release version (`0.2.0`) | Rust (`Cargo.toml`, authored) | Domain fact about the product, not derivable from the environment. |
+| Deciding *what a build is* (release / dev / dirty) | Rust | Domain classification. The frontend must not infer "this is a dev build" from a string shape — it renders what it is told. |
+| Rendering the About block, copy-to-clipboard | Frontend | Pure presentation. |
+
+Borderline row: the release/dev/dirty classification could be done in the
+frontend by pattern-matching the describe string. It goes to Rust because that is
+a *rule about what constitutes a release build*, and it would have to change if
+the tagging scheme changed — the litmus test in the template puts that in Rust.
+Send a typed enum, not a string for the frontend to parse.
+
+## Design
+
+### `build.rs`
+
+```rust
+fn main() {
+    emit_build_provenance();
+    tauri_build::build()
+}
+
+fn emit_build_provenance() {
+    let describe = std::process::Command::new("git")
+        .args(["describe", "--tags", "--always", "--dirty"])
+        .output()
+        .ok()
+        .filter(|o| o.status.success())
+        .and_then(|o| String::from_utf8(o.stdout).ok())
+        .map(|s| s.trim().to_string())
+        .unwrap_or_else(|| "unknown".to_string());
+
+    println!("cargo:rustc-env=JELLYTAU_GIT_DESCRIBE={describe}");
+
+    // Rebuild when HEAD moves or a ref is written, so the string does not go
+    // stale across commits. Guarded: these paths do not exist in a git-less
+    // source tarball, and emitting rerun-if-changed for a missing path would
+    // force a rebuild every time.
+    for p in [".git/HEAD", ".git/refs"] {
+        if std::path::Path::new("../").join(p).exists() {
+            println!("cargo:rerun-if-changed=../{p}");
+        }
+    }
+}
+```
+
+🔴 **`build.rs` must never fail the build.** Every git call is
+`.ok()`-swallowed; a missing git binary, a shallow clone, or a source tarball all
+yield `"unknown"`. A build that breaks because git is absent would be a worse bug
+than the one this fixes.
+
+Note the `../` prefixes: `build.rs` runs with CWD at `src-tauri/`, so the repo's
+`.git` is one level up.
+
+### The provenance type
+
+```rust
+/// TRACES: DR-093
+#[derive(specta::Type, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct BuildInfo {
+    /// Authored release version (Cargo.toml).
+    pub version: String,
+    /// `git describe --tags --always --dirty`, or "unknown".
+    pub git_describe: String,
+    /// What kind of build this is — classified in Rust, not inferred by the UI.
+    pub kind: BuildKind,
+}
+
+/// TRACES: DR-093
+#[derive(specta::Type, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub enum BuildKind {
+    /// Built from a clean, exactly-tagged commit in release mode.
+    Release,
+    /// Release-mode build that is not on a clean tag (e.g. master, or dirty).
+    Untagged,
+    /// debug_assertions build.
+    Development,
+    /// Git state unavailable at build time.
+    Unknown,
+}
+```
+
+Classification:
+
+```rust
+let kind = if cfg!(debug_assertions) {
+    BuildKind::Development
+} else if describe == "unknown" {
+    BuildKind::Unknown
+} else if describe.contains('-') {   // "v0.2.0-3-gcb79a37" or "...-dirty"
+    BuildKind::Untagged
+} else {
+    BuildKind::Release
+};
+```
+
+### Command
+
+```rust
+/// TRACES: DR-093
+#[tauri::command]
+#[specta::specta]
+pub fn get_build_info() -> BuildInfo { … }
+```
+
+No parameters, so the camelCase param rule does not apply; the struct fields do
+need `#[serde(rename_all = "camelCase")]` (above). Regenerate `bindings.ts`.
+
+Also log the provenance once at startup, next to the existing init logging —
+that is what makes a user-submitted log file self-identifying, which is most of
+the value.
+
+### Settings › About
+
+A new block at the bottom of `src/routes/settings/+page.svelte`, rendering
+version, describe string, and a badge for non-release builds. One
+copy-to-clipboard button that yields a paste-ready block for bug reports:
+
+```
+JellyTau 0.2.0 (v0.2.0-3-gcb79a37-dirty, development)
+linux x86_64
+```
+
+Platform/arch come from the existing Tauri APIs; do not shell out.
+
+### Removing one version file
+
+`tauri.conf.json`'s `"version"` field can be omitted, in which case Tauri falls
+back to the Cargo version. That takes the bump from three files to two.
+
+**Verify before adopting**: confirm the Android `versionName`/`versionCode` and
+the NSIS installer version still resolve correctly with the field absent —
+Android packaging in particular reads the Tauri config. If either regresses,
+keep the field and drop this part; it is a convenience, not the point of the
+spec.
+
+## Out of scope
+
+- Deriving the *release* version from git tags (see Motivation).
+- A build-time timestamp. It defeats reproducible builds and adds little over
+  the commit SHA.
+- CI provenance/attestation, SBOM, signing.
+- Displaying the Jellyfin server version (separate concern, already available
+  from `/System/Info`).
+
+## Acceptance criteria
+
+- [ ] `cargo build` succeeds with git absent, from a shallow clone, and from a source tarball with no `.git` — yielding `"unknown"` in each case, never a build failure.
+- [ ] A tagged clean release build reports `BuildKind::Release`; `bun run tauri dev` reports `Development`; a dirty tree reports `Untagged` (release mode) with `-dirty` in the describe string.
+- [ ] The describe string changes after a new commit without a manual `cargo clean` (rerun-if-changed works).
+- [ ] Provenance is logged once at startup.
+- [ ] Settings › About renders version + describe + build-kind badge, with working copy-to-clipboard.
+- [ ] 🔴 CI checkouts that build a shippable artifact set `fetch-depth: 0`, or their artifacts are knowingly stamped `unknown`. Currently only `publish-docs.yml` sets it; `build-release.yml` has five checkouts and `build-and-test.yml` two, all of which would report `unknown` as-is.
+- [ ] **No toolchain installed in CI** — git is already present in the builder image; nothing new is added.
+- [ ] `bun run check`, `bun run test`, `bun run check:boundary` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] `bindings.ts` regenerated.
+- [ ] DR-093 allocated in `requirements.md`; new code carries `// TRACES:`.
+
+## Testing
+
+**Rust**: the classification is pure and must be extracted from the command as
+`classify_build(describe: &str, debug: bool) -> BuildKind` so it can be tested
+directly. Cover: `"v0.2.0"` → `Release`; `"v0.2.0-3-gcb79a37"` → `Untagged`;
+`"v0.2.0-dirty"` → `Untagged`; `"unknown"` → `Unknown`; `debug = true` → always
+`Development` regardless of describe.
+
+`build.rs` itself is not unit-testable. Verify its failure path manually by
+building with `PATH` stripped of git, and from a `git archive` tarball — both
+must succeed with `"unknown"`.
+
+**Frontend**: assert the About block renders each `BuildKind` correctly, and that
+it renders the backend-supplied kind rather than re-deriving it from the string
+(a test that passes a `Release` kind with a `-dirty` describe and asserts the
+badge follows the *kind* would catch that regression).
+
+## TRACES
+
+- `build.rs` provenance emission → `// TRACES: | DR-093`
+- `BuildInfo` / `BuildKind` / `classify_build` → `// TRACES: | DR-093`
+- `get_build_info` command → `// TRACES: | DR-093`
+- Settings About block → `// TRACES: | DR-093`
+- `classify_build` tests → `UT-BUILD-1`
+- Allocate **DR-093** in `requirements.md` ("Build provenance: git describe and
+  build profile surfaced in-app and in logs"). Next free DR at time of writing
+  is DR-093.
+
+## Notes for the implementer
+
+- Do the `build.rs` + command + logging first; the About UI is the smaller half
+  and the logging alone delivers most of the diagnostic value.
+- The `fetch-depth: 0` change is the easiest part to forget and the one that
+  makes CI artifacts useless if missed — it is why that acceptance box is
+  flagged. Weigh it per workflow: test-only jobs do not need it.
+- Do not add a build timestamp "while you are in there" — see Out of scope.
+- A parallel Claude session may be active — `git diff` before "repairing"
+  unexpected changes.

docs/specs/catalog-index-search.md

@@ -0,0 +1,334 @@
+# Spec: Locally-indexed search
+
+**Status:** Implemented
+**Requirements:** UR-065 → DR-108, DR-109, DR-110, DR-111; IR-030
+**UX spec:** [ux-flows.md §6.1](../ux-flows.md) (search surface is unchanged)
+**Revises:** [scoped-search.md](scoped-search.md) and
+[scoped-search-boundary.md](scoped-search-boundary.md) — scope semantics are
+untouched; this changes only *which corpus* the cache leg searches.
+
+## Summary
+
+Search stops depending on a per-keystroke round trip to Jellyfin. The local
+SQLite catalog — which is already synced and already FTS5-indexed — becomes the
+corpus the instant leg of search reads, so results appear as fast as SQLite can
+answer, online or offline. A background indexer keeps that catalog fresh on a
+schedule instead of only at app start, prunes content deleted on the server, and
+covers the item types search groups results by. The server query stays, demoted
+to a background reconciliation that merges in late results for anything indexed
+since the last pass.
+
+## Motivation
+
+The pieces are already built and simply not wired together:
+
+- [`sync_full_catalog`](../../src-tauri/src/commands/catalog.rs) already walks
+  every library `Recursive=true` and persists items with `synced_at`.
+- `items_fts` (schema.rs migration 001) already indexes `name`, `overview`,
+  `album_name`, `album_artist`, `artists`, `series_name` with keep-in-sync
+  triggers.
+- `repository_search` is already two-phase — synchronous cache result, then a
+  spawned server query merged in via the `search-event`.
+
+What breaks the chain is that the cache leg is hard-restricted to *downloaded*
+items. `OfflineRepository::search` wraps its FTS query in a `downloaded_items`
+CTE requiring `d.status = 'completed'`:
+
+```sql
+FROM items i
+ JOIN items_fts fts ON fts.rowid = i.rowid
+ INNER JOIN downloaded_items di ON i.id = di.id
+ WHERE i.server_id = ? AND items_fts MATCH ?
+```
+
+So for a user with no downloads, phase 1 returns nothing on every query, and
+every debounced keystroke falls through to a full `Recursive=true` server
+request with `Limit=10000`. The populated local index is never read.
+
+`get_items` does not have this problem — it gates a third `synced_at IS NOT NULL`
+branch on `include_catalog_browse()` (offline.rs, the "Show all server media"
+toggle). The asymmetry is the bug: **offline you can already browse the whole
+catalog but cannot search it.**
+
+Three further defects found while confirming the above:
+
+1. **The FTS index grows without bound.** `save_to_cache` uses
+   `INSERT OR REPLACE INTO items`, but `recursive_triggers` is never enabled
+   (`storage/mod.rs` sets only `foreign_keys` and `journal_mode`). SQLite fires
+   `AFTER DELETE` triggers on a REPLACE *only* with recursive triggers on — so
+   `items_ad` never runs, the old FTS row is orphaned, and because `items.id` is
+   a `TEXT PRIMARY KEY` the replacement row takes a **new rowid** and inserts a
+   second FTS entry. Every sync appends a duplicate index. Results stay correct
+   (the `INNER JOIN … ON fts.rowid = i.rowid` hides orphans, and no rowid is ever
+   reused because nothing is deleted) but `MATCH` degrades permanently.
+2. **Server-side deletions never propagate.** There is no `DELETE FROM items`
+   anywhere in the codebase. The local catalog is append-only, so media removed
+   from the server would stay searchable forever — tolerable when the cache was
+   only a browse accelerator, not acceptable when it is the search corpus.
+3. **The index omits types search groups by.** `CATALOG_ITEM_TYPES` is
+   `MusicAlbum, Movie, Series, Season, Episode, Audio, BoxSet` — no
+   `MusicArtist`, no `Playlist`, and People live in a separate `people` table
+   with no FTS at all. UR-060 mandates Artists and People result groups, so today
+   those can *only* come from the server.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Which corpus search reads (downloads-only vs full synced catalog) | **Rust** | Sync/availability policy over domain data. Changes if Jellyfin's API or the offline rules change, not if the UI is redesigned. Reuses the existing `include_catalog_browse()` flag so search and browse cannot diverge again. |
+| Index freshness policy — TTL, when a re-index is due, skip-while-offline | **Rust** | Explicitly named as domain policy in [SPEC-REVIEW-CHECKLIST.md](SPEC-REVIEW-CHECKLIST.md) ("reachability/sync policy"). It is currently frontend-driven in `offlineCatalog.ts`; this spec moves it. |
+| Which Jellyfin item types get indexed (`CATALOG_ITEM_TYPES`) | **Rust** | Textbook domain taxonomy — a category→item-type set. Must never appear in `src/`. |
+| Reconciling a crawl against local rows (what to prune) | **Rust** | Operates on domain data and depends on crawl completeness semantics. |
+| FTS query construction, ranking, scope→type expansion | **Rust** | Already there (`search_rank.rs`, `SearchScope::item_types()`); unchanged by this spec. |
+| Rendering a "catalog last indexed N ago" hint and any re-index button | **Frontend** | Pure presentation of a backend-supplied timestamp. |
+| Debounce interval, result group order, scope chips | **Frontend** | Input handling and view preference; changes only if the UI is redesigned. |
+
+Borderline call, recorded: the **TTL value itself** (how many hours before a
+re-index is due) could be argued as a user preference and therefore frontend. It
+is placed in Rust because the frontend must not be able to decide *whether the
+cache is authoritative* — that is the same class of decision as
+`include_catalog_browse`, which already lives in Rust. If the TTL later becomes
+user-configurable it stays a Rust-owned setting the frontend edits through a
+command, not a frontend constant. Borderline defaults to Rust.
+
+## Design
+
+### 1. Search the full synced catalog (DR-108)
+
+`OfflineRepository::search` mirrors `get_items` exactly: rename the CTE to
+`available_items` and add the same third branch, gated on the same flag.
+
+```rust
+let catalog_branch = if include_catalog_browse() {
+    "UNION
+
+        -- Synced catalog: fast online search, or the offline 'Show all
+        -- server media' view. Mirrors get_items; see set_include_catalog_browse.
+        SELECT DISTINCT i.id
+        FROM items i
+        WHERE i.synced_at IS NOT NULL"
+} else {
+    ""
+};
+```
+
+No new IPC surface and no frontend change: `set_include_catalog_browse` is
+already called with `true` when online or when the offline toggle is on, and
+`false` only when offline with the toggle off. Search inherits the correct
+behaviour in all three states, and the "search is restricted to downloads" case
+survives for users who deliberately asked for downloads-only.
+
+Also fix, in the same function, the `type_filter` built by **string
+interpolation** of `include_item_types` rather than bound parameters. It is
+currently safe only because callers pass `SearchScope`-derived values, but
+`SearchOptions.include_item_types` is settable directly from the frontend (as
+`GenericMediaListPage` does). Bind the values.
+
+Phase 2 (the server query) is unchanged and still merges via `search-event`, so
+content added to the server since the last index still surfaces — just late
+rather than first.
+
+### 2. Scheduled background indexer (DR-109, IR-030)
+
+A Rust-owned task replaces the frontend's startup-only trigger.
+
+```rust
+/// How long a full-catalog index stays fresh before a re-index is due.
+const CATALOG_INDEX_TTL: Duration = Duration::from_secs(6 * 60 * 60);
+```
+
+Behaviour:
+
+- On app setup, spawn a tokio task that ticks every 30 min.
+- Each tick: if a repository is active **and** the server is reachable **and**
+  `now - last_catalog_sync > CATALOG_INDEX_TTL`, run a full index pass.
+- On the existing `ConnectivityMonitor` reconnect signal, evaluate the same
+  staleness condition immediately rather than waiting for the next tick.
+- Never run two passes concurrently (the existing `syncInProgress` guard moves
+  into Rust as an `AtomicBool`).
+
+`last_catalog_sync` is already written to `app_settings` by `sync_full_catalog`
+and is currently read only for a UI hint; this makes it load-bearing.
+
+`RepositoryManager` (`commands/repository.rs`) is a `HashMap<String, …>` with no
+notion of an active handle, so the task has nothing to run against. Add:
+
+```rust
+pub struct RepositoryManager {
+    repositories: Arc<Mutex<HashMap<String, Arc<HybridRepository>>>>,
+    active: Arc<Mutex<Option<String>>>,   // set in create(), cleared in destroy()
+}
+```
+
+Progress is reported with a **kebab-case** event (per the project convention):
+
+```rust
+// event name: "catalog-index-event"
+#[derive(specta::Type, Serialize, Clone)]
+#[serde(rename_all = "camelCase")]
+pub struct CatalogIndexEvent {
+    pub state: CatalogIndexState,   // #[serde(tag = "type")] Idle | Running | Complete | Failed
+    pub libraries_done: usize,
+    pub libraries_total: usize,
+    pub items_indexed: usize,
+}
+```
+
+`sync_full_catalog` stays a command so the UI can still force a pass; it and the
+scheduler share one internal `run_index_pass()`.
+
+### 3. Index hygiene — no orphans, and deletions propagate (DR-110)
+
+**Orphan growth.** Replace `INSERT OR REPLACE INTO items (…)` in `save_to_cache`
+with a true upsert:
+
+```sql
+INSERT INTO items (id, server_id, …) VALUES (…)
+ON CONFLICT(id) DO UPDATE SET
+    name = excluded.name, overview = excluded.overview, …,
+    synced_at = excluded.synced_at
+```
+
+This preserves the rowid (which `items_fts` keys on via `content_rowid`) and
+fires `items_au` instead of silently orphaning a row. Preferred over
+`PRAGMA recursive_triggers = ON` because it also stops the rowid churn, and the
+three FTS triggers are the only triggers in the schema so nothing else depends
+on REPLACE semantics.
+
+A new migration `021_rebuild_items_fts` clears the orphans already accumulated on
+existing installs:
+
+```sql
+INSERT INTO items_fts(items_fts) VALUES('rebuild');
+```
+
+**Deletions.** After a library crawls *successfully and completely*, reconcile:
+delete local rows for that library whose `id` was not seen in the crawl. Two
+constraints the implementation must respect:
+
+- Skip any item with a completed download — the user has the file; removing the
+  row would orphan it. Prune only synced-but-not-downloaded rows.
+- Only sweep libraries whose crawl succeeded. `sync_full_catalog` is
+  deliberately best-effort per library, and `items.parent_id` is
+  `ON DELETE CASCADE` — sweeping on a partial crawl would cascade a whole series
+  away because one request timed out.
+
+### 4. Index the types search groups by (DR-111)
+
+Add `MusicArtist` and `Playlist` to `CATALOG_ITEM_TYPES`.
+
+People need a different mechanism: they live in `people` (`id`, `server_id`,
+`name`, `overview`, `primary_image_tag`, `synced_at`), populated incidentally by
+item-detail fetches, with no FTS table. Migration `022_people_fts` adds one
+mirroring the `items_fts` pattern:
+
+```sql
+CREATE VIRTUAL TABLE IF NOT EXISTS people_fts USING fts5(
+    name, overview, content='people', content_rowid='rowid'
+);
+-- plus people_ai / people_ad / people_au triggers
+```
+
+`OfflineRepository::search` UNIONs `people_fts` matches into its result set as
+`Person`-typed items when the resolved scope permits them (i.e. when
+`include_item_types` is `None` — `SearchScope::All`). `search_rank.rs` already
+handles `MediaKind::Person`, so ranking needs no change.
+
+## Out of scope
+
+- **Incremental indexing** (e.g. Jellyfin's `MinDateLastSaved`). A full crawl is
+  what makes the deletion sweep in §3 sound — it yields the authoritative id set
+  per library. An incremental pass cannot detect deletions, so it would need a
+  separate reconciliation strategy. Worth revisiting if full crawls prove too
+  slow on large libraries; measure first.
+- **Changing search UX** — scope chips, group order, the debounce, and the
+  `/search` route are untouched.
+- **Removing the server leg.** Phase 2 stays.
+- The two dead search implementations (`storage_search_items` in
+  `commands/storage/mod.rs`, `offline_search` in `commands/offline.rs`) — both
+  registered in `lib.rs` and exported to `bindings.ts`, neither called from the
+  frontend. Deleting them is correct but is cleanup, not this feature; file
+  separately so this spec's diff stays reviewable.
+- `GenericMediaListPage` passing raw `includeItemTypes` and re-implementing the
+  store's request-id/event protocol. A real boundary smell, tracked separately.
+
+## Acceptance criteria
+
+- [ ] With a synced catalog and **zero downloads**, typing a query returns
+      results from the local index before any server request completes.
+- [ ] Offline with "Show all server media" **on**, search returns the full
+      catalog (non-downloaded entries greyed out, matching browse).
+- [ ] Offline with the toggle **off**, search returns downloaded media only —
+      the behaviour that exists today.
+- [ ] Re-running a full index pass N times does not grow `items_fts` row count
+      beyond the `items` row count.
+- [ ] An item deleted server-side disappears from local search after one index
+      pass; a **downloaded** item deleted server-side does not.
+- [ ] A library that fails mid-crawl prunes nothing.
+- [ ] Searching an artist or actor name returns results with the server
+      unreachable.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] `bun run check:boundary` passes.
+- [ ] New requirement-implementing code carries `// TRACES:` comments.
+- [ ] `bindings.ts` regenerated (new `CatalogIndexEvent` type).
+
+## Testing
+
+Per CLAUDE.md, each defect gets a **failing test first**.
+
+Rust (`cargo test`), against an in-memory DB seeded with synced-but-not-
+downloaded items:
+
+- `search` returns synced items when `include_catalog_browse()` is true, and
+  only downloaded items when false. *Fails today* — the current CTE returns
+  empty in the first case.
+- Upserting the same item twice leaves exactly one `items_fts` row. *Fails
+  today.*
+- The sweep removes a vanished synced item, retains a vanished downloaded item,
+  and no-ops for a library whose crawl errored.
+- `type_filter` binds parameters — a type string containing a quote does not
+  alter the query.
+- Staleness: a `last_catalog_sync` inside the TTL does not trigger a pass; one
+  outside it does; offline never does.
+- `people_fts` matches surface as `Person` items under `SearchScope::All` and
+  are excluded under `Music`/`Movies`/`Tv`.
+
+Frontend (`vitest`): the catalog-index event maps to the staleness hint; no
+change to the search store's request-id/stale-response handling, which stays
+covered by its existing tests.
+
+## TRACES
+
+| Piece | Tag |
+|---|---|
+| `OfflineRepository::search` availability CTE | `// TRACES: UR-065 \| DR-108` |
+| Background indexer task + scheduling | `// TRACES: UR-065 \| DR-109, IR-030` |
+| `save_to_cache` upsert + FTS rebuild migration | `// TRACES: UR-065 \| DR-110` |
+| Deletion reconciliation | `// TRACES: UR-065 \| DR-110` |
+| `CATALOG_ITEM_TYPES` widening + `people_fts` | `// TRACES: UR-065, UR-060 \| DR-111` |
+
+## Notes for the implementer
+
+- **A parallel Claude session may be active in this repo.** Run `git diff`
+  before "repairing" changes you did not make (CLAUDE.md gotchas).
+- The frontend's `offlineCatalog.ts` startup trigger should be **removed**, not
+  left alongside the Rust scheduler — two independent triggers with one
+  `syncInProgress` guard each is how double-crawls happen.
+- `downloads` has a relaxed FK to `items` (migration 005). Verify the deletion
+  sweep's interaction with it before enabling the sweep, and check whether
+  `parent_id`'s `ON DELETE CASCADE` reaches further than intended.
+- The existing 100 ms `cache_with_timeout` in `hybrid.rs` returns *empty* on
+  timeout rather than erroring. Once the cache leg is the primary path, that
+  budget may need raising — an FTS query over a large catalog on cold page cache
+  can exceed it, and the failure mode is a silently empty result.
+- Keep `SearchScope` semantics as-is: `All => None` (no filter), deliberately
+  not a union, so People and folders are not filtered out (DR-063).
+- Noted but deliberately not fixed here: `pushCatalogVisibility` in
+  `offlineCatalog.ts` derives the flag as `connected || showCatalog` — the
+  frontend computing an availability *policy*, even though the flag itself is
+  Rust-stored. DR-108 depends on that derivation being correct and it is, so
+  this spec leaves it alone. Once DR-109 has moved sync policy into Rust, the
+  derivation belongs there too, with the frontend pushing only the raw user
+  toggle. Folding it into this change would enlarge the diff for no behavioural
+  gain — but do not add *new* policy on the frontend side of that line.

docs/specs/favorites-browsing.md

@@ -0,0 +1,403 @@
+# Spec: Favourites — marking, browsing, and sync
+
+**Status:** Implemented
+**Requirements:** UR-067, UR-068, UR-069 → DR-113 … DR-120; JA-033, JA-034
+(allocated in [requirements.md](../requirements.md); tests UT-099 … UT-107).
+Note: UR-066/DR-112/IR-031 were claimed by the concurrent safe-area work while
+this spec was being written, so the ids here start one higher than first drafted.
+Existing: UR-017 → DR-021, JA-017, JA-018 (the toggle itself, already built).
+**UX spec:** [ux-flows.md](../ux-flows.md) §3.2 (full-player favourite), §5.2
+(album detail favourite), §5B.3 (movie detail hero: *Play / Download / Favorite*)
+— all three already specify favourite affordances that **do not exist in the
+build**. This spec closes those, and adds a new §5C for the Favourites browse
+surface.
+**Supersedes / revises:** nothing.
+
+## Summary
+
+JellyTau can favourite an item but can never show you what you favourited. The
+heart is mounted in exactly one place (the mini player), no query anywhere asks
+Jellyfin or the local database for favourites, and favourites marked on any other
+client are invisible here. This spec adds the read side (a Favourites page, home
+carousels, an in-library filter), puts the heart on detail pages and media cards,
+teaches the backend to ingest server-side favourite state, and drains favourite
+toggles made while offline.
+
+## Background: what exists today
+
+Verified in code, 2026-08-04. The **write** path is real and mostly correct; the
+**read** path does not exist at all.
+
+1. **Toggling works, from one place only.**
+   [FavoriteButton.svelte](../../src/lib/components/FavoriteButton.svelte) is
+   mounted solely in
+   [MiniPlayer.svelte:381](../../src/lib/components/player/MiniPlayer.svelte#L381).
+   Nothing else in `src/` renders it — so the only favouritable item in the app
+   is the one currently playing.
+
+2. **The toggle's plumbing is sound.**
+   [favorites.ts](../../src/lib/services/favorites.ts) writes local first
+   (`storage_toggle_favorite`,
+   [storage/mod.rs:908](../../src-tauri/src/commands/storage/mod.rs#L908) — sets
+   `user_data.is_favorite` + `pending_sync = 1`), then POST/DELETEs
+   `/Users/{uid}/FavoriteItems/{id}`
+   ([online.rs:1652](../../src-tauri/src/repository/online.rs#L1652)) only when
+   connected. Leave this design intact.
+
+3. **`MediaItem.user_data` is always `None` from the server.**
+   `JellyfinItem` has no `UserData` field, and `to_media_item` hardcodes
+   [`user_data: None`](../../src-tauri/src/repository/online.rs#L656) with the
+   comment *"User data not included in basic item responses"*. The only
+   populated `user_data` in the app comes from
+   [series_progress.rs](../../src-tauri/src/repository/series_progress.rs#L244)
+   and the local read in
+   [offline.rs:115](../../src-tauri/src/repository/offline.rs#L115). **Nothing
+   ingests server favourite state**, which is why the mini player has to fetch
+   `storageGetPlaybackProgress` per track to colour one heart
+   ([MiniPlayer.svelte:75-93](../../src/lib/components/player/MiniPlayer.svelte#L75-L93)).
+
+4. **No favourites query exists.**
+   [`GetItemsOptions`](../../src-tauri/src/repository/types.rs#L278) has no
+   favourites field; `Filters=IsFavorite` appears nowhere; no SQL selects
+   `is_favorite = 1`; there is no `/library/favorites` route and no favourites
+   carousel in [home.ts](../../src/lib/stores/home.ts).
+
+5. **Offline favourites are silently lossy.** Offline `mark_favorite` /
+   `unmark_favorite` are no-ops
+   ([offline.rs:1620](../../src-tauri/src/repository/offline.rs#L1620)), so an
+   offline toggle survives only as a local row with `pending_sync = 1` — and
+   **nothing ever drains that flag**. `syncService.queueFavorite`
+   ([syncService.ts:91](../../src/lib/services/syncService.ts#L91)) exists with
+   no callers.
+
+## Motivation
+
+Favouriting is a promise: the app takes the input and shows a "Added to
+favorites" toast, then discards it as far as the user can tell. Three of the UX
+flows already specify favourite buttons that were never built, and the one that
+was built (mini player) writes to a store nothing reads. Either the feature gets
+its read side or the heart should be removed — this spec takes the first option.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Favourites **scope** → set of Jellyfin item types | Rust | Domain taxonomy. Changes when Jellyfin adds/renames a type, never when the UI is redesigned. Reuses the canonical `SearchScope::item_types()` ([types.rs:324](../../src-tauri/src/repository/types.rs#L324)) — the exact leak class of [scoped-search-boundary.md](scoped-search-boundary.md). |
+| Cross-library favourites query (`Filters=IsFavorite`, `Recursive`, paging, sort field) | Rust | Query shaping against the Jellyfin API is domain logic; the endpoint's contract changes with the server, not the UI. |
+| Offline favourites SQL (join `user_data`, downloaded/catalog gating) | Rust | Storage + domain. Must obey the existing catalog-browse gate (DR-080) which the frontend cannot see. |
+| Deserialising Jellyfin `UserData` into `MediaItem.user_data` | Rust | Provider payload mapping. |
+| Mirroring server favourite state into the local `user_data` table | Rust | Cache/sync policy. |
+| Conflict rule: a local row with `pending_sync = 1` beats the server value | Rust | Business rule about which write wins; nothing to do with rendering. |
+| Draining pending favourite toggles on reconnect | Rust | Sync policy, and it must run whether or not any view is mounted — a frontend-driven drain dies with the component. Consistent with *reachability from real traffic* (DR-055). |
+| Which surfaces show favourites, tab order, row placement on home | Frontend | Pure presentation; changes only if the UI is redesigned. |
+| Heart placement, animation, toast, haptics, empty-state copy | Frontend | Presentation. |
+| In-session optimistic heart state shared across views | Frontend | View state, not persisted truth; the durable write already goes to Rust. |
+
+**Borderline, and the tie-breaker used:** *which* scopes appear as tabs (All /
+Movies / Shows / Music) is a presentation choice — the frontend picks which
+`SearchScope` values to offer. What each scope *means* is Rust's. The frontend
+sends the enum value and never names an item type in connection with favourites.
+Single-type pages (`itemType: "Movie"` on the Movies list page) stay as they are;
+this rule targets category taxonomy, not every mention of a type.
+
+## Design
+
+### 1. Server user data reaches `MediaItem` (Rust, DR-113, JA-034)
+
+Jellyfin returns `UserData` on `/Users/{uid}/Items*` responses. Add the field to
+`JellyfinItem` and map it in `to_media_item`, replacing the hardcoded `None`:
+
+```rust
+// in JellyfinItem
+#[serde(alias = "UserData")]
+pub user_data: Option<JellyfinUserData>,
+```
+
+`JellyfinUserData` deserialises `IsFavorite`, `Played`, `PlaybackPositionTicks`,
+`PlayCount`, `LastPlayedDate` into the existing
+[`UserData`](../../src-tauri/src/repository/types.rs#L44) type (which already
+carries `is_favorite` and already serialises camelCase, so `bindings.ts` needs no
+new type — only regeneration). Add `UserData` to the `Fields=` list in `get_items`
+/ `get_item` so the shape is explicit rather than relying on the default.
+
+Wire shape, unchanged from today's `UserData`:
+
+```ts
+item.userData?.isFavorite // boolean | null | undefined
+```
+
+Delete the now-false `// User data not included in basic item responses` comment.
+
+### 2. Local mirror of server favourites (Rust, DR-114)
+
+Choke point: `save_to_cache(parent_id, &items)` in
+[offline.rs](../../src-tauri/src/repository/offline.rs) — every server result
+that gets cached (including via
+[`cache_items_from_server`](../../src-tauri/src/repository/hybrid.rs#L124) and
+the background cache refresh) passes through it.
+
+For each item carrying `user_data.is_favorite`, upsert:
+
+```sql
+INSERT INTO user_data (user_id, item_id, is_favorite, synced_at, pending_sync)
+VALUES (?, ?, ?, ?, 0)
+ON CONFLICT(user_id, item_id) DO UPDATE SET
+  is_favorite = excluded.is_favorite,
+  synced_at   = excluded.synced_at
+WHERE user_data.pending_sync = 0;      -- local unsynced change wins
+```
+
+The `WHERE` on the conflict clause is the whole conflict rule: a toggle made
+offline is never overwritten by a stale server value before it has been pushed.
+
+### 3. Favourites queries (Rust, DR-115, DR-116, JA-033)
+
+**(a) In-library filter** — one new field on `GetItemsOptions`:
+
+```rust
+#[serde(skip_serializing_if = "Option::is_none")]
+pub favorites_only: Option<bool>,
+```
+
+- online `get_items`: append `&Filters=IsFavorite` when true.
+- offline `get_items`: add `INNER JOIN user_data ud ON ud.item_id = i.id AND ud.user_id = ? AND ud.is_favorite = 1`, composed with the existing `available_items` CTE so the downloads-only gate still applies.
+
+Frontend sends `{ favoritesOnly: true }` (camelCase — nested struct field, needs
+the existing `#[serde(rename_all = "camelCase")]` on `GetItemsOptions`, already
+present).
+
+**(b) Cross-library favourites** — a new trait method, because favourites span
+libraries and `get_items` is `ParentId`-shaped:
+
+```rust
+/// TRACES: UR-067 | DR-115 | JA-033
+async fn get_favorites(
+    &self,
+    scope: SearchScope,
+    options: Option<GetItemsOptions>,
+) -> Result<SearchResult, RepoError>;
+```
+
+```rust
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_get_favorites(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    scope: SearchScope,
+    options: Option<GetItemsOptions>,
+) -> Result<SearchResult, String>
+```
+
+Frontend call (command name matches the Rust fn exactly; top-level params
+auto-camelCase; `SearchScope` is `#[serde(rename_all = "camelCase")]` so the wire
+values are `"all" | "music" | "movies" | "tv"`):
+
+```ts
+await commands.repositoryGetFavorites(handle, "movies", { limit: 100 });
+```
+
+- **online**: `/Users/{uid}/Items?Filters=IsFavorite&Recursive=true&SortBy=SortName&SortOrder=Ascending` + `&IncludeItemTypes=…` from `scope.item_types()` (omit entirely on `None`, per that function's contract) + the standard `Fields=`.
+- **offline**: `items ⨝ user_data (is_favorite = 1)`, type filter from the same `scope.item_types()`, honouring `include_catalog_browse()`.
+- **hybrid**: same cache-first race as `get_items`, **including the DR-080 rule** — with the catalog-browse gate off, an empty offline result is authoritative and must not fall through to the server. Getting this wrong reproduces Defect B from [offline-downloaded-only-filter.md](offline-downloaded-only-filter.md).
+
+**The cache-first result arrives stale, and there is no second payload.** On a
+cache hit, `hybrid::get_items` returns the local rows and refreshes the cache in
+a background task whose result the frontend never sees — fine for a library
+listing that changes daily, wrong for favourites, where the *point* is that
+another client just changed something. Favourites is the second read path (after
+search) that needs the deferred update, so the background refresh in
+`get_favorites` must emit the same `favorites-changed` event as §4 when the
+server's favourite set differs from what was returned:
+
+```
+favorites-changed  →  { itemIds: string[] }   // union of ids whose is_favorite flipped
+```
+
+Both producers (background refresh, reconnect drain) emit the identical payload,
+and the frontend has one handler that refreshes the `favorites` store. Without
+this, a favourite marked on another client appears in JellyTau only on the
+*second* visit to the page.
+
+### 4. Draining offline toggles (Rust, DR-120)
+
+On the offline→online transition already detected by `ConnectivityMonitor`,
+select `user_data WHERE pending_sync = 1 AND is_favorite IS NOT NULL`, POST or
+DELETE `/Users/{uid}/FavoriteItems/{id}` per row, then set `pending_sync = 0` and
+`synced_at`. Failures leave the row pending for the next transition.
+
+Emit a kebab-case event when anything changed, so open views refresh without
+polling:
+
+```
+favorites-changed  →  { itemIds: string[] }
+```
+
+`syncService.queueFavorite` is dead code once this lands — delete it or point it
+at the backend drain; do not leave two competing queues.
+
+### 5. Frontend surfaces (DR-117, DR-118, DR-119)
+
+**Favourites page** — new route `/library/favorites`:
+- Scope tabs *All / Movies / Shows / Music* via the existing `LibraryViewTabs`; each tab sends a `SearchScope` value, nothing more.
+- Renders through `LibraryGrid` + `MediaCard` (tracklist for Music→tracks if the tab is later split; not in this pass).
+- Entry points: a card on the library overview ([library/+page.svelte](../../src/routes/library/+page.svelte)) and "See all" on the home rows.
+- Empty state per tab: "Nothing favourited yet — tap the heart on anything you like."
+
+**Home carousels** — `favoriteMovies`, `favoriteShows`, `favoriteMusic` added to
+[home.ts](../../src/lib/stores/home.ts), each `repositoryGetFavorites(scope, { limit: 20 })`,
+rendered after *Recently Added* and **only when non-empty** (no empty rows on a
+fresh install).
+
+**In-library filter** — a favourites toggle in the header of
+[GenericMediaListPage](../../src/lib/components/library/GenericMediaListPage.svelte)
+and the Movies/TV landing pages, passing `favoritesOnly: true` into the existing
+`repo.getItems(...)` options. Session-scoped state; not persisted (a persisted
+filter that hides most of a library is a support call waiting to happen).
+
+**Hearts** — mount `FavoriteButton`:
+- Movie / series detail hero button row, beside the download buttons ([library/[id]/+page.svelte:528-553](../../src/routes/library/%5Bid%5D/+page.svelte#L528-L553)) — closes ux-flows §5B.3.
+- `EpisodeFocusView`, `ArtistDetailView`, `PlaylistDetailView`, album detail — closes ux-flows §5.2.
+- `MediaCard` artwork overlay (top-right). Suppressed on `isServerOnly` cards, and must not fight the existing long-press/scroll-guard handlers ([MediaCard.svelte:60-70](../../src/lib/components/library/MediaCard.svelte#L60-L70)) — the heart is its own button and stops propagation.
+
+**Shared optimistic state** — a small `favorites` store (`Map<string, boolean>`
+overlay + `favorites.set(id, value)`), so un-hearting an item on the Favourites
+page removes it from the grid and from any home row without a refetch, and a
+heart tapped on a card is reflected on the detail page. Resolution order:
+
+```
+favorites store override  ??  item.userData?.isFavorite  ??  false
+```
+
+`toggleFavorite()` updates the store alongside its existing local + server
+writes; the `favorites-changed` event refreshes it. This removes the mini
+player's per-track `storageGetPlaybackProgress` fetch once items carry
+`userData`.
+
+### 6. Offline behaviour
+
+Toggling offline keeps working exactly as now (local write + `pending_sync`), and
+now actually reaches the server on reconnect (§4). The Favourites page offline
+shows favourites among downloaded/cached items, subject to the existing
+catalog-browse gate. The offline repo's no-op `mark_favorite`/`unmark_favorite`
+stay no-ops — the local write plus the drain is the offline path.
+
+## Out of scope
+
+- Favouriting people, genres, or collections; favourite **playlists** are included only insofar as they fall under the Music scope.
+- Sorting by "date favourited" — Jellyfin does not expose it. Favourites sort by name.
+- A dedicated bottom-nav tab for favourites (reachable from library overview + home).
+- Building a playlist or download batch from favourites.
+- Reconciling favourites for items that no longer exist on the server.
+- Splitting the Music tab into albums/artists/tracks sub-tabs.
+
+## Acceptance criteria
+
+- [ ] Favouriting is possible from movie, series, episode, album, artist and playlist detail pages, and from media cards in any grid.
+- [ ] A favourite marked in another Jellyfin client shows a filled heart in JellyTau without toggling it here.
+- [ ] `/library/favorites` lists favourites across libraries, filtered by the All/Movies/Shows/Music tabs.
+- [ ] Home shows favourite rows for movies, shows and music, and shows no row when a category has none.
+- [ ] Movies/TV/Music list pages can be filtered to favourites only.
+- [ ] Un-hearting an item on one surface updates the others without a manual refresh.
+- [ ] A favourite toggled while offline reaches the server after reconnect (verified against a real server or a fake repository).
+- [ ] Offline, the Favourites page respects the "Show all server media" gate — with it off, an empty result stays empty and does not fall through to the server.
+- [ ] No item-type set appears in `src/` in connection with favourites; the frontend sends `SearchScope` only.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] `bun run check:boundary` passes (necessary, not sufficient — see CLAUDE.md).
+- [ ] New requirement-implementing code carries `// TRACES:` comments.
+- [ ] `bindings.ts` regenerated from Rust, not hand-edited.
+
+## Testing
+
+**🔴 §4 (the pending-sync drain) is a bug fix — failing test first.** Write a
+test that toggles a favourite with the repository offline, transitions to online,
+and asserts the server call happened; watch it fail before writing the drain.
+
+Rust (`cd src-tauri && cargo test`):
+
+| Test | Covers |
+|------|--------|
+| UT-099 | A Jellyfin item JSON fixture with `UserData.IsFavorite: true` maps to `MediaItem.user_data.is_favorite == Some(true)` |
+| UT-100 | `online::get_favorites` builds an endpoint with `Filters=IsFavorite`, `Recursive=true`, and the scope's `IncludeItemTypes`; `SearchScope::All` omits the type filter entirely |
+| UT-101 | `offline::get_favorites` returns only `is_favorite = 1` rows, respects the scope type filter, and returns nothing extra when the catalog-browse gate is off |
+| UT-102 | `save_to_cache` mirror does **not** overwrite a row with `pending_sync = 1` |
+| UT-103 | Drain pushes pending rows, clears `pending_sync`, sets `synced_at`, and leaves failed rows pending |
+| UT-104 | `get_items` with `favorites_only: true` filters both online (endpoint) and offline (SQL) |
+| UT-107 | The background refresh in `hybrid::get_favorites` emits `favorites-changed` with the flipped ids, and emits nothing when the server set matches the cache |
+
+Frontend (`bun run test`):
+
+| Test | Covers |
+|------|--------|
+| UT-105 | `favorites` store override precedence: store value beats `userData.isFavorite` beats `false` |
+| UT-106 | Un-hearting removes the item from a favourites list view (pure logic extracted to a `.ts` module, per the TrackList/episodeStrip pattern) |
+| IT-0xx | `repositoryGetFavorites` param naming — add to [tauriIntegration.test.ts](../../src/lib/utils/tauriIntegration.test.ts): camelCase top-level params, scope serialised as `"movies"` etc. |
+
+Any component logic worth testing gets extracted into a plain `.ts` module first
+(`favoritesView.ts`), rather than tested through the component.
+
+## TRACES
+
+| Piece | Tag |
+|-------|-----|
+| `JellyfinUserData` + `to_media_item` mapping | `// TRACES: UR-069 \| DR-113, JA-034 \| UT-099` |
+| `save_to_cache` user_data mirror | `// TRACES: UR-069 \| DR-114 \| UT-102` |
+| `get_favorites` (trait, online, offline, hybrid) + command | `// TRACES: UR-067 \| DR-115, JA-033 \| UT-100, UT-101` |
+| `GetItemsOptions.favorites_only` handling | `// TRACES: UR-067 \| DR-116 \| UT-104` |
+| `/library/favorites` route + tabs | `// TRACES: UR-067 \| DR-117` |
+| Home favourite carousels | `// TRACES: UR-067 \| DR-118` |
+| `FavoriteButton` mounts + `favorites` store | `// TRACES: UR-068 \| DR-119 \| UT-105, UT-106` |
+| Pending-favourite drain + `favorites-changed` | `// TRACES: UR-069 \| DR-120 \| UT-103` |
+
+New requirement rows to add to [requirements.md](../requirements.md):
+
+- **UR-067** — Browse favourited media across libraries (page, home rows, in-library filter).
+- **UR-068** — Mark/unmark favourites from browse and detail surfaces, not only the player.
+- **UR-069** — Favourite state stays consistent with the server in both directions.
+- **DR-113 … DR-120** — as tabled above.
+- **JA-033** — Query favourite items (`Filters=IsFavorite`).
+- **JA-034** — Read `UserData` from item responses.
+
+## Implementation notes (as built)
+
+Two things landed differently from the design above, both forced by where the
+`AppHandle` lives:
+
+1. **The `favorites-changed` event is emitted from the command layer, not the
+   repository.** `HybridRepository` has no `AppHandle` — the same reason
+   `search-event` is emitted from `repository_search`. `repository_get_favorites`
+   therefore does the two-phase read itself (cache leg returned, server leg
+   spawned) and diffs the two id sets via `changed_favorite_ids`, which is
+   extracted and unit-tested (UT-107) rather than buried in the spawn.
+2. **The drain hooks the existing `connectivity:reconnected` event** via
+   `app.listen` in `commands/favorites.rs`, rather than reaching into
+   `ConnectivityMonitor` (which knows nothing about repositories). It drains
+   through a narrow `FavoriteSink` trait so it can be tested against a recording
+   double instead of a forty-method `MediaRepository` mock.
+
+3. **The command falls back to `HybridRepository::get_favorites` when nothing is
+   cached.** The two-phase read alone paints "Nothing favourited yet" on a fresh
+   install and corrects it a server round trip later, which is a wrong answer
+   shown to the user. An empty cache leg therefore defers to the repository's
+   own cache-first-then-server read. That read was also fixed to *save through*
+   on a server hit — without it the page re-queried the server on every visit
+   and the DR-114 mirror was never filled by this path.
+
+Also as built: `DatabaseService` is not object-safe (generic methods), so the
+drain takes `Arc<RusqliteService>` like the rest of the storage code, and
+`get_items`' endpoint construction was extracted to `build_get_items_endpoint`
+so the `favorites_only` filter could be asserted without an HTTP server.
+
+**Not built:** the full-player heart. ux-flows §3.2 lists one among the full
+player's secondary controls and it remains unbuilt — recorded as a known
+deviation in ux-flows §5C.5 rather than silently dropped.
+
+## Notes for the implementer
+
+- A parallel Claude session may be active in this repo — `git diff` before "repairing" unexpected changes.
+- Do **not** try to reuse `get_items` with an empty `ParentId` for cross-library favourites; that endpoint is built as `?ParentId={}` ([online.rs:731](../../src-tauri/src/repository/online.rs#L731)) and an empty value is not a reliable "all libraries" request. Use `get_favorites`.
+- `SearchScope` is reused rather than a new `FavoritesScope` so there is one taxonomy expansion in the codebase, not two that can drift. If the name grates once favourites ship, rename the type across search + favourites in one commit — don't fork it.
+- `SearchScope::All` returns `None` from `item_types()` **on purpose**; callers must omit `IncludeItemTypes` entirely rather than sending a union (see the doc comment at [types.rs:316](../../src-tauri/src/repository/types.rs#L316)).
+- Ship order that keeps each step demonstrable: §1+§2 (state becomes visible) → §5 hearts (marking becomes possible) → §3+§5 browse surfaces (finding becomes possible) → §4 drain.
+- Regenerate `bindings.ts` after the Rust types change; never hand-edit it.

docs/specs/libmpv2-migration.md

@@ -0,0 +1,169 @@
+# Spec: Migrate to libmpv2 and declare the project licence
+
+**Status:** Proposed
+**Requirements:** UR-003 → IR-003 (revises the MPV integration); no new user-facing behaviour
+**UX spec:** n/a
+**Supersedes / revises:** dependency and licensing housekeeping identified in [playback-backend-unification.md](playback-backend-unification.md)
+
+## Summary
+
+Two related pieces of housekeeping that block or complicate later work:
+
+1. Replace the abandoned `libmpv` crate (pinned to a git branch) with the
+   maintained `libmpv2`.
+2. Add a `LICENSE` file. The project has none, which leaves its legal status
+   undefined while it links GPL-licensed libmpv.
+
+Neither changes user-visible behaviour. Both are prerequisites for
+[windows-native-audio-backend.md](windows-native-audio-backend.md).
+
+## Motivation
+
+### The dependency is dead
+
+```toml
+# src-tauri/Cargo.toml
+libmpv = { git = "https://github.com/ParadoxSpiral/libmpv-rs.git", branch = "master" }
+```
+
+- crates.io `libmpv` 2.0.1 was published **2020-09-29**.
+- The upstream repo's last commit was **2023-01-08**; nothing since was released.
+- We pin a git *branch*, so builds are not reproducible — the same lockfile-less
+  checkout can resolve differently over time, and CI has no protection if the
+  branch moves or the repo disappears.
+
+`libmpv2` (kohsine/libmpv2-rs) is a maintained fork of exactly this crate:
+6.0.0 released **2026-05-12**, ~23.5k recent downloads against the original's
+~1.1k, releases roughly quarterly since 2024.
+
+### The project has no licence
+
+There is no `LICENSE`/`COPYING` file and `src-tauri/Cargo.toml` has no `license`
+field. The project is open source and will never be commercial, so this is purely
+an omission — but it matters because we link libmpv, and "no licence" defaults to
+*all rights reserved*, which is incompatible with distributing a GPL-derived
+work.
+
+## Design
+
+### Part 1 — licence
+
+**Use GPLv3.** This is forced, not chosen:
+
+- mpv's default build is **GPLv2-or-later**, so the combined work must be
+  GPL-compatible.
+- Apache-2.0 is **GPLv2-incompatible** (patent-termination and indemnification
+  clauses) but GPLv3-compatible.
+- A scan of the dependency tree found Apache-2.0-**only** crates with no
+  alternative arm — most importantly **`tao`** (Tauri's own windowing crate),
+  plus `sync_wrapper`, `gethostname`, and `ring` (Apache-2.0 AND ISC).
+
+`tao` is unavoidable in a Tauri app, so GPLv2 is unavailable. Exercising mpv's
+"or later" option puts the combination at **GPLv3**.
+
+Actions:
+- Add `LICENSE` containing the GPLv3 text.
+- Add `license = "GPL-3.0-or-later"` to `src-tauri/Cargo.toml` and `license` to
+  `package.json`.
+- Note in the README that the binary links libmpv (GPLv2+) and FFmpeg.
+
+Because the project is open source, we use mpv's **default GPL build** — no
+`-Dgpl=false`, no LGPL FFmpeg build, and none of the LGPL §6 relinking analysis
+that a proprietary app would need. We keep VAAPI/VDPAU/X11 and every GPL FFmpeg
+filter.
+
+🔴 Never build FFmpeg with `--enable-nonfree` — that produces a binary that is
+**unredistributable under any licence**, open source or not.
+
+### Part 2 — libmpv → libmpv2
+
+```toml
+# Linux (and later Windows, per the Windows audio spec)
+libmpv2 = "=6.0.0"
+```
+
+Pin exactly: `libmpv2` has broken its API in **every** major release.
+
+Breaking changes to expect, from the changelog:
+
+| Version | Change | Impact here |
+|---|---|---|
+| 4.0.0 | Removed command helper methods — call `mpv.command(...)` directly | Low; we already use `command`/`set_property` |
+| 5.0.0 | Removed `mpv_node` support entirely (properties return strings; parse JSON yourself); `EventContext` folded into `Mpv`; `ProtocolContext` → `Protocol` | **Medium** — `start_event_loop` uses `create_event_context()`; check whether that call still exists |
+| 6.0.0 | `RenderContext::new()` → `Mpv::create_render_context()`; `'static` bound on `OpenGLInitParams`; render context now borrows `Mpv` (fixes a use-after-free) | **None** — we do not use the render API |
+
+The last row matters: we run mpv audio-only (`video = no`), so the entire render
+surface is irrelevant to us. Consider disabling the default `render` feature to
+reduce build surface.
+
+The main porting work is the event loop in `mpv_backend.rs` — `wait_event`,
+`disable_deprecated_events`, and the `FileLoaded` / `PlaybackRestart` /
+`PropertyChange` / `EndFile` handling, given 5.0.0 folded `EventContext` into
+`Mpv`.
+
+Everything else — `set_property` calls, the `af` filter graph, the 250ms position
+thread, the seek-suppression window — should port unchanged.
+
+## Layer assignment
+
+No logic moves. This is a dependency swap plus a licence file; the
+`PlayerBackend` trait boundary is untouched.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| mpv event → `PlayerStatusEvent` mapping | Rust (unchanged) | Already correct; only the binding API beneath it changes. |
+
+## Out of scope
+
+- Any behaviour change. If playback behaves differently after this, that is a bug.
+- Windows support — separate spec, but this must land first.
+- Adopting the render API. We are audio-only on mpv.
+- Re-licensing decisions beyond adding the file the project already implies.
+
+## Acceptance criteria
+
+- [ ] `LICENSE` (GPLv3) present; `license` field set in `Cargo.toml` and `package.json`.
+- [ ] A full dependency-licence audit has been run (`cargo install cargo-license && cargo license`) and confirms no GPLv3-incompatible dependency. *(The scan behind this spec resolved 441 of 575 crates from the local registry cache; the remaining 134 are unverified.)*
+- [ ] `libmpv` git dependency removed; `libmpv2` pinned to an exact version.
+- [ ] Linux audio playback works identically: play/pause/seek/volume, queue advance, gapless, EQ, normalization, sleep timer.
+- [ ] Position updates still arrive at 250ms; the 150ms post-seek suppression still prevents the jump-to-zero glitch.
+- [ ] `EndFile` still emits `PlaybackEnded` only for EOF (not STOP/QUIT/ERROR) — autoplay depends on this.
+- [ ] Builder image updated if the libmpv dev package requirement changed; **no toolchain install added to any CI step**.
+- [ ] `bun run check`, `bun run test`, `bun run check:boundary` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+
+## Testing
+
+The existing `mpv_backend_test.rs` plus the `build_af_filter`,
+`eq_filter_entries`, and `normalize_filter_entry` tests are the regression net —
+they must pass unchanged, since none of them touch the binding API.
+
+The event loop has no unit tests and is where the risk concentrates. Verify
+manually on Linux:
+
+1. Play → pause → play; confirm position does not flash to 0:00 (the known
+   playing-event regression).
+2. Seek mid-track; confirm no jump-to-zero within 150ms.
+3. Let a track end naturally; confirm autoplay advances (exercises `EndFile` EOF).
+4. Press stop; confirm autoplay does **not** advance.
+5. Sleep-timer expiry; confirm it stops without triggering autoplay.
+
+Cases 3–5 are the ones most likely to break silently, and each corresponds to a
+bug already fixed once in this codebase.
+
+## TRACES
+
+- `MpvBackend` construction / event loop → existing `// TRACES: UR-003 | IR-003`, unchanged
+- No new requirement IDs; this is a dependency migration.
+
+## Notes for the implementer
+
+- Do this **before** the Windows audio backend.
+- Read the 4.0/5.0/6.0 changelogs before writing code — the crate has broken API
+  in every major release, most recently two months before this spec.
+- The crates.io `repository` field for `libmpv2` points at `kohsine/libmpv-rs`,
+  but the repo was renamed to **`libmpv2-rs`**; the old raw URLs 404.
+- `libmpv2-sys` ships pregenerated bindings and vendored headers, so no libclang
+  is needed at build time — relevant to keeping the builder image thin.
+- A parallel Claude session may be active — `git diff` before "repairing"
+  unexpected changes.

docs/specs/library-mosaic.md

@@ -0,0 +1,125 @@
+# Spec: Library mosaic (library overview + home shortcuts)
+
+**Status:** Implemented
+**Requirements:** UR-075 → DR-174, DR-175 (with UR-067 → DR-117 extended)
+**UX spec:** [ux-flows.md](../ux-flows.md) §5C.2 (Favourites)
+
+## Summary
+
+The library overview and the home "Your Libraries" strip stop being fixed-shape
+grids and become a **mosaic**: rows share one height, and each tile is as wide as
+its own artwork is. A square music cover, a 16:9 library backdrop and a 2:3
+poster sit in the same row at their own proportions instead of all three being
+cropped into whichever box the grid picked. Favourites gain a tile per category,
+placed beside the library that category belongs to, alongside the existing
+cross-library entry.
+
+## Motivation
+
+Every surface here shows artwork of more than one shape. The grid resolved that
+by choosing one shape and cropping to it — and the home strip said so out loud:
+
+> Uniform 16:9 artwork so music (square) and video libraries line up at the same
+> height in this mixed row.
+
+Lining them up is right; cropping the covers to do it is not. Holding the
+**height** fixed and letting the **width** vary achieves the same alignment with
+no crop at all, which is the whole idea of a justified layout.
+
+Favourites had one entry for everything. With per-category tiles, "my favourite
+albums" is one tap from the library page rather than a tap plus a tab.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Collection type → favourites category (`movies` → Movies, `livetv` → none) | **Rust** | Jellyfin vocabulary. It changes when Jellyfin renames a collection type, never when this page is redesigned — the same test that put `SearchScope::item_types` in Rust. Shipping it in Svelte would have re-created the leak [scoped-search-boundary.md](scoped-search-boundary.md) exists to document. |
+| Which scopes exist at all (`SearchScope`) | **Rust** | Already there; unchanged. |
+| Row packing: heights, widths, justification, clamping | Frontend | Geometry of a rendered page. It changes when the layout is redesigned and never when the API does. |
+| Assumed artwork shape before the image loads (music = square, else wide) | Frontend | The shape of a *picture*, not a taxonomy — and it is only a starting guess, overruled by the decoded image. |
+| Tile labels, order, and showing a category's tile once | Frontend | Pure presentation: wording and placement. |
+
+Borderline row: the "assumed artwork shape" is a per-collection-type default, and
+any per-collection-type table deserves suspicion. The tie-breaker: it does not
+decide *what a category means* or what is fetched — it seeds a pixel dimension
+that the loaded bitmap immediately corrects. Getting it wrong costs one re-pack,
+not a wrong result. The scope mapping, which does decide what is fetched, went to
+Rust.
+
+## Design
+
+### Wire
+
+`Library` gains one optional field, derived at construction:
+
+```rust
+pub struct Library {
+    pub id: String,
+    pub name: String,
+    pub collection_type: String,
+    pub image_tag: Option<String>,
+    pub favorites_scope: Option<SearchScope>,   // ← new
+}
+
+impl SearchScope {
+    pub fn for_collection_type(collection_type: &str) -> Option<SearchScope>;
+}
+```
+
+```ts
+type Library = { …; favoritesScope?: SearchScope | null }
+```
+
+`Library::new` derives it, so the four construction sites (online views, two
+offline cache reads, tests) cannot forget it. `None` is *omitted* from the JSON,
+not sent as null. No new command, no new event.
+
+### Layout
+
+`src/lib/components/library/mosaic.ts` — pure, no DOM:
+
+- `layoutMosaic(items, { containerWidth, targetHeight, gap })` → rows of tiles
+  with pixel boxes. Tiles join a row until the height needed to fill the width
+  drops to the target; the row closes there and is justified to the container
+  width, the rounding remainder absorbed by its widest tile. The **last row is
+  not justified** (one leftover tile would inflate into a banner) — it sits at
+  the target height, left-aligned.
+- `layoutMosaicStrip(items, height)` → the same rule as one fixed-height row, for
+  a horizontally scrolling shelf.
+- `mosaicTargetHeight(containerWidth)` → the row height chosen when the caller
+  doesn't pick one. Bounded so a phone still fits two tiles across and a desktop
+  doesn't turn each library into a billboard.
+- Ratios are clamped to a band (0.5–2.5) so one panorama can't own a row.
+
+`MosaicGrid.svelte` supplies the two things only the DOM knows — the measured
+container width (`bind:clientWidth`) and the artwork's decoded ratio — and
+renders the caller's `tile` snippet. `CachedImage` gained an `onNaturalSize`
+callback for the second. Measured ratios are committed in one debounced batch
+(120 ms): artwork arrives over several hundred milliseconds and re-packing per
+image would shuffle the grid under the pointer.
+
+`MosaicTile.svelte` draws one tile at an exact pixel box, with its label written
+**over** the bottom of the artwork. A caption below the box would add height the
+layout didn't compute, and a caption that wrapped to two lines would break the
+row alignment the mosaic exists to provide.
+
+### Composition
+
+`libraryMosaic.ts` (pure, tested) builds the tile list: the cross-library
+favourites entry first, then each library followed by its own category tile. A
+category appears **once** — two movie libraries share one favourites list, so a
+tile each would be two tiles to the same place. A library whose `favoritesScope`
+is absent (Live TV, channels, books) gets no tile rather than one opening an
+unfiltered list.
+
+Home uses the same tiles in `layout="strip"` but **without** the favourites tiles:
+home already carries Favourite Movies / Shows / Music rows of its own, and a
+second entry point in the strip above them would be redundant.
+
+## Out of scope
+
+- The item grids inside a library (`/library/movies`, `/library/music/albums`, …).
+  Those show one item type each, so a uniform grid crops nothing; the mosaic buys
+  them nothing but reflow.
+- Backdrop/collage artwork for libraries with no image of their own.
+- Reordering or pinning libraries.

docs/specs/playback-backend-unification.md

@@ -0,0 +1,226 @@
+# Spec: Playback backend unification — findings and strategy
+
+**Status:** Accepted (analysis; no code changes)
+**Requirements:** IR-004, UR-031, UR-032, UR-033 — revises the "Platform Playback Backend Parity" issue in requirements.md
+**UX spec:** n/a
+**Supersedes / revises:** informs [android-native-video-spike.md](android-native-video-spike.md), [android-audio-settings-parity.md](android-audio-settings-parity.md), [windows-native-audio-backend.md](windows-native-audio-backend.md)
+
+## Summary
+
+This spec records the outcome of an investigation into unifying JellyTau's
+playback backends (Linux/MPV, Android/ExoPlayer, Windows/webview) onto a single
+engine with hardware acceleration everywhere. **The conclusion is that video
+cannot be unified onto a native engine, and should not be attempted.** Audio
+*can* be, and that is where the remaining specs direct effort.
+
+No code changes follow from this spec directly. It exists so the decision is
+written down with its evidence, and so a future session does not re-run the same
+investigation.
+
+## Motivation
+
+The requirements doc carries a "Platform Playback Backend Parity" issue noting
+that audio settings work on Linux but not Android, and proposing eventual
+convergence. The natural next question — "should we just run one engine
+everywhere?" — needed answering before spending effort on per-backend patches.
+
+The investigation also surfaced that several statements in requirements.md and in
+code comments are factually wrong. Those corrections are part of the deliverable.
+
+## Findings
+
+### 1. The current architecture is not what the docs describe
+
+| Platform | Audio | Video |
+|----------|-------|-------|
+| Linux | MPV (native, **audio-only**) | webview `<video>` + hls.js |
+| Android | ExoPlayer (native) | **webview `<video>` + hls.js** |
+| Windows | webview `<audio>` | webview `<video>` + hls.js |
+
+Two surprises:
+
+- **MPV never decodes video.** `mpv_backend.rs` sets `video = no` and
+  `audio-display = no` at construction. Linux video has always been the webview.
+  Correspondingly, `player_play_item` deliberately does *not* load into MPV on
+  Linux (it calls `set_current_item`, which only updates the queue).
+- **Android video is also the webview.** `createAdapter()` in
+  `src/lib/player/adapters/index.ts` hardcodes `const effectiveKind = "html5"`
+  and does `void backendKind`, discarding the `use_html5_element` signal that
+  `get_player_status` computes in Rust. `NativePlayerAdapter` is dead code, and
+  ExoPlayer's `SurfaceView` path in `JellyTauPlayer.kt` is unreachable.
+
+So video is *already* unified — on HTML5, everywhere, by accident of that
+hardcode — and on the path without hardware decoding on Android.
+
+### 2. Native video cannot be composited with a Tauri webview
+
+This is the load-bearing finding. It is **not** an mpv limitation; it defeats
+every candidate engine identically:
+
+- **mpv**: `tauri-plugin-libmpv`'s own platform table reads Linux ⚠️
+  *"Experimental. Window embedding is not working."*
+- **GStreamer** (wry discussion #284, 2024): *"Gstreamer was rendering above the
+  surface and covering all html elements."*
+- **libVLC** (tauri discussion #6343, 2024): *"I had to render the webview in a
+  child window though because vlc kept rendering on top of it."*
+
+Root cause, from Tauri maintainer amrbashir (tauri#9220, 2024-03-30):
+
+> "we are limited to using Webkit2GTK on Linux and that requires a GTK window.
+> While possible to add a GTK widget as a child X11 window inside raw X11 window,
+> this is however a bit hacky and **it is not possible on Wayland at all**."
+
+WebKitGTK, WebView2, and Android WebView each draw into their own compositor
+surface. A native video surface is either entirely above or entirely below the
+webview; it cannot interleave with HTML. Every working example in the ecosystem
+is the same hack — a separate child window position-synced to a
+`getBoundingClientRect()` div — which breaks on resize, scroll, and any UI drawn
+over the video. For JellyTau that means the controls, subtitle overlay, and
+mini-player.
+
+The most recent comment on tauri#6343 (2026-05-23) confirms it is still unsolved:
+
+> "I'm faking it and the window is not truly embedded, basically when the parent
+> moves or resizes I reset the position and size of the libmpv window to align it
+> with an HTML div."
+
+**The principle to carry forward: audio can unify on a native engine; video
+cannot, because video needs a surface and the webview owns the surface.**
+
+### 3. mpv would regress streaming quality
+
+mpv has **no adaptive bitrate**. It delegates HLS to FFmpeg's demuxer, which
+selects one variant at open time and never adapts; mpv#3548 (2016) requested ABR
+and it never landed. `--hls-bitrate` is a static picker defaulting to `max`.
+
+The webview path already has real ABR via hls.js. Moving video to mpv would be a
+**downgrade** on every platform — no graceful degradation on weak networks, and
+quality changes requiring teardown and reload.
+
+### 4. Crossfade is architecturally blocked on mpv
+
+mpv's audio chain is single-stream. FFmpeg's `acrossfade` is an `N→A` filter
+requiring two input streams, so there is no second input to feed it. Real
+crossfade needs **two libmpv instances** with manually ramped volumes. Upstream
+maintainer response (mpv#4512, closed three minutes after opening):
+
+> "No. I also find crossfading stupid and complex, so the likeliness of that
+> happening is low."
+
+GStreamer *could* do it via `audiomixer`. mpv cannot, at any reasonable cost.
+
+### 5. Engine comparison summary
+
+| Criterion | mpv | GStreamer | libVLC |
+|-----------|-----|-----------|--------|
+| Webview compositing | ❌ Linux broken | ❌ same wall | ❌ same wall |
+| Adaptive bitrate HLS | ❌ none | ✅ adaptivedemux2 | ✅ adaptive module |
+| Rust bindings | ⚠️ `libmpv2` active; our pin is dead | ✅ `gstreamer-rs` excellent | ❌ `vlc-rs` abandoned (2018) |
+| Windows cross-MSVC | ⚠️ prebuilt DLL | ❌ pkg-config vs cargo-xwin | ❌ no better |
+| Android packaging | ✅ Maven AAR (used by Findroid) | ⚠️ Cerbero/NDK, painful | ✅ mature AAR |
+| ASS/SSA subtitles | ✅ libass built in | ✅ libass | ✅ libass |
+| Crossfade | ❌ impossible | ✅ `audiomixer` | ⚠️ unclear |
+
+Every candidate fails the first row, which is the disqualifying one.
+
+### 6. Two further options ruled out
+
+**Webview `<audio>`/`<video>` everywhere** (i.e. delete the native audio backends
+too) is dead on Android: `navigator.mediaSession` is *deliberately compiled out*
+of Android WebView (Chromium CL 2613133003), so lockscreen/media-notification
+control would be impossible. Chromium has also never shipped `audioTracks`. It
+remains fine for Windows *video*, which is what we already do.
+
+**FFmpeg-direct / Rust-native** (`ffmpeg-next`, `rsmpeg`, Symphonia) is not
+close: the safe bindings do not expose hardware decode at all, `ffmpeg-next` is
+self-declared maintenance-only, and Symphonia lacks HE-AAC and gapless AAC. This
+is a multi-person-year path to reach parity with what we already have.
+
+### 7. If libmpv is ever revisited on Android
+
+Recorded so the next investigation starts from evidence rather than repeating the
+search. The `dev.jdtech.mpv:libmpv` AAR — maintained by Findroid's author, i.e.
+another Jellyfin Android client — was inspected directly:
+
+- `libmpv.so` exports the full 54-function `mpv_*` C API with **zero `Java_`
+  symbols**; JNI is a separate optional ~19 KB `libplayer.so`. So it is drivable
+  from Rust without a Java shim. (This is precisely what disqualifies libVLC,
+  whose Android video path hard-requires a Java `AWindow` jobject.)
+- ~23 MB/ABI, versus libVLC's ~46 MB/ABI.
+- 🔴 **The published AAR is built `--enable-gpl --enable-version3` — it is
+  GPLv3**, not LGPL. Fine for us (see [libmpv2-migration.md](libmpv2-migration.md)),
+  but it would be a hard constraint for anyone shipping closed source, and an
+  LGPL rebuild would be your own build to own.
+- Top unverified risk if anyone tries this: whether `libmpv2-sys` can
+  cross-compile for `aarch64-linux-android` against that prebuilt `.so`. No
+  working example of `libmpv2` on Android was found.
+
+None of this changes the verdict — the cost is the MediaSession/foreground-service
+rewrite, not the bindings.
+
+## Decision
+
+1. **Do not unify video onto a native engine.** Video stays in the webview with
+   hls.js on all platforms. This is not a compromise — it is the configuration
+   that falls out of the compositing constraint, and it is the only one that
+   gives us ABR for free.
+2. **Android native video is worth a bounded spike anyway** — not for
+   unification, but because ExoPlayer's `SurfaceView` path already exists and
+   would restore hardware decode plus ASS/SSA subtitles. See
+   [android-native-video-spike.md](android-native-video-spike.md).
+3. **Audio parity is the real gap** and is achievable without touching any of the
+   above. See [android-audio-settings-parity.md](android-audio-settings-parity.md)
+   and [windows-native-audio-backend.md](windows-native-audio-backend.md).
+4. **Migrate the dead libmpv pin** regardless of any of this. See
+   [libmpv2-migration.md](libmpv2-migration.md).
+
+## Corrections to existing docs
+
+These are factual errors found during the investigation. Fixing them is in scope
+for this spec.
+
+| Location | Says | Actually |
+|----------|------|----------|
+| `requirements.md` UR-031 (line ~44) | "Done (Linux only)" | Not implemented on any platform. |
+| `requirements.md` DR-034 (line ~196) | "Done (Linux only)" | Not implemented anywhere — `mpv_backend.rs` has a bare `// TODO: Implement crossfade via MPV audio filters if needed`. Architecturally blocked on mpv (finding 4). |
+| `requirements.md` parity matrix | Crossfade ✅ Linux / ❌ Android | ❌ / ❌ |
+| `requirements.md` parity matrix | (no EQ row) | EQ is also Linux-only — `build_af_filter`/`eq_filter_entries` exist only in `mpv_backend.rs`. Same root cause, same fix. |
+| `nativeAdapter.ts:11-14` | Native Android video "blocked upstream by tauri#10152" | tauri#10152 is a stale *feature request*, dead since 2024-07-01. The capability shipped in tauri commit `27d01834` (2024-09-02). Not a blocker. |
+
+## Layer assignment
+
+No new logic. The one boundary observation worth recording:
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Which video backend a platform uses (`use_html5_element`) | Rust | Already correctly computed in `get_player_status`. The frontend currently *discards* it — that is the bug, not the design. Restoring it means the frontend consumes a backend decision rather than making its own. |
+
+## Out of scope
+
+- Any code change. This spec is analysis; the sibling specs carry the work.
+- iOS/macOS. Not current targets.
+- Replacing hls.js.
+
+## Acceptance criteria
+
+- [ ] `requirements.md` DR-034 status corrected; parity matrix updated (crossfade ❌/❌, EQ row added).
+- [ ] Stale tauri#10152 comment in `nativeAdapter.ts` corrected.
+- [ ] The four sibling specs exist and are linked from here.
+
+## Testing
+
+n/a — documentation only.
+
+## TRACES
+
+No new code. Requirement text changes only; DR-034's status line is the one
+substantive edit.
+
+## Notes for the implementer
+
+- The evidence above was gathered in July 2026. The compositing constraint has
+  been stable since 2021 (wry#284) and is maintainer-declared unfixable, so it is
+  unlikely to change soon — but if someone revisits this, tauri#6343 and wry#284
+  are the threads to re-read first.
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes.

docs/specs/playback-docs-corrections.md

@@ -0,0 +1,198 @@
+# Spec: Playback documentation corrections
+
+**Status:** Proposed
+**Requirements:** revises the status of DR-034; corrects the parity matrix in [requirements.md](../requirements.md)
+**UX spec:** n/a
+**Supersedes / revises:** implements the "Corrections to existing docs" section of [playback-backend-unification.md](playback-backend-unification.md)
+
+## Summary
+
+Fix four factual errors in the requirements doc and the player source comments,
+all found while investigating backend unification. Each claims something the code
+does not do. Small change, but they are actively misleading: two of them assert a
+feature is implemented when it is implemented nowhere, and one cites an upstream
+blocker that no longer exists.
+
+Documentation and comments only — no behaviour change.
+
+## Motivation
+
+These errors compound. DR-034 reads "Done (Linux only)", so a future session
+planning Android parity would reasonably assume crossfade exists on Linux and
+only needs porting — when in fact it is unimplemented everywhere *and*
+architecturally blocked on the engine it supposedly runs on. Likewise the
+tauri#10152 comment has been discouraging work on Android native video since the
+upstream capability shipped in September 2024.
+
+## The corrections
+
+### 1. DR-034 status is wrong
+
+`requirements.md` line ~196:
+
+```
+| DR-034 | Crossfade engine with configurable duration (0-12s) | Player | UR-031 | Done (Linux only) |
+```
+
+The code:
+
+```rust
+// src-tauri/src/player/mpv_backend.rs, in set_audio_settings
+// TODO: Implement crossfade via MPV audio filters if needed
+```
+
+That is the entire crossfade implementation. `crossfade_duration` is plumbed
+through `AudioSettings` and clamped to 0–12s, but no backend ever acts on it.
+
+**Change to:** `Not implemented (blocked on MPV — see playback-backend-unification.md)`
+
+Worth stating *why* in the requirements entry, because it is not a scheduling
+gap: mpv's audio chain is single-stream, and FFmpeg's `acrossfade` is an `N→A`
+filter needing two inputs. Real crossfade requires two libmpv instances with
+manually ramped volumes. Upstream declined the feature (mpv#4512).
+
+### 1b. UR-031 status is wrong for the same reason
+
+`requirements.md` line ~44:
+
+```
+| UR-031 | Crossfade between audio tracks | Low | Done (Linux only) |
+```
+
+Same error one level up: the *user* requirement is also marked done. Since no
+backend implements crossfade, UR-031 is not satisfied on any platform.
+
+**Change to:** `Not implemented (blocked — see DR-034)`
+
+Note line ~517 of the same file (`UR-031 (Crossfade), UR-032 (Gapless),
+UR-033 (Normalization) only work on Linux`) inherits the error — crossfade works
+nowhere, so it should read UR-032/UR-033 only.
+
+### 2. Parity matrix crossfade row is wrong
+
+```
+| Crossfade | ✅ | ❌ | Gap |
+```
+
+**Change to** `| Crossfade | ❌ | ❌ | Not implemented |` — it is not a
+platform-parity gap, it is an unbuilt feature.
+
+### 3. Parity matrix is missing the equalizer
+
+The matrix lists crossfade, gapless, and normalization but omits the EQ, which
+has the same Linux-only shape and the same root cause (`ExoPlayerBackend` not
+overriding `set_audio_settings`). `build_af_filter` and `eq_filter_entries` exist
+only in `mpv_backend.rs`; there is no equalizer code in the Android tree.
+
+**Add:** `| Equalizer (10-band) | ✅ | ❌ | Gap |`
+
+### 4. `nativeAdapter.ts` cites a stale blocker
+
+`src/lib/player/adapters/nativeAdapter.ts:11-14` states native Android video is
+blocked upstream by tauri#10152 (transparent webview / SurfaceView compositing).
+
+tauri#10152 is open but **dead since 2024-07-01**, and it is a *feature request*
+that `WebviewWindowBuilder::transparent` was desktop-only — not a report that
+compositing is broken. The capability shipped in tauri commit `27d01834`
+(2024-09-02), which moved `transparent()` into the cross-platform impl block with
+only the tao call `#[cfg(desktop)]`-fenced. It landed as a clippy cleanup, so the
+issue was never closed. Separately, the black/white-screen bug (tauri#8381,
+tauri#9408) was a broken JNI signature for `setBackgroundColor`, fixed in wry
+0.39.4; we ship wry 0.55.x.
+
+**Change to:** a comment stating the adapter is currently unreachable because
+`createAdapter` hardcodes the HTML5 kind, that transparency is no longer an
+upstream blocker, and that
+[android-native-video-spike.md](android-native-video-spike.md) tracks whether
+SurfaceView compositing actually works. Be explicit that *nobody has
+demonstrated* SurfaceView-behind-WebView on Tauri Android — nothing upstream
+blocks it, and nothing upstream proves it.
+
+### 5. Platform capability is signalled three incompatible ways
+
+Not a doc error — a real inconsistency found during the same investigation, worth
+recording here even though fixing it needs its own change.
+
+Which backend a platform uses is currently expressed three ways:
+
+1. Rust `#[cfg]` gates in `player/mod.rs` and `create_player_backend` — the truth.
+2. The `useHtml5Element` / `VideoBackend` value from `get_player_status` — which
+   the frontend discards (see the spike spec).
+3. **Frontend user-agent sniffing** in `src/lib/services/webviewAudio.ts:30-41`:
+
+```ts
+const ua = navigator.userAgent.toLowerCase();
+const isAndroid = ua.includes("android");
+const isLinux = ua.includes("linux") && !isAndroid;
+return !isAndroid && !isLinux;
+```
+
+The comment says it is "matching the Rust cfg gate" — i.e. the frontend
+re-derives a backend decision from the user-agent string and hopes it stays in
+sync. That is the frontend deciding *which backend exists*, which is domain
+knowledge, not presentation. It also breaks silently the moment a new target is
+added or a webview's UA changes.
+
+**This is a boundary leak of the same family the spec-review checklist exists to
+catch**, even though `check:boundary`'s tripwire (item-type arrays) does not
+match it. Rust already computes the answer; the frontend should consume it.
+
+Not fixed by this spec — it is behavioural, not documentation. It should be
+folded into the spike spec's factory rework, where the same
+"consume Rust's decision instead of re-deriving it" change is already in scope.
+
+### Also worth fixing while here
+
+`requirements.md` IR-004 reads "In Progress (basic playback works, audio settings
+missing)". That stays accurate until
+[android-audio-settings-parity.md](android-audio-settings-parity.md) lands, but
+the "Future Fix" list in the parity issue proposes
+`ConcatenatingMediaSource` for crossfade — **deprecated in current Media3**. Drop
+that suggestion; the modern approach is a custom `AudioProcessor`.
+
+## Layer assignment
+
+No logic. Documentation and comments only.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| — | — | No logic introduced or moved by this spec. |
+
+## Out of scope
+
+- Implementing crossfade. This spec only stops claiming it exists.
+- Implementing Android audio settings — see the parity spec.
+- Running the Android video spike — see that spec.
+- Rewriting the architecture docs. `docs/architecture/05-platform-backends.md`
+  should be re-read for the same class of error, but that is a larger pass.
+
+## Acceptance criteria
+
+- [ ] DR-034 status corrected, with the blocking reason stated.
+- [ ] UR-031 status corrected (line ~44), and the "only work on Linux" line (~517) no longer lists crossfade.
+- [ ] Parity matrix: crossfade ❌/❌; equalizer row added.
+- [ ] `ConcatenatingMediaSource` suggestion removed from the "Future Fix" list.
+- [ ] `nativeAdapter.ts` comment corrected and pointing at the spike spec.
+- [ ] `bun run check` and `bun run test` pass (a comment change still touches TS).
+- [ ] `bun run traces:markdown` re-run if requirement text changed.
+
+No Rust changes, so the `cargo` gates do not apply.
+
+## Testing
+
+None beyond the standard gates — no behaviour changes. Confirm
+`bun run traces:markdown` regenerates cleanly, since DR-034's row is referenced
+by the traceability matrix.
+
+## TRACES
+
+No code implementing requirements changes; no TRACES comments to add or update.
+The DR-034 row in `docs/traceability.md` will regenerate with the corrected text.
+
+## Notes for the implementer
+
+- Do **not** silently delete DR-034. The requirement (UR-031 crossfade) is still
+  wanted; it is the *status* that is wrong. Keeping the row with an honest status
+  and a reason is the point.
+- A parallel Claude session may be active — `git diff` before "repairing"
+  unexpected changes.

docs/specs/player-facade-enforcement.md

@@ -0,0 +1,258 @@
+# Spec: Enforce the unified player boundary
+
+**Status:** Proposed
+**Requirements:** DR-095 (new); relates to UR-005 and the unified-player-boundary
+principle in CLAUDE.md and [02-svelte-frontend.md](../architecture/02-svelte-frontend.md)
+**UX spec:** n/a — refactor, no user-visible change.
+**Supersedes / revises:** n/a
+
+## Summary
+
+The stated principle is that UI controls playback **only** through
+`playerController` ([src/lib/player/index.ts](../../src/lib/player/index.ts)),
+never by calling `commands.player*` directly. There are **52 direct call sites
+outside** that facade. This spec routes the genuine playback-control calls
+through the facade, narrows the principle's wording so it stops forbidding
+things it never meant to forbid, and adds the lint rule that keeps it true —
+because this rule is the one design principle in the audit with **no automated
+check at all**, and it is also the one that drifted furthest.
+
+## Motivation
+
+Direct `commands.player*` usage outside `src/lib/player/`, by file:
+
+| File | Sites |
+|---|---|
+| [queue.ts](../../src/lib/stores/queue.ts) | 10 |
+| [player/[id]/+page.svelte](../../src/routes/player/[id]/+page.svelte) | 9 |
+| [VideoPlayer.svelte](../../src/lib/components/player/VideoPlayer.svelte) | 8 |
+| [settings/+page.svelte](../../src/routes/settings/+page.svelte) | 5 |
+| [sleepTimer.ts](../../src/lib/stores/sleepTimer.ts) / [auth.ts](../../src/lib/stores/auth.ts) / [autoplay.ts](../../src/lib/api/autoplay.ts) | 4 each |
+| [preload.ts](../../src/lib/services/preload.ts) | 3 |
+| [library/[id]](../../src/routes/library/[id]/+page.svelte), [playerEvents.ts](../../src/lib/services/playerEvents.ts), [playbackMode.ts](../../src/lib/stores/playbackMode.ts) | 1–2 each |
+
+These are **not** equivalent violations, and treating them as one number is why
+the rule has been easy to ignore. Three distinct groups:
+
+**(a) Genuine violations — playback control with a facade method that already
+exists.** `playerStop` ×6, `playerPlayTracks` ×4, `playerSeek` ×2,
+`playerPlayAlbumTrack` ×2, `playerNext`, `playerPrevious`, `playerSkipTo`,
+`playerToggleShuffle`, `playerCycleRepeat`, `playerRemoveFromQueue`,
+`playerMoveInQueue`, `playerAddTrackById`, `playerAddTracksByIds`,
+`playerSetSubtitleTrack`, `playerPlayItem`. The facade exposes `stop()`,
+`seek()`, `next()`, `previous()`, `skipTo()`, `toggleShuffle()`,
+`cycleRepeat()`, `removeFromQueue()`, `moveInQueue()`, `addTrackById()`,
+`addTracksByIds()`, `setSubtitleTrack()`, `playTracks()`, `playAlbumTrack()`,
+`playItem()` — every one of these has a facade equivalent that is simply not
+being called. `queue.ts` is the starkest case: it imports `commands` directly
+and re-implements ten methods the facade already provides.
+
+**(b) Playback control with no facade method.** `playerPlayQueue`,
+`playerGetQueue`, `playerGetStatus`, `playerEnterBackgroundAudio`,
+`playerExitBackgroundAudio`, `playerSetSleepTimer`, `playerCancelSleepTimer`,
+`playerPlayNextEpisode`, `playerCancelAutoplayCountdown`. In scope for the
+principle, but currently *impossible* to comply with — the facade has no surface
+for them. A rule that cannot be followed is not being broken so much as it is
+unfinished.
+
+**(c) Not playback control.** `playerConfigureJellyfin` ×3,
+`playerDisableJellyfin`, `playerGet/SetAudioSettings`,
+`playerGet/SetVideoSettings`, `playerGetEqPresets`,
+`playerGet/SetAutoplaySettings`, `playerGet/SetCacheConfig`,
+`playerPreloadUpcoming`. These are configuration and lifecycle calls that happen
+to live under the `player_` command prefix. The principle is about *who is
+authoritative for playback state* — settings CRUD isn't that.
+
+The audit's read: the rule as written is violated 52 times, which makes real
+drift indistinguishable from acceptable usage, and that ambiguity is what lets
+group (a) persist. Note also that the principle **is** well-honoured where it
+matters most — the read side is clean, with UI reading state exclusively from
+the facade's re-exported stores. The write side is what drifted.
+
+## Layer assignment
+
+Frontend-internal refactor. No domain logic moves and nothing new crosses IPC —
+the same Rust commands are called, through one module instead of many.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Playback command dispatch (adapter routing: native vs HTML5) | Frontend — `src/lib/player/` **only** | Presentation-layer plumbing, but must be centralised: the facade picks between the native backend and the HTML5 `<video>` adapter. A caller bypassing it silently skips that routing. |
+| Playback *authority* (position, pause, rate, track changes) | **Rust / the player** | Unchanged. The player is authoritative; UI is a consumer. This spec does not touch that direction. |
+| Queue mutation commands | Frontend facade → Rust | Rust owns queue state; the facade is the single call path to it. |
+| Player settings CRUD (EQ, video, autoplay, cache) | Frontend, **outside** the facade | Configuration, not playback control — read/written on a settings page with no adapter routing. Explicitly carved out below. |
+| Backend→frontend event handling | `playerEvents.ts` | Already correct. It is the facade's own plumbing, not a bypassing consumer. |
+
+No Jellyfin taxonomy is involved, so no boundary-leak risk.
+
+## Design
+
+### 1. Narrow the principle to what it actually means
+
+Amend CLAUDE.md and [02-svelte-frontend.md](../architecture/02-svelte-frontend.md):
+
+> **Unified player boundary.** UI controls **playback** — transport, queue
+> mutation, track selection, playback initiation — *only* through
+> `playerController`. Player **configuration** commands (`player_*_settings`,
+> `player_configure_jellyfin`, `player_*_cache_config`, `player_preload_upcoming`)
+> are ordinary IPC and may be called directly from settings surfaces.
+
+This is a clarification, not a relaxation: it makes group (c) explicitly fine so
+that a violation count means something. A rule with 52 nominal violations, most
+of them acceptable, provides no signal.
+
+### 2. Fill the facade gaps (group b)
+
+Add to `playerController`, each a thin pass-through preserving current
+behaviour:
+
+```ts
+playQueue, getQueue, getStatus,
+enterBackgroundAudio, exitBackgroundAudio,
+setSleepTimer, cancelSleepTimer,
+playNextEpisode, cancelAutoplayCountdown,
+```
+
+Do this **first** — group (a) cannot be fully migrated while callers still need
+a direct import for a neighbouring call, and a file that imports `commands` for
+one reason will keep using it for others.
+
+### 3. Migrate group (a)
+
+Mechanical: replace `commands.playerX(...)` with `playerController.x(...)`.
+Highest-value first: `queue.ts` (10 sites, all direct facade equivalents), then
+`player/[id]/+page.svelte`, `VideoPlayer.svelte`, `sleepTimer.ts`,
+`playbackMode.ts`, `library/[id]/+page.svelte`.
+
+Two sites need care rather than substitution:
+
+- **`playerEvents.ts`** (`playerOnPlaybackEnded`, `playerStop` in the error
+  path). This module *is* the facade's event plumbing — the counterpart to
+  `index.ts`, inside the boundary conceptually though not by directory. Treat
+  `src/lib/services/playerEvents.ts` as **inside** the boundary and exempt it,
+  rather than making it call the facade that calls back into it. Record this in
+  the lint config with the reason.
+- **`VideoPlayer.svelte`** — registers its own adapter via `setActiveAdapter`.
+  Its `playerStop`/`playerPlayItem` calls interact with adapter lifecycle, and
+  CLAUDE.md's gotcha ("no lifecycle calls after an `await` in `onMount`") applies.
+  Migrate this file **last and on its own**, so an Android seek regression is
+  bisectable to one commit.
+
+### 4. Add the lint rule (the part that makes it stick)
+
+The audit's finding was that principles with working checks held up and
+principles without them drifted. This principle has no check. Add
+`scripts/check-player-boundary.sh`, wired as `bun run check:player-boundary` and
+into `test-all.sh`:
+
+```sh
+# Playback-control commands that MUST go through the facade.
+CONTROL='player(Play|Pause|Toggle|Stop|Seek|Next|Previous|SkipTo|ToggleShuffle|CycleRepeat|RemoveFromQueue|MoveInQueue|SetVolume|ToggleMute|SetSubtitleTrack|SeekVideo|SwitchAudioTrack|PlayTracks|PlayAlbumTrack|PlayItem|PlayQueue|AddTrackById|AddTracksByIds|GetQueue|GetStatus|EnterBackgroundAudio|ExitBackgroundAudio|SetSleepTimer|CancelSleepTimer|PlayNextEpisode|CancelAutoplayCountdown|OnPlaybackEnded)'
+
+# Inside the boundary: the facade and its event plumbing.
+EXEMPT='^src/lib/player/|^src/lib/services/playerEvents\.ts$'
+```
+
+Flag `commands.$CONTROL` in non-test `src/` files outside `EXEMPT`. Config
+commands are deliberately absent from the list, matching §1 — so the check
+encodes the narrowed rule rather than the aspirational one.
+
+An ESLint `no-restricted-syntax` rule would give better editor feedback, but the
+project has no ESLint config; a shell check matches the existing
+`check:boundary` precedent and adds no dependency.
+
+## Out of scope
+
+- Changing playback *behaviour* — pure refactor.
+- The one-directional state principle (audited clean; UI reads from facade
+  stores only).
+- Moving settings CRUD behind the facade (§1 explicitly carves it out).
+- Introducing ESLint.
+- Refactoring `VideoPlayer.svelte`'s 2079 lines generally, beyond its facade
+  call sites.
+- The `commands.player*` calls **inside** `src/lib/player/` — that is the
+  facade doing its job.
+
+## Acceptance criteria
+
+- [ ] `playerController` exposes the group-(b) methods listed in §2.
+- [ ] `grep -rn "commands\.player" src/ --include='*.ts' --include='*.svelte' | grep -v '^src/lib/player/' | grep -v 'playerEvents\.ts' | grep -v '\.test\.' | grep -v bindings.ts`
+      returns **only** configuration commands per §1 — no transport, queue, or
+      playback-initiation call.
+- [ ] `queue.ts` no longer imports `commands` from bindings.
+- [ ] `bun run check:player-boundary` exists, is wired into `test-all.sh`, and
+      passes.
+- [ ] The check **fails** when a `commands.playerStop()` is added to a non-exempt
+      file — verify explicitly, as with the other gates in this batch.
+- [ ] The check does **not** fail on `commands.playerSetAudioSettings()` in
+      `settings/+page.svelte` (the §1 carve-out works).
+- [ ] CLAUDE.md and `02-svelte-frontend.md` carry the narrowed wording, including
+      the config carve-out and the `playerEvents.ts` exemption with its reason.
+- [ ] **No behavioural change**: audio and video playback, queue reorder,
+      shuffle/repeat, sleep timer, background audio, and autoplay all behave as
+      before on **both Linux and Android**.
+- [ ] Android seek and `onMount` lifecycle still correct after the
+      `VideoPlayer.svelte` migration (the known-fragile path).
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `bun run check:boundary` passes.
+- [ ] Changed code carries `// TRACES:` comments.
+- [ ] No Rust change, so no `bindings.ts` regeneration.
+
+## Testing
+
+**Frontend** (`bun run test`):
+- Extend the existing facade tests to cover each new group-(b) method: it
+  forwards to the right command with the right arguments, and routes to the
+  active adapter where applicable.
+- `queue.ts` tests: assert calls land on `playerController`, not `commands`. Mock
+  the facade — a test that mocks `commands` would pass either way and guard
+  nothing.
+- Keep `tauriIntegration.test.ts` and the other IPC param-naming tests green;
+  they cover the camelCase rule this refactor must not disturb.
+
+**Manual** (no automated coverage for these paths):
+- Linux: play/pause/seek/next/prev, queue reorder, shuffle, repeat, sleep timer,
+  transcoded video (HLS), background audio enter/exit.
+- Android: the same, plus lockscreen/MediaSession controls, and **seek after
+  entering the player** — the specific regression CLAUDE.md warns about.
+
+Because this is a pure refactor, the strongest signal is that no test *changes
+expectation*. A test needing its assertions rewritten means behaviour moved —
+investigate rather than update it.
+
+## TRACES
+
+Allocate in `requirements.md`:
+
+- **DR-095** — "UI playback control is routed exclusively through the
+  `playerController` facade (`src/lib/player/`), with `playerEvents.ts` inside
+  the boundary as its event plumbing and player *configuration* commands
+  explicitly outside it; enforced by `scripts/check-player-boundary.sh`."
+  Category: Player. Traces to UR-005. Status: Done on merge.
+
+```typescript
+// src/lib/player/index.ts
+// TRACES: UR-005 | DR-095
+```
+
+New facade tests take `@req-test: UT-089` onward (next free UT is **UT-089**;
+coordinate if landing alongside the sibling specs, which draw from the same
+pool).
+
+## Notes for the implementer
+
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes (CLAUDE.md §Gotchas).
+- **Order matters**: §2 (fill gaps) → §3 (migrate, `VideoPlayer.svelte` last and
+  alone) → §4 (add the check). Adding the check first turns `master` red.
+- 🔴 **`VideoPlayer.svelte`**: no lifecycle calls after an `await` in `onMount` —
+  it flips to HTML5 mode and breaks Android seek. Do not let a mechanical
+  substitution introduce an `await` before a lifecycle call.
+- The facade's `requireHandle()` may throw where a raw `commands` call did not.
+  Check each migrated call site's error handling rather than assuming the
+  try/catch still covers the same cases.
+- `playbackMode.ts` interacts with remote-mode routing (`play_on_session` vs
+  local MPV). Verify remote casting still works after migrating its
+  `playerPlayTracks` call.
+- This spec is deliberately the *lowest* priority of the audit batch: it is the
+  largest diff and the only one carrying real regression risk, while the
+  traceability gate is a few lines and restores a dead safety net.

docs/specs/read-through-media-cache.md

@@ -0,0 +1,219 @@
+# Spec: Two-path media — selectable playback bitrate, independent whole-file download
+
+**Status:** Proposed
+**Requirements:** UR-070, UR-071 → DR-121, DR-122, DR-123, DR-124, DR-125; IR-032
+**UX spec:** player quality selector — needs a `ux-flows.md` section before build
+**Related:** [catalog-index-search.md](catalog-index-search.md),
+[downloads-as-offline-library.md](downloads-as-offline-library.md)
+
+## Summary
+
+Two things that are today tangled become explicitly separate:
+
+- **The playback path** streams at a bitrate the viewer can change from the
+  player. It is ephemeral and its rendition is volatile.
+- **The download path** fetches the whole file at one canonical quality, in the
+  background, independently of whatever playback is doing.
+
+Bytes fetched for playback are kept **only** when the playback rendition happens
+to be the same artifact the download path would produce — i.e. direct play.
+Otherwise playback bytes are discarded and the download path does its own fetch.
+
+## Motivation
+
+The appealing version of this — "stream and download at once, switch when enough
+has arrived" — breaks the moment the viewer can change bitrate. A capture taken
+while the rendition changes underneath it is a splice of two encodings: not a
+playable file, and not something that can be honestly recorded as a download.
+Once bitrate is selectable, one stream cannot serve both jobs.
+
+Separating the paths also removes the thing that made the original idea
+expensive: there is no mid-playback source swap to engineer, because the download
+never has to take over the live session. It lands on disk and is used at the next
+natural boundary — next episode, or next time the item is played.
+
+What exists already and is *not* this: `SmartCache` predictively downloads *other*
+items, `player_preload_upcoming` warms the next one, and
+`refresh_queue_local_sources` swaps queue entries to local at boundaries. All of
+it concerns items you are not currently playing.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Available bitrate options for an item | **Rust** | Derived from Jellyfin's media sources and playback-info negotiation; changes with the API. |
+| Mapping a chosen bitrate to transcode parameters | **Rust** | Domain vocabulary. `get_video_download_url` already owns the quality→params mapping; playback must reuse it, not restate it. |
+| Deciding whether playback bytes are keepable (direct play vs transcode) | **Rust** | Depends on the negotiated session. |
+| Canonical download quality | **Rust** | Policy over domain data. |
+| Cache eviction, storage budget, sparse-range bookkeeping | **Rust** | Storage policy. |
+| Promotion to a `downloads` row, and what invalidates a cache entry | **Rust** | Domain state. |
+| Rendering the quality selector; remembering the last choice | **Frontend** | Presentation and a view preference. The *list* comes from Rust. |
+| WiFi-only / opt-in toggles | **Frontend collects, Rust enforces** | The control is UI; the gate must hold even if the UI never calls. |
+
+Borderline, recorded: the **default** playback bitrate could look like a user
+preference (frontend). It goes to Rust because it must be reconcilable with what
+the server can actually produce for a given media source — a preference the
+backend has to validate is not a preference the frontend can own alone. The
+frontend stores the user's *choice*; Rust decides what that choice resolves to.
+
+## Design
+
+### DR-121 — Bitrate selection in the player
+
+The player exposes the qualities Rust reports for the current item. Changing it
+re-negotiates the stream URL at the new quality and resumes at the current
+position. This is a deliberate, user-initiated interruption — a brief rebuffer is
+expected and acceptable, unlike the involuntary swap the earlier design would
+have needed.
+
+Constraints that must not be broken:
+
+- On Linux, video playback must keep using the HLS `master.m3u8` URL. CLAUDE.md
+  records that returning `stream.mp4` means transcoded playback never starts.
+  A quality change re-negotiates *within* HLS.
+- The quality→transcode-parameter mapping already exists in
+  `get_video_download_url` ([online.rs:1702-1717](../../src-tauri/src/repository/online.rs#L1702-L1717)).
+  Playback must call into the same mapping. Two copies of that table will drift.
+- Track selection (audio/subtitle) already survives a stream re-negotiation
+  elsewhere in the player; a quality change must preserve it too.
+
+### DR-122 — The playback path is ephemeral
+
+Playback bytes are not persisted unless DR-124 says they are keepable. No partial
+capture is ever retained across a quality change: on change, any in-flight capture
+for that session is abandoned and its partial file deleted.
+
+### DR-123 — The download path is independent
+
+Downloading the whole file is a separate operation through the existing download
+manager, at one canonical quality (default `original`, the direct static copy),
+using `/Videos/{id}/stream.mp4` — progressive and Range-capable, which is what
+the resumable download worker relies on. It is unaffected by what playback is
+doing, and playback is unaffected by it.
+
+Once complete it becomes an ordinary download row, so everything already built on
+top of downloads — offline browsing, `refresh_queue_local_sources`, the Downloads
+page — picks it up with no further work.
+
+**Prerequisite:** downloaded *video* is currently never played locally.
+`repository_get_video_stream_url` goes straight to the online repo and
+[player/[id]/+page.svelte:316](../../src/routes/player/[id]/+page.svelte#L316)
+calls it with no local check — so a completed video download is still streamed.
+This must be fixed or the whole feature is invisible for video.
+
+### DR-124 — Keep playback bytes only when they *are* the download
+
+Capture is enabled only where the played bytes and the canonical download artifact
+are the same thing — a **direct-play** session. Then:
+
+| Path | Mechanism |
+|---|---|
+| Android / ExoPlayer | `SimpleCache` + `CacheDataSource`, keyed by item id **and** media-source id so renditions never collide. LRU evictor sharing the existing smart-cache budget — not a second budget over the same disk. |
+| Linux audio / MPV | `stream-record`, set through the existing `set_property` plumbing. |
+| Linux video (HLS transcode) | **Not captured.** Segments are not a file; assembling one needs ffmpeg, which is not a dependency and which CI is forbidden from installing at job time. The download path (DR-123) covers this case instead. |
+
+Two abandonment rules, both of which must delete the partial rather than promote
+it:
+
+- **Seek during an mpv capture.** `stream-record` is documented as intended for
+  linear streams; seeking breaks the recording. Straight-through listening
+  captures, scrubbing does not.
+- **Any quality change** (DR-122).
+
+### DR-125 — Promotion, rendition, and invalidation
+
+A capture is promoted to a `downloads` row (`status = 'completed'`) only when it
+covers the whole resource. Partial captures stay cache and remain evictable.
+
+A new `downloads.source_rendition` column records the negotiated
+quality/container/codec of whatever produced the bytes; `NULL` for rows fetched by
+the existing paths, which are always `original`. This is what makes an "upgrade to
+original" action possible later, and what stops a 720p capture and a 4K download
+being indistinguishable rows.
+
+**Invalidation.** A quality change never touches a file that already exists —
+neither a permanent download nor a completed temporary one. Both remain valid
+copies of the rendition they hold, and deleting either would throw away bytes
+already paid for.
+
+What a quality change *does* invalidate is an **in-flight** capture or background
+download of cached media: it is abandoned and restarted at the newly chosen
+quality, because a capture spanning a rendition change is a splice of two
+encodings rather than a playable file (DR-122).
+
+So the rule is about *ongoing* work, not stored files. Nothing in this spec
+deletes user data.
+
+### Gating
+
+Capture and background download obey the existing WiFi-only gate and storage
+budget, and are off unless opted in. Enforcement is in Rust.
+
+## Out of scope
+
+- **Mid-playback switch onto a completing download.** Two independent paths make
+  it unnecessary; the download is used from the next boundary.
+- **Backfilling the unplayed remainder of a capture.** Watch 40 minutes and you
+  have 40 minutes; completing it needs sparse-range bookkeeping and a resumable
+  tail fetch. The DR-123 download path already produces a complete file, which is
+  the reason this can wait.
+- **Bundling ffmpeg** to make transcoded video capturable. Real option, large
+  packaging decision, its own proposal.
+- **Routing Linux video playback through `stream.mp4`.** Regresses a documented,
+  hard-won fix.
+
+## Acceptance criteria
+
+- [ ] The player offers the qualities Rust reports, and changing one resumes at
+      the same position with audio/subtitle selection preserved.
+- [ ] A quality change abandons any in-flight capture and leaves no partial file.
+- [ ] A quality change never deletes a `downloads` row.
+- [ ] A completed background download of a video is *played from disk* on the next
+      play (the DR-123 prerequisite).
+- [ ] A direct-play session played start-to-finish leaves a complete local file
+      with no second fetch; replaying it fetches no media bytes.
+- [ ] Seeking during an mpv capture abandons it; no truncated file is promoted.
+- [ ] A transcoded Linux video session is never captured, and never partially
+      promoted.
+- [ ] Promoted rows record their rendition; existing paths still record
+      `NULL`/`original`.
+- [ ] Gates hold with the setting off *and* with the frontend never sending it.
+- [ ] Eviction cannot delete bytes backing a promoted download row.
+- [ ] `bun run check`, `bun run test`, `cargo fmt`, `cargo clippy`,
+      `bun run test:rust`, `bun run check:boundary` pass; `bindings.ts`
+      regenerated if Rust types changed.
+
+## Testing
+
+Rust, table-driven and pure where possible: quality→params resolution shared with
+the download path; keepability (direct play vs transcode vs gate off); promotion
+(complete → promoted, partial → not, seek-abandoned → not, quality-changed → not);
+invalidation (evicts cache, never a download row); rendition round-trip.
+
+Android: instrumented — a played direct-play item yields cache entries, and a
+replay issues no media network request.
+
+Frontend: the quality list renders from backend data with no item-type or
+codec taxonomy in `src/`; the selector's remembered choice is a view preference.
+
+## TRACES
+
+| Piece | Tag |
+|---|---|
+| Quality selector + re-negotiation | `// TRACES: UR-070 \| DR-121` |
+| Ephemeral playback / capture abandonment | `// TRACES: UR-070 \| DR-122` |
+| Independent whole-file download + local video playback fix | `// TRACES: UR-071 \| DR-123, IR-032` |
+| ExoPlayer cache / mpv stream-record / keepability | `// TRACES: UR-071 \| DR-124` |
+| Promotion, `source_rendition`, invalidation | `// TRACES: UR-071 \| DR-125` |
+
+## Notes for the implementer
+
+- **A parallel Claude session is active in this repo.** `git diff` before
+  "repairing" anything you did not write.
+- Do not duplicate the quality→transcode-parameter table. Call the existing one.
+- Reuse the smart-cache storage budget; two budgets over one disk is how devices
+  fill up.
+- The `downloads` FK to `items` is relaxed (migration 005) — exercise promotion
+  for an item that was never cached.
+- Build DR-123's local-playback fix first. Without it nothing in this spec is
+  observable for video.

docs/specs/req-coverage-script-removal.md

@@ -0,0 +1,161 @@
+# Spec: Remove the broken `check-req-coverage.sh`
+
+**Status:** Implemented
+**Requirements:** supports DR-093 (see [traceability-gate-repair.md](traceability-gate-repair.md))
+**UX spec:** n/a — developer tooling.
+**Supersedes / revises:** n/a
+
+## Summary
+
+`scripts/check-req-coverage.sh` is broken, orphaned, and actively misleading: it
+reports `Total Requirements: 1`, zeros in every category, and then prints
+**"✨ All requirements have implementations!"**. Nothing references it — not CI,
+not `package.json`, not the docs. This spec deletes it, with a narrowly-scoped
+alternative (repair it) documented and rejected below.
+
+## Motivation
+
+Running it today produces:
+
+```
+Category Breakdown:
+UR:    0 requirements
+IR:    0 requirements
+DR:    0 requirements
+JA:    0 requirements
+
+Summary:
+Total Requirements:       1
+✅ Fully Implemented:     0 (0%)
+
+✨ All requirements have implementations!
+```
+
+Every number is wrong (the real totals are UR 61, IR 29, DR 89, JA 32), and the
+concluding message is the *opposite* of a warning — a developer running this to
+sanity-check coverage is told everything is fine.
+
+This is worse than having no script. It is a trap, and it sits in `scripts/`
+next to tools that do work, with nothing marking it as dead.
+
+Verification that it is genuinely orphaned:
+
+```console
+$ grep -rn "check-req-coverage" . --include='*.yml' --include='*.json' \
+      --include='*.sh' --include='*.md' | grep -v node_modules
+(no output)
+```
+
+## Layer assignment
+
+Developer tooling only; no application logic and nothing crosses the IPC
+boundary.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Requirement-coverage reporting | Build tooling — `extract-traces.ts` | One tool should own coverage analysis. A second, divergent implementation is how the two answers ("1 requirement" vs "211") came to disagree unnoticed. |
+
+## Design
+
+**Delete `scripts/check-req-coverage.sh`.**
+
+Coverage reporting is owned by [scripts/extract-traces.ts](../../scripts/extract-traces.ts),
+which is correct, is what CI runs, and gains a first-class local coverage mode
+in [traceability-gate-repair.md](traceability-gate-repair.md):
+
+```bash
+bun run traces:coverage   # the supported way to check coverage locally
+```
+
+Then check the sibling scripts for the same rot. `scripts/` also contains
+`check-test-coverage.sh` and `find-req-implementations.sh`, neither of which is
+referenced from `package.json`. An unreferenced script is never run and so rots
+silently — that is the actual failure mode being fixed here, and fixing only the
+one instance found by audit leaves the others to be rediscovered later.
+
+### Findings (investigation, 2026-07)
+
+All three scripts turned out to share a **single root cause**, and all three are
+deleted:
+
+| Script | Defect |
+|---|---|
+| `check-req-coverage.sh` | Reads `README.md`, which has held **zero** requirement rows since they moved to `docs/requirements.md` → `total_reqs=1`, every category 0, "✨ All requirements have implementations!" Also greps `src-tauri/` unscoped. |
+| `check-test-coverage.sh` | Greps `src-tauri/` unscoped — including **40 GB** of `target/` build artifacts. Hangs indefinitely; produces no output at all. |
+| `find-req-implementations.sh` | Same unscoped `src-tauri/` grep. Same hang. |
+
+So none of them were subtly wrong — two could never terminate, and the third
+inverted its own conclusion.
+
+They were nonetheless *salvageable*: scoping the greps to `src-tauri/src` and
+repointing at `docs/requirements.md` would be a few lines, and the `@req:` /
+`@req-test:` tags they read are still present in the tree (**146** and **76**
+occurrences).
+
+**Decision: delete all three anyway.** The tags are an undocumented parallel
+convention — `@req:` appears in no doc, and CLAUDE.md describes only `TRACES:`.
+Repairing the scripts would re-establish a second traceability system to keep in
+sync with the first, which is the same two-sources-of-truth condition that let
+"1 requirement" and "211 requirements" coexist unnoticed. `TRACES:` plus the
+repaired coverage engine ([traceability-gate-repair.md](traceability-gate-repair.md))
+already cover this ground.
+
+The existing `@req:` / `@req-test:` comments are left in place: they are
+harmless as prose, several encode genuinely useful test intent, and stripping
+222 comments across the tree is a large diff with no functional gain. They are
+simply no longer read by any tool.
+
+### Alternative considered: repair rather than delete
+
+Rejected. The script's output format duplicates what `traces:markdown` already
+generates, it has no tests, no caller, and no documented purpose distinct from
+`extract-traces.ts`. Repairing it recreates the two-sources-of-truth condition
+that produced the contradiction. If a shell-based coverage check is ever wanted,
+it should shell out to `traces:json` and `jq` rather than re-parse
+`requirements.md` independently.
+
+## Out of scope
+
+- The CI workflow denominators — [traceability-gate-repair.md](traceability-gate-repair.md).
+- Any change to `extract-traces.ts`'s output (that spec owns it).
+- Auditing scripts that *are* referenced from `package.json` — they run
+  regularly and would fail visibly.
+
+## Acceptance criteria
+
+- [ ] `scripts/check-req-coverage.sh` no longer exists.
+- [ ] `grep -rn "check-req-coverage" .` (excluding `node_modules` and this spec)
+      returns nothing — no dangling reference in CI, docs, or `package.json`.
+- [ ] `scripts/check-test-coverage.sh` and `find-req-implementations.sh` have each
+      been run and either wired into `package.json` or deleted; the decision and
+      reason are recorded in `scripts/README.md`. **Outcome: all three deleted —
+      see Findings.**
+- [ ] `scripts/README.md` documents `bun run traces:coverage` as the supported
+      way to check requirement coverage locally.
+- [ ] `bun run test:all` passes (confirms nothing invoked the deleted script).
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `bun run check:boundary` passes.
+
+## Testing
+
+No unit tests — this is a deletion. Verification is the grep in the acceptance
+criteria plus a green `bun run test:all`, which exercises the script paths that
+actually run.
+
+## TRACES
+
+No new requirement. The deletion is covered by **DR-093**
+([traceability-gate-repair.md](traceability-gate-repair.md)), which establishes
+`extract-traces.ts` as the single owner of coverage reporting. Note the removal
+in that DR's text when both land.
+
+## Notes for the implementer
+
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes (CLAUDE.md §Gotchas).
+- Land this **after** or alongside [traceability-gate-repair.md](traceability-gate-repair.md),
+  so `bun run traces:coverage` exists before the broken script is removed and
+  developers are never left without a coverage command.
+- Check `docs/traceability-ci.md` and `docs/traces-quick-ref.md` for prose
+  references to the deleted script; the grep above covers `.md`, but read the
+  surrounding sentence rather than deleting the line mechanically.

docs/specs/scoped-search-boundary-implementation.md

@@ -0,0 +1,254 @@
+# Spec: Land the scoped-search boundary fix (implementation)
+
+**Status:** Stage 1 Implemented — Stage 2 (result-side grouping) outstanding
+**Requirements:** UR-049, UR-050 | DR-063, DR-066, DR-067 (existing — no new IDs)
+**UX spec:** n/a — zero user-visible change is the point (see Acceptance criteria).
+**Supersedes / revises:** implements [scoped-search-boundary.md](scoped-search-boundary.md),
+which specified this fix but was never built. That spec remains the **design
+authority**; this one is the delivery plan and status correction.
+
+## Summary
+
+[scoped-search-boundary.md](scoped-search-boundary.md) diagnosed a domain-taxonomy
+leak, specified the fix in full detail, and became the justification for the
+project's boundary rule in CLAUDE.md, the `check:boundary` tripwire, and the
+spec-review checklist. **The fix was never implemented.** The leak it describes
+is still live in `main`. This spec exists to close that gap and to correct the
+record — the codebase currently enforces a rule against a violation it still
+contains.
+
+## Motivation
+
+The mapping the rule forbids is present and in use:
+
+```ts
+// src/lib/utils/searchScope.ts:29-32
+const SCOPE_ITEM_TYPES: Record<Exclude<SearchScope, "all">, string[]> = {
+  music: ["MusicAlbum", "MusicArtist", "Audio", "Playlist"],
+  movies: ["Movie"],
+  tv: ["Series", "Episode"],
+};
+```
+
+This is not dead code. [library.ts:262](../../src/lib/stores/library.ts#L262)
+calls `scopeItemTypes(scope)` and puts the result straight into
+`options.includeItemTypes`. Meanwhile there is **no `SearchScope` anywhere in
+`src-tauri/`**:
+
+```console
+$ grep -rn "SearchScope" src-tauri/src --include='*.rs'
+(no output)
+```
+
+Three things make this the highest-value item found in the design-principles
+audit:
+
+1. **The rule's own founding incident is unremediated.** CLAUDE.md cites this
+   spec as "the incident this rule came from." A rule whose originating
+   violation is still shipping is not credible.
+2. **The tripwire cannot see it.** `bun run check:boundary` passes — it greps for
+   a multi-type array literal *at the query site*, and this one is assigned to a
+   named const and dereferenced elsewhere. Broadening the tripwire is specified
+   separately in [boundary-tripwire-hardening.md](boundary-tripwire-hardening.md);
+   note that hardening it **without** landing this fix would turn `master` red.
+3. **The spec's own acceptance criterion fails today.** "Adding a hypothetical
+   new type to a scope requires editing only Rust" — adding a type to the Music
+   scope right now requires editing `searchScope.ts`.
+
+## Layer assignment
+
+Unchanged from [scoped-search-boundary.md](scoped-search-boundary.md) §Design;
+restated so this spec is reviewable on its own.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Scope → Jellyfin item types (`music` → `MusicAlbum`, `MusicArtist`, `Audio`, `Playlist`) | **Rust** | Domain vocabulary. Changes if Jellyfin adds/renames an item type — the litmus test's "yes" case. This is the leak being fixed. |
+| Result item → search group bucketing | **Rust** | Same taxonomy, result side. Classifying a `MediaItem` as a Song vs Album is Jellyfin vocabulary, not layout. |
+| `All` sends no filter at all (≠ union of enumerated types) | **Rust** | A query-shaping rule with a correctness consequence (Person/folder results would be silently dropped). Belongs with the expansion it qualifies. |
+| Group display order, labels, reordering, persistence | Frontend | Pure presentation — changes only if the UI is redesigned. Explicitly retained frontend-side. |
+| `resolveSearchScope(pathname)` — route → initial scope | Frontend | Routing/navigation, no Jellyfin vocabulary. Stays exactly as-is. |
+| Chip labels (`SCOPE_LABELS`), scope order (`SEARCH_SCOPES`) | Frontend | Display strings over an opaque enum. |
+| `GROUP_SCOPE` (which group belongs to which scope) | **Delete** | Borderline taxonomy, made redundant: once Rust filters by scope, out-of-scope groups arrive empty and drop via the empty-omit rule. Borderline defaults to Rust; here it defaults to *gone*. |
+
+The `SearchScope` and `SearchGroupId` **types** come to the frontend from
+generated `bindings.ts`. Naming an opaque enum variant is not taxonomy; knowing
+what item types it expands to is.
+
+## Design
+
+**Follow [scoped-search-boundary.md](scoped-search-boundary.md) §Design as
+written** — `SearchScope` enum + `item_types()` in `repository/types.rs`,
+`SearchOptions.scope`, `SearchGroupId`/`SearchGroup`/`GroupedSearchResult`,
+scope-wins precedence, `All` → `None` → no filter. It is not restated here;
+duplicating it would create two drifting copies of the same design.
+
+This spec adds only the delivery sequencing that the original left implicit.
+
+### Staging: land it in two reviewable pieces
+
+The original bundles the query side and the result side into one change. That is
+a large diff touching Rust types, `bindings.ts`, the store, and a component, with
+the `search-event` dual-payload hazard in the middle. Split it:
+
+**Stage 1 — query side (closes the leak).**
+`SearchScope` enum, `SearchOptions.scope`, command resolves scope →
+`include_item_types` in Rust, `library.ts` sends `{ scope }`, delete
+`SCOPE_ITEM_TYPES` and `scopeItemTypes()`. Result grouping stays as it is.
+
+After Stage 1 the actual boundary violation is gone and
+[boundary-tripwire-hardening.md](boundary-tripwire-hardening.md) can land safely.
+
+**Stage 2 — result side.** `SearchGroupId`/`SearchGroup`/`GroupedSearchResult`,
+Rust bucketing, both payloads converted, `composeSearchGroups()` shrunk,
+`GROUP_ITEM_TYPES`/`groupItemTypes()`/`GROUP_SCOPE` deleted.
+
+Both stages are required for the original spec's acceptance criteria to pass;
+Stage 1 alone leaves `GROUP_ITEM_TYPES` in the frontend. **Stage 1 is not a
+stopping point** — it is a review boundary. Do not mark the parent spec
+Implemented until Stage 2 lands.
+
+### Stage 1 — delivered (July 2026)
+
+- `SearchScope` enum + `item_types()` in [repository/types.rs](../../src-tauri/src/repository/types.rs);
+  `All` → `None` → no filter.
+- `SearchOptions.scope` with `resolve_scope()`; scope wins over
+  `include_item_types`, which stays for the non-search `get_items` callers.
+- `repository_search` resolves the scope **once, before** the cache/server split,
+  so both phases filter identically.
+- `SCOPE_ITEM_TYPES` and `scopeItemTypes()` deleted; `searchScope.ts` now
+  re-exports `SearchScope` from the generated bindings instead of a hand-written
+  union.
+- [library.ts](../../src/lib/stores/library.ts) sends `{ scope }`.
+- 8 Rust tests (`search_scope_tests`); the frontend suite now asserts the
+  *opaque scope* is sent rather than an item-type list.
+
+Verified: adding `"AudioBook"` to the Music scope changed **zero** files under
+`src/` — the criterion that failed before this work.
+
+**Stage 2 remains open**: `GROUP_ITEM_TYPES` / `groupItemTypes()` (result-side
+bucketing, single-type-per-group) are still in `searchScope.ts`, and both search
+payloads still carry a flat `MediaItem[]` rather than `GroupedSearchResult`.
+
+### 🔴 The `search-event` dual payload (Stage 2)
+
+The original flags this as "the single largest part of the change and the
+easiest to half-do." Restating because it is the one thing that silently breaks:
+search resolves **twice** — the command returns instant cache results, then the
+merged cache+server union arrives via `search-event`. Both payloads must carry
+`GroupedSearchResult`. Convert one and the UI flickers between shapes as server
+results land.
+
+Write the failing test for the *event* payload first — the command return is the
+obvious half, the event is the half that gets forgotten.
+
+### Note on `SearchOptions.scope` and specta
+
+`SearchOptions` is already `#[serde(rename_all = "camelCase")]` with
+`skip_serializing_if = "Option::is_none"`. Add `scope: Option<SearchScope>`
+following that pattern so `All`/absent omits the key. Regenerate `bindings.ts`
+— `SearchOptions` there is currently
+`{ limit?, includeItemTypes?, searchTerm? }` and must gain `scope?`. Never
+hand-edit it.
+
+## Out of scope
+
+- Redesigning anything in [scoped-search-boundary.md](scoped-search-boundary.md).
+  If implementation shows the design wrong, revise **that** spec, don't fork it.
+- Online/offline `include_item_types` **filtering** — already correct; only the
+  source of the type list moves.
+- Ranking within or across groups (DR-090 territory).
+- Chip UX, scope persistence, group-order persistence — unchanged.
+- The two lesser type-set sites in `DownloadedBrowse.svelte` and
+  `GenericMediaListPage.svelte`, handled in
+  [boundary-tripwire-hardening.md](boundary-tripwire-hardening.md).
+- Broadening the tripwire itself — same sibling spec.
+
+## Acceptance criteria
+
+Inherits every criterion from [scoped-search-boundary.md](scoped-search-boundary.md)
+§Acceptance criteria. Additionally:
+
+- [ ] `grep -rn "SearchScope" src-tauri/src --include='*.rs'` returns matches —
+      the enum exists in Rust (it does not today).
+- [ ] `grep -n "SCOPE_ITEM_TYPES\|scopeItemTypes\|GROUP_ITEM_TYPES\|groupItemTypes" src/lib/utils/searchScope.ts`
+      returns nothing.
+- [ ] `grep -rn "scopeItemTypes" src/` returns nothing — including the
+      `library.ts` import and call site.
+- [ ] `SearchOptions` in `bindings.ts` includes `scope`; regenerated, not
+      hand-edited.
+- [ ] **Behaviour is byte-identical for the user**: same scoping, same groups,
+      same order, same empty-group omission, offline included. This spec is a
+      pure refactor — any visible change is a defect.
+- [ ] `All` scope sends no `includeItemTypes` (asserted in a Rust test, not by
+      inspection).
+- [ ] Adding a type to the Music scope requires editing **only** Rust —
+      demonstrate by making the edit and confirming no `src/` file changes.
+- [ ] `scoped-search-boundary.md` status flips to **Implemented**, and
+      `scoped-search.md`'s "frontend only, no Rust changes" framing gets a
+      banner pointing at the corrected design.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] `bun run check:boundary` passes.
+- [ ] Changed code carries `// TRACES:` comments (IDs below).
+
+## Testing
+
+Follow [scoped-search-boundary.md](scoped-search-boundary.md) §Testing. Emphases:
+
+**Rust** (`cargo test`):
+- `SearchScope::item_types()` per scope; `All` → `None`.
+- Scope resolution happens **before** the online/offline split, so both paths
+  get the same filter — a regression here is invisible until someone searches
+  offline.
+- `scope` set + `include_item_types` set → scope wins (the documented
+  precedence; assert it rather than trusting the doc).
+- Stage 2: mixed `Vec<MediaItem>` buckets correctly; unknown types dropped;
+  canonical group order; **the `search-event` payload is the grouped shape**.
+
+**Frontend** (`bun run test`):
+- `resolveSearchScope()` tests in `searchScope.test.ts` must pass **unchanged** —
+  they cover the part that is not moving, and are the regression net proving the
+  refactor didn't disturb routing.
+- `library.ts` sends `{ scope }` and never `includeItemTypes` for search.
+- `composeSearchGroups()` over fixture `SearchGroup[]` with no `.type`
+  inspection in the implementation.
+
+**Offline parity:** run a scoped search with the server unreachable and confirm
+identical grouping. The offline repository path honours `include_item_types`
+independently, and this is the case most likely to be missed.
+
+## TRACES
+
+No new requirement IDs — this implements existing ones. Retag as the code moves:
+
+```rust
+// src-tauri/src/repository/types.rs
+/// TRACES: UR-049 | DR-063
+pub enum SearchScope { … }
+```
+
+```typescript
+// src/lib/utils/searchScope.ts — keep the file header; it retains
+// resolveSearchScope + group-order presentation logic.
+// TRACES: UR-049, UR-050 | DR-063, DR-066, DR-067
+```
+
+Update DR-063's text in `requirements.md` to state that scope expansion is owned
+by Rust, so the requirement stops describing the leaked design. New Rust tests
+take `@req-test: UT-089` onward (next free UT is **UT-089**).
+
+## Notes for the implementer
+
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes (CLAUDE.md §Gotchas).
+- **Read [scoped-search-boundary.md](scoped-search-boundary.md) first.** This
+  spec is deliberately thin on design; that one is the authority.
+- Sequence with the sibling specs: **Stage 1 here → then
+  [boundary-tripwire-hardening.md](boundary-tripwire-hardening.md)**. Hardening
+  the tripwire first turns `master` red on a known-unfixed violation.
+- `git log --oneline -- docs/specs/scoped-search-boundary.md` is worth a look
+  before starting — understanding why the fix stalled may surface a constraint
+  the spec didn't record.
+- The user-visible-change count for this spec is zero. If QA reports a
+  difference in search results, that is a bug in the refactor, not an
+  improvement.

docs/specs/scoped-search.md

@@ -8,8 +8,14 @@
 > is being moved into Rust. The **user-facing behaviour and UX in this spec are
 > unchanged**; only where the scope→item-type mapping and result bucketing live
 > changes. Read the boundary spec before touching search code.
+>
+> **Progress:** the scope→item-type mapping now lives in Rust
+> (`SearchScope::item_types()`); the frontend sends an opaque scope. Result-side
+> bucketing (`GROUP_ITEM_TYPES`) is still frontend-side — see
+> [scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md)
+> §Stage 2.
 
-**Status:** Implemented (boundary revision pending — see banner above)
+**Status:** Implemented (boundary revision: query side done, result side pending)
 **Scope:** Frontend only. No Rust changes required. *(Revised — see banner.)*
 **Requirements:** UR-049 → DR-063, DR-064, DR-065; UR-050 → DR-066, DR-067
 (see [requirements.md](../requirements.md)).

docs/specs/series-current-episode-navigation.md

@@ -0,0 +1,239 @@
+# Spec: series navigation lands on the current episode
+
+**Status:** Accepted
+**Requirements:** UR-062 → DR-101, DR-102, DR-103, DR-104, DR-107; UR-063 → DR-105; UR-064 → DR-106
+**UX spec:** [ux-flows.md §5B.1](../ux-flows.md), [§5B.2](../ux-flows.md), [§5B.4](../ux-flows.md), [§5B.5](../ux-flows.md)
+
+## Summary
+
+Opening a TV series lands you where you actually are in it: the seasons render
+as collapsible sections with **only the current season expanded**, the current
+episode highlighted and scrolled into view, and the hero button opens that
+episode's focus view (labelled `Resume S2E4` / `Play S1E1`) instead of the first
+season. A season stops being a destination of its own — every route that used to
+land on `/library/<seasonId>` now lands on the series with that season in view,
+so the full cross-season episode list is always reachable in one place. Watch
+history can be erased per series and per season. Separately, each video library
+collapses from three routes (landing, all-titles, genres) to one route with
+in-page tabs.
+
+## Motivation
+
+Two problems, reported together.
+
+**1. Series navigation dead-ends at season 1.** The series detail page's Play
+button resolved its target as `$libraryItems[0]` — the first *season* child,
+ordered by `SortName` — and navigated to `/player/<seasonId>`. The player route
+classifies `season` as a container kind and bounces it back to
+`/library/<seasonId>`. So Play on a series played nothing; it navigated you to
+the season-1 page. Opening a series without pressing Play rendered every season
+stacked but scrolled to the top, so a viewer 4 seasons deep had to scroll past
+everything they had already watched.
+
+The backend has been able to answer "where is this viewer in this show" the
+whole time: `repository_get_next_up_episodes(handle, series_id, limit)` is wired
+end-to-end to `/Shows/NextUp?SeriesId=`. **Both frontend call sites pass
+`undefined` for `series_id`** — the per-series capability existed and was never
+used.
+
+**2. Seasons are an accidental page.** There is no season route. `/library/
+<seasonId>` falls through the detail page's `kind` chain into the generic
+"Contents" poster grid, which contradicts ux-flows §5A.2 (episodes in a season
+must render as a row list). Worse, clicking an episode from that grid opens a
+*bare* Episode page, which §5B.1 explicitly forbids. Four call sites fed it: the
+episode breadcrumb, `handleItemClick case "season"`, the TV landing page, and
+the broken Play button above.
+
+**3. Too many video library routes.** Seven routes serve two media types, and the
+naming does not even agree with itself: `/library/tv` + `/library/tv/shows` +
+`/library/shows/genres` versus `/library/movies` + `/library/movies/all` +
+`/library/movies/genres`. The genre routes do not share a prefix, which
+`searchScope.ts:45` carries an apologetic comment about. The two "all" pages are
+27-line config wrappers over the same `GenericMediaListPage`.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Which episode is "current" for a series (resume → next-up → first unwatched → first) | **Rust** | Domain policy over Jellyfin user-data semantics. It changes if Jellyfin changes what `UserData.is_played` means, if Next Up's rules change, or if we decide a 98%-watched episode counts as finished. It does not change if the UI is redesigned. |
+| Gathering a series' episodes across all seasons in broadcast order | **Rust** | Jellyfin's shape (episodes hang off season folders, except when a series is flat and they hang off the series) is provider vocabulary. The frontend already reimplemented this fan-out *and* its flat-series fallback; that is domain knowledge that leaked. |
+| Ordering rule for "series order" (season index, then episode index, specials last) | **Rust** | Season 0 = specials is a Jellyfin convention, not a layout choice. |
+| Scrolling the current episode into view; the highlight ring and `Up next` badge | Frontend | Pure presentation. Changes only if the page is redesigned. |
+| Which seasons start expanded | Frontend | Consumes the backend's answer (`currentEpisode`) to decide layout. The *decision* about where the viewer is stays in Rust; only "and therefore this section opens" is here. |
+| What "erase watch history" means (played flag + resume position, recursive over a container) | **Rust** | Jellyfin user-data semantics. Changes if the server's mark-unplayed behaviour changes; unaffected by any UI redesign. |
+| Refusing to clear history while offline | **Rust** | A data-integrity rule, not a disabled button: history cleared only locally would be undone by the next sync. The UI disabling the button is a courtesy on top. |
+| Play button *label* (`Resume S2E4` vs `Play S1E1`) | Frontend | Rendering a decision the backend already made (the returned episode plus its resume position). |
+| Which route Play navigates to | Frontend | Navigation is presentation. |
+| Redirecting `/library/<seasonId>` to the series anchor | Frontend | Route topology. |
+| Episode-strip window size (3 before / 6 after) | Frontend | A layout constant; §5B.2 owns it. |
+| Library page tabs and the `?view=` param | Frontend | View preference and route topology. |
+
+Borderline row — **the strip's cross-season *ordering*** is Rust (it is series
+order, above), but the *window* taken from that ordered list is frontend. The
+tie-breaker: the list handed to the frontend is already correct and complete;
+choosing how much of it fits on screen is layout.
+
+## Design
+
+### Rust: the current-episode policy
+
+Two new pieces, split so the policy is unit-testable without a repository.
+
+**Pure policy** — `src-tauri/src/repository/series_progress.rs`:
+
+```rust
+/// Series order: season index asc, then episode index asc. Specials (season 0)
+/// sort after every numbered season rather than before season 1.
+pub fn sort_series_order(episodes: &mut [MediaItem]);
+
+/// The episode a viewer should land on, given everything already fetched.
+/// Order: in-progress episode → Next Up → first unwatched → first episode.
+pub fn pick_current_episode(
+    episodes: &[MediaItem],   // series order
+    next_up: &[MediaItem],
+    resume: &[MediaItem],
+) -> Option<MediaItem>;
+```
+
+Why that order:
+
+- **In-progress wins** because a partially-watched episode is literally where
+  the viewer stopped; Next Up would skip past it. Ties break toward the earliest
+  in series order, so a viewer who dipped into a later episode still resumes the
+  one they are actually working through.
+- **Next Up second** because it is the server's own answer, and it accounts for
+  history we do not cache.
+- **First unwatched third** — the offline repository returns an empty vec for
+  Next Up (`offline.rs:1247`), so without this fallback the whole feature would
+  be online-only. This is the offline path, not dead code.
+- **First episode last** so a never-watched series lands on S1E1 rather than
+  nothing.
+
+A `resume`/`next_up` entry that is not among `episodes` is still honoured — it
+comes from the same server and may carry an id the season fan-out missed — but
+it must belong to this series.
+
+**Fetch + command** — `src-tauri/src/commands/repository.rs`:
+
+```rust
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_get_series_episodes(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    series_id: String,
+) -> Result<Vec<MediaItem>, String>
+
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_get_series_current_episode(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    series_id: String,
+) -> Result<Option<MediaItem>, String>
+```
+
+Frontend params are camelCase (`{ handle, seriesId }`) per the Tauri v2 rule.
+
+`repository_get_series_episodes` performs the fan-out the frontend used to do:
+`get_items(series_id)` → seasons → `get_items(season_id)` per season, plus the
+flat-series fallback (a series whose children are episodes, not seasons), then
+`sort_series_order`. `repository_get_series_current_episode` calls it, adds
+`get_next_up_episodes(Some(series_id), Some(1))` and
+`get_resume_items(Some(series_id), Some(10))`, and applies `pick_current_episode`.
+Both tolerate a failing Next Up (offline) by treating it as empty rather than
+failing the whole call.
+
+### Frontend: series page
+
+- `loadItem()` calls `repositoryGetSeriesEpisodes` once instead of fanning out
+  over seasons itself, and `repositoryGetSeriesCurrentEpisode` for the anchor.
+  Season *headers* still come from `get_items(seriesId)`; the page groups the
+  returned episodes under them by `parentIndexNumber`.
+- No `?episode=` param → series view, `SeasonSection` receives
+  `currentEpisodeId`, `EpisodeRow` renders the highlight and scrolls itself into
+  view (`scrollIntoView({ block: "center" })`, the existing `focused` mechanism,
+  now distinguishing *focused* from *current*).
+- Seasons are collapsible and **only the current season is expanded**
+  (`initialExpandedSeasons`). Without this a ten-season show renders every
+  episode of every season at once and buries the one the viewer came for. A
+  collapsed season still shows its episode count and watched count, so progress
+  is legible without expanding. Toggle state is local and not persisted — it is
+  a reading position, not a preference.
+- Hero Play → `goto(/library/<seriesId>?episode=<currentId>)`, i.e. the Episode
+  Focus View, where an explicit Play/Resume starts playback. This follows
+  ux-flows §5B.5's "tap opens, never commits" rule: Play on a *container* is
+  navigation; Play on a *leaf* (the focus view, a movie) commits.
+- Clicking an episode in a season section → `?episode=` swap, not
+  `/player/<id>`. §5B.1.
+
+### Frontend: seasons are not a destination
+
+`/library/<seasonId>` resolves the season's `seriesId` and redirects to
+`/library/<seriesId>#season-<indexNumber>`; `SeasonSection` renders that anchor
+id. A season with no `seriesId` (deep link into a stale cache) keeps the old
+generic rendering as a fallback so the user is never stranded. Inbound links
+updated: episode breadcrumb, `handleItemClick case "season"`, the TV landing
+page's `case "Season"`, and `DownloadedBrowse`.
+
+### Erasing watch history
+
+```rust
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_clear_watch_history(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    item_id: String,
+) -> Result<(), String>
+```
+
+`OnlineRepository` maps it to `DELETE /Users/{userId}/PlayedItems/{itemId}` —
+Jellyfin's mark-unplayed, which clears the played flag *and* zeroes the resume
+position, and which the server applies recursively to a folder. One call
+therefore handles a whole series or a single season; no per-episode fan-out.
+`OfflineRepository` returns `RepoError::Offline` rather than clearing locally,
+because divergent local history is undone by the next sync.
+
+`ClearHistoryButton` is shared by the series hero (`scope="series"`) and each
+`SeasonSection` header (`scope="season"`). It confirms first — there is no undo —
+disables itself while the server is unreachable, and reloads the page on success
+so the recomputed current episode is what the viewer sees. Clearing a whole
+series therefore returns it to S1E1, which is the same path a never-watched
+series takes through `pick_current_episode`.
+
+### Frontend: one route per video library
+
+`/library/tv` and `/library/movies` each gain `?view=browse|all|genres` tabs,
+rendering the existing `GenericMediaListPage` / `GenericGenreBrowser` components
+inline. `?view=` is omitted for `browse` (the default) to keep URLs clean —
+the same convention `searchRouteUrl` uses for the `all` scope.
+
+The four legacy routes become redirect-only `+page.ts` loads:
+
+| Legacy | Redirects to |
+|--------|--------------|
+| `/library/tv/shows` | `/library/tv?view=all` |
+| `/library/shows/genres` | `/library/tv?view=genres` |
+| `/library/movies/all` | `/library/movies?view=all` |
+| `/library/movies/genres` | `/library/movies?view=genres` |
+
+They are kept (rather than deleted) because `GenreTags` builds links to them and
+users may have them in history. `resolveSearchScope` keeps its `/library/shows`
+branch for the same reason.
+
+The "Browse" tile grid at the bottom of both landing pages is removed — the tabs
+replace it, and the tiles were a second navigation affordance to the same two
+destinations the carousels' "Show all" links already reach.
+
+## Out of scope
+
+- **Cross-season autoplay.** `player/mod.rs:fetch_next_episode_for_item` is
+  still season-bounded, so autoplay stops at a season boundary. Fixing it should
+  reuse `repository_get_series_episodes`, but it touches the playback state
+  machine and the Android JNI advance path (see the `AutoplayDecision` deadlock
+  note in CLAUDE.md) and belongs in its own change.
+- **Music library routes.** `/library/music/*` has five sub-routes with the same
+  shape; the same consolidation applies but is not done here.
+- **Marking a series' progress** (mark-watched / mark-unwatched from the series
+  page).

docs/specs/streaming-bitrate-cap.md

@@ -0,0 +1,146 @@
+# Spec: streaming bitrate cap
+
+**Status:** Implemented
+**Requirements:** UR-074 → DR-162 (partially serves UR-070)
+**UX spec:** n/a — the controls reuse existing patterns (Settings → Video Playback, and the player's track menus).
+
+## Summary
+
+The viewer picks a bandwidth ceiling for video — from `Original` (no client
+limit) down to 720 kbps — and every video the app opens is fetched within it,
+live TV included. The choice is made once in Settings and persists across
+restarts; a single video can be moved to another ceiling from the player, which
+re-opens the stream and resumes where it was without changing the saved default.
+
+## Motivation
+
+Every video URL the app built carried a fixed allowance —
+`MaxStreamingBitrate=20000000`, `VideoBitrate=18000000` — the `PlaybackInfo`
+negotiation asked for 20 Mbps, and the device profile advertised
+`999999999`, which invites the server to direct-play a source of any size. On a
+metered or slow connection there was no lever at all short of not watching.
+
+The related UR-070 asks for something adjacent but different: a list of the
+renditions *the server can produce for this item*. That needs per-item
+`MediaSources` negotiation and is still proposed. What was missing first is
+cruder and more valuable: a device-wide budget that holds regardless of what is
+playing.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| What a quality step *is* — total ceiling, audio share, resolution cap | Rust | Jellyfin encoding vocabulary. It changes if Jellyfin's transcoder or parameter binding changes, not if the UI is redesigned. Exactly the shape of `EqPreset::gains()`. |
+| Splitting the ceiling between video and audio | Rust | A domain rule about what the server is being asked to produce; getting it wrong overshoots the user's cap. |
+| Choosing `MaxHeight` for a bitrate | Rust | An encoding judgement (how many pixels a budget can carry), not a display preference. |
+| Where the cap is applied (URL builders, `PlaybackInfo`, live TV, audio handoff) | Rust | All four are backend concerns, and the frontend must not have to know that a cap has more than one enforcement point. |
+| Whether a mid-playback change needs a stream reload, and performing it | Rust | Same decision the audio-track switch already delegates: the backend knows the playback mode and owns the queue. |
+| Persisting the default | Rust | Application state in `app_settings`, alongside every other durable setting. |
+| Rendering the picker, menu placement, which control is highlighted | Frontend | Pure presentation. |
+
+The frontend holds one string — the serde token for the chosen variant — and
+labels/details it received from Rust. It never encodes a bitrate, a resolution
+or a parameter name.
+
+## Design
+
+`StreamingQuality` (`src-tauri/src/settings.rs`) is the ladder: `Original`,
+`Mbps20`, `Mbps10`, `Mbps8`, `Mbps4`, `Mbps2`, `Mbps1`, `Kbps720`, serialised
+camelCase (`"mbps10"`). Each step answers `max_bitrate()`, `audio_bitrate()`,
+`video_bitrate()` (= total − audio), `max_height()`, `label()`, `detail()`.
+
+The active ceiling is a process-wide `RwLock<StreamingQuality>` in
+`repository/online.rs`, read by every builder. Process-wide rather than a field
+on `OnlineRepository` because it is a preference about *this device's
+connection*: it must survive a repository rebuilt on re-login, and the URL
+builders and the negotiation have to agree on it or the cap leaks. This mirrors
+`offline::INCLUDE_CATALOG_BROWSE`.
+
+Enforcement points — all four are required:
+
+| Point | What the cap sets |
+|-------|-------------------|
+| `get_video_stream_url` (HLS transcode) | `MaxStreamingBitrate`, `VideoBitrate`, `AudioBitrate`, `MaxHeight` |
+| `get_playback_info` | request `MaxStreamingBitrate`, and the device profile's `MaxStreamingBitrate`/`MaxStaticBitrate` |
+| `open_live_stream` | `MaxStreamingBitrate` |
+| `build_audio_only_stream_url_for_video` | `min(cap audio, 384 kbps)` |
+
+The negotiation is the one that matters most. `MaxStaticBitrate` is what makes
+the server refuse to *direct play* a source fatter than the ceiling; without it
+a 30 Mbps remux is served untouched and no URL parameter downstream can reduce
+it.
+
+IPC:
+
+```rust
+player_get_streaming_qualities() -> Vec<(StreamingQuality, String, String)>  // variant, label, detail
+player_set_stream_quality(repository_handle, quality, use_html5,
+                          current_position, media_source_id, audio_stream_index)
+    -> StreamQualityResponse           // #[serde(tag = "strategy")]: native | reloadStream
+```
+
+`VideoSettings` gains `streaming_quality` (`#[serde(default)]`, so settings
+persisted before the field existed load as uncapped).
+`player_set_video_settings` applies it and writes it to `app_settings`;
+`restore_streaming_quality` reads it back in the Tauri `setup` hook via
+`tauri::async_runtime::spawn`, defaulting to uncapped if anything fails.
+
+`StreamQualityResponse` keeps its Rust field names on the wire (`new_url`) —
+tauri-specta only camelCases the `strategy` tag. The facade
+(`playerController.setStreamQuality`) dispatches `reloadSource` for
+`reloadStream` and does nothing for `native`, because the backend has already
+reloaded itself.
+
+Mid-playback the change applies to the current video **and** becomes the process
+ceiling for what follows, but it is not persisted: the in-player menu is a "this
+film, this connection" control and Settings owns the durable default.
+
+## Out of scope
+
+- Per-item rendition lists from the server's `MediaSources` (UR-070's other half).
+- Connection-aware caps (separate WiFi/cellular ceilings). One cap, all connections.
+- Adaptive/automatic selection from measured throughput.
+- Download quality, which already has its own preset vocabulary (UR-071/DR-123).
+
+## Acceptance criteria
+
+- [x] `bun run check` passes.
+- [x] `cargo fmt` clean, `cargo clippy` clean, Rust tests pass.
+- [x] `bun run test` passes.
+- [x] `bun run check:boundary` passes — no bitrate/resolution numbers in `src/`.
+- [x] New code carries `// TRACES:` comments.
+- [x] `bindings.ts` regenerated from Rust.
+- [x] A capped step changes what the URL asks for; the uncapped default is byte-identical to the previous behaviour.
+
+## Testing
+
+Rust (`cargo test`):
+
+- `test_video_stream_url_applies_bitrate_cap` — all four parameters at `Mbps2`.
+- `test_video_stream_url_uncapped_keeps_legacy_allowance` — `Original` is unchanged and adds no `MaxHeight`.
+- `test_audio_only_stream_url_takes_the_lower_of_cap_and_default`.
+- `test_streaming_quality_budget_is_internally_consistent`, `..._ladder_descends`, `..._round_trips_through_json`.
+
+The ceiling is process-wide, so tests that depend on it serialise on a guard
+(`QualityFixture`) that restores `Original` on drop — including the two
+pre-existing stream-URL tests, which would otherwise see another test's cap.
+
+`get_playback_info` and `open_live_stream` need a live server and are not unit
+tested; their behaviour is the enum's `max_bitrate()`, which is.
+
+## TRACES
+
+- `StreamingQuality`, `VideoSettings.streaming_quality` — `UR-074 | DR-162`
+- URL builders / negotiation / live TV — `UR-004, UR-074 | DR-140, DR-162`
+- Audio-only handoff — `UR-040, UR-074 | DR-162`
+- Commands, facade, Settings UI, player menu — `UR-074 | DR-162`
+- Tests — `UT-156`, `UT-157`
+
+## Notes for the implementer
+
+- `videoBitRate` with a capital R is the *download* endpoint's binding quirk
+  (DR-123). The streaming endpoint used here binds `VideoBitrate`/
+  `MaxStreamingBitrate` as spelled above — do not "correct" one to the other.
+- A parallel Claude session may be active in this repo; `git diff` before
+  repairing unexpected changes. DR-160/161 were claimed by such a session while
+  this feature was in flight, which is why it is DR-162.

docs/specs/traceability-gate-repair.md

@@ -0,0 +1,238 @@
+# Spec: Repair the traceability coverage gate
+
+**Status:** Implemented
+**Requirements:** DR-093 → supports the traceability practice described in CLAUDE.md
+**UX spec:** n/a — developer tooling, no user-facing surface.
+**Supersedes / revises:** n/a
+
+## Summary
+
+The CI traceability gate has been passing unconditionally for an unknown length
+of time because it divides traced-requirement counts by **hardcoded denominators
+that no longer match [requirements.md](../requirements.md)**. It currently
+reports **158% overall coverage** (and `JA 24 / 3 = 800%`), so the 50% threshold
+is mathematically unreachable and the job cannot fail. This spec makes the gate
+derive its denominators from `requirements.md` at run time, so it reports the
+real number (**85%** today) and can actually fail again.
+
+## Motivation
+
+`.gitea/workflows/traceability-check.yml` hardcodes `UR/39, IR/24, DR/48, JA/3`
+and `TOTAL_REQS=114`. The real counts are **UR 61, IR 29, DR 89, JA 32 — 211
+total**. Requirements were added over time; the divisors were never updated.
+
+The consequence is not a cosmetic reporting bug. The gate is the *only*
+automated defence for the traceability practice, and it is dead:
+
+```
+CI today:  181 / 114 = 158%  → threshold 50% can never trip
+Reality:   181 / 211 =  85%  → healthy, but unguarded
+```
+
+Coverage could collapse to 30% and CI would still print a green
+"✅ Coverage is acceptable". An audit of the design principles found that every
+principle with a *working* automated check is in good shape, and the ones that
+drifted are exactly the ones whose checks were broken or too narrow — this is
+the clearest instance.
+
+A second, related defect is handled in a sibling spec: `scripts/check-req-coverage.sh`
+is separately broken and orphaned (see
+[req-coverage-script-removal.md](req-coverage-script-removal.md)).
+
+## Layer assignment
+
+This spec touches only CI/build tooling — no application logic crosses the
+Rust/Svelte boundary. The table is filled in for completeness.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Counting requirement IDs defined in `requirements.md` | Build tooling (`scripts/`) | Neither runtime layer; it is repo metadata analysis. Belongs beside `extract-traces.ts`, not in the workflow YAML, so it is runnable and testable locally. |
+| Counting *traced* requirement IDs | Build tooling — existing `extract-traces.ts` | Already implemented and correct; this spec consumes it rather than duplicating it. |
+| Threshold policy (the 50% number) | CI workflow | Deployment policy, not analysis. Keeping it in YAML lets it be tuned without touching the script. |
+
+No frontend or Rust logic is added, so no taxonomy leak is possible.
+
+## Design
+
+### 1. Denominators come from `requirements.md`, not literals
+
+`requirements.md` defines requirements in markdown tables with a stable leading
+cell, e.g.:
+
+```
+| DR-001 | Player state machine (idle, loading, …) | Player | UR-005 | Done |
+| UR-002 | Access media when online or offline | High | Done |
+```
+
+Extend [scripts/extract-traces.ts](../../scripts/extract-traces.ts) to also emit
+the *defined* counts, so one tool owns both sides of the fraction and CI does no
+arithmetic on stale literals. Add a `defined` key to the JSON report:
+
+```jsonc
+{
+  "byType":   { "UR": [...], "IR": [...], "DR": [...], "JA": [...] },  // traced (existing)
+  "defined":  { "UR": 61, "IR": 29, "DR": 89, "JA": 32 },              // NEW
+  "coverage": { "covered": 181, "total": 211, "percent": 85 },         // NEW
+  "requirements": { ... },                                             // existing
+  "totalTraces": 318, "totalFiles": …, "timestamp": "…"                // existing
+}
+```
+
+Parsing rule for a *defined* requirement: a line in `docs/requirements.md`
+matching `^\|\s*(UR|IR|DR|JA)-\d{3}\s*\|` — the ID must be the table's first
+cell. This deliberately does **not** count IDs mentioned in the `Traces To`
+column or in prose, which is why a naive `grep -o` over the whole file
+overcounts.
+
+`defined` counts IDs that exist in the spec; `byType` counts IDs that appear in
+a `TRACES:` comment somewhere in the source. Coverage is
+`|byType ∩ defined| / |defined|`.
+
+> **Intersection, not raw length.** A `TRACES:` comment naming an ID that
+> `requirements.md` does not define (a typo, or a requirement later deleted)
+> must **not** inflate the numerator — that is how a ratio exceeds 100% in the
+> first place. Such IDs are reported separately as `orphaned` so they get fixed
+> rather than silently counted or silently dropped.
+
+```jsonc
+"orphaned": ["DR-097"]   // traced in code but not defined in requirements.md
+```
+
+### 2. The workflow consumes the computed number
+
+Replace the arithmetic in `.gitea/workflows/traceability-check.yml` (lines
+46–76) with reads of the precomputed fields:
+
+```sh
+COVERAGE=$(jq '.coverage.percent' traces-report.json)
+COVERED=$(jq '.coverage.covered' traces-report.json)
+TOTAL_REQS=$(jq '.coverage.total'   traces-report.json)
+
+for T in UR IR DR JA; do
+  TRACED=$(jq --arg t "$T" '.byType[$t] | length' traces-report.json)
+  DEFINED=$(jq --arg t "$T" '.defined[$t]'        traces-report.json)
+  echo "   $T: $TRACED / $DEFINED"
+done
+
+MIN_THRESHOLD=50
+[ "$COVERAGE" -lt "$MIN_THRESHOLD" ] && { echo "❌ …"; exit 1; }
+```
+
+No hardcoded denominator survives anywhere in the workflow.
+
+### 3. A self-check so this cannot silently rot again
+
+The root cause was a number that drifted with nothing watching it. Add a
+guard that fails the job on an arithmetically impossible result:
+
+```sh
+if [ "$COVERAGE" -gt 100 ]; then
+  echo "❌ Coverage > 100% — the gate is miscomputing; orphaned IDs: $(jq -c '.orphaned' traces-report.json)"
+  exit 1
+fi
+```
+
+A >100% reading is now a hard failure rather than a green tick.
+
+### 4. Local parity
+
+Add a script so the gate is runnable outside CI:
+
+```jsonc
+"traces:coverage": "bun run scripts/extract-traces.ts --format coverage"
+```
+
+Prints the same table CI prints and exits non-zero below threshold.
+
+### Threshold
+
+Keep `MIN_THRESHOLD=50` in this spec. Real coverage is 85%, so raising the bar
+is tempting, but doing it in the same change that repairs the gate conflates
+"restore the safety net" with "tighten the policy" — if the build then fails, it
+is ambiguous which change caused it. Ratcheting is deliberately deferred to
+follow-up work once the honest number has been observed on `master` for a few
+builds.
+
+## Out of scope
+
+- Raising `MIN_THRESHOLD` above 50 (see above).
+- Fixing/removing `scripts/check-req-coverage.sh` — [req-coverage-script-removal.md](req-coverage-script-removal.md).
+- Adding TRACES comments to raise the actual coverage number.
+- Changing the `TRACES:` comment format or the extractor's parsing of it.
+- The PR "modified files missing TRACES" step (lines 78–126), which is advisory
+  by design and stays advisory.
+
+## Acceptance criteria
+
+- [ ] `bun run traces:json` emits `defined`, `coverage`, and `orphaned` keys.
+- [ ] `coverage.total` equals the count of requirement IDs defined in
+      `requirements.md` (**211** at time of writing), not a literal.
+- [ ] `coverage.percent` reports **85** (±1 for rounding) on the current tree —
+      i.e. the honest number, not 158.
+- [ ] No hardcoded requirement denominator (`39`, `24`, `48`, `3`, `114`) remains
+      in `.gitea/workflows/traceability-check.yml`. Verify:
+      `grep -nE '/ *(39|24|48|3|114)\b' .gitea/workflows/traceability-check.yml`
+      returns nothing.
+- [ ] Adding a new requirement row to `requirements.md` **lowers** reported
+      coverage until it is traced (proves the denominator is live).
+- [ ] A `TRACES:` comment naming an undefined ID appears in `orphaned` and does
+      **not** raise `coverage.percent`.
+- [ ] The job fails if coverage is forced below 50% (test by temporarily raising
+      `MIN_THRESHOLD` to 99 locally) — proving the gate can fail again.
+- [ ] The job fails if coverage computes >100%.
+- [ ] `bun run check` and `bun run test` pass.
+- [ ] `bun run check:boundary` passes.
+- [ ] New requirement-implementing code carries `// TRACES:` comments.
+- [ ] No Rust types changed, so no `bindings.ts` regeneration needed.
+
+## Testing
+
+`extract-traces.ts` currently has no test coverage. Add
+`scripts/extract-traces.test.ts` (vitest) over fixture strings rather than the
+live `requirements.md`, so the tests do not change meaning as requirements are
+added:
+
+- **UT:** counts a well-formed table row as a defined requirement.
+- **UT:** does **not** count an ID appearing only in the `Traces To` column or
+  in prose — the specific overcounting bug this parse rule avoids.
+- **UT:** coverage is the intersection — a traced-but-undefined ID lands in
+  `orphaned` and does not inflate the numerator.
+- **UT:** coverage of an empty trace set is 0%, not a divide-by-zero.
+- **UT:** all-traced fixture reports exactly 100%, never above.
+
+CI behaviour is verified by the acceptance criteria above (the forced-failure
+check is the important one — a gate nobody has watched fail is not known to
+work).
+
+## TRACES
+
+Allocate in `requirements.md`:
+
+- **DR-093** — "Traceability coverage gate derives requirement denominators from
+  `requirements.md` at run time (not hardcoded literals), computes coverage as
+  the intersection of traced and defined IDs, reports IDs traced but undefined
+  as orphaned, and fails on an impossible >100% result." Category: Tooling.
+  Status: Done on merge.
+
+Tag:
+
+```typescript
+// scripts/extract-traces.ts
+// TRACES: | DR-093
+```
+
+Tests carry `@req-test: UT-089 …` onward (next free UT is **UT-089**).
+
+## Notes for the implementer
+
+- A parallel Claude session may be active in this repo — run `git diff` before
+  "repairing" unexpected changes (CLAUDE.md §Gotchas).
+- **Do not add tooling to the CI image for this.** `jq` and `bun` are already in
+  `jellytau-builder`; this spec needs nothing else. Installing a system package
+  in a workflow step violates the hard CI rule in CLAUDE.md.
+- Keep `traces:json`'s existing keys intact — `release-notes.ts` and
+  `traces:markdown` consume the same report, and the CI workflow uploads it as
+  an artifact. This is an additive change.
+- The `head -50 docs/traceability.md` and artifact-upload steps are unaffected.
+- Expect the first green build after this change to print a *lower* number than
+  before (85% vs 158%). That is the fix working, not a regression.

docs/specs/windows-native-audio-backend.md

@@ -0,0 +1,209 @@
+# Spec: Windows native audio backend
+
+**Status:** Proposed
+**Requirements:** UR-003, UR-027, UR-032, UR-033 → DR-030, DR-035, DR-036; new IR-030
+**UX spec:** n/a — Settings › Audio already renders the controls
+**Supersedes / revises:** acts on the "audio can unify, video cannot" conclusion in [playback-backend-unification.md](playback-backend-unification.md)
+
+## Summary
+
+Give Windows a real native audio backend instead of the current webview
+`<audio>` shim. Windows is the only platform where audio playback has no decoder
+of its own: `WebviewAudioBackend` hands a URL to a frontend `<audio>` element and
+relays transport commands. It cannot set volume, cannot apply any audio setting,
+and reports state only via DOM events.
+
+Audio needs no rendering surface, so **none of the webview-compositing problems
+that block unified video apply here.** This is the cleanest available win.
+
+## Motivation
+
+`WebviewAudioBackend` was a deliberate stopgap ("audio-only playback for
+platforms without a native audio backend"), and it works — but it has a hard
+functional gap. From `webview_audio_backend.rs`:
+
+```rust
+fn set_volume(&mut self, volume: f32) -> Result<(), PlayerError> {
+    // ...stores locally only; there is no ControlCommand action for volume
+}
+```
+
+So volume changes never reach the element; the frontend has to observe the player
+store and apply volume itself. `set_audio_settings` likewise stores values that
+nothing consumes — EQ, normalization, and gapless are all inert on Windows.
+
+Meanwhile the backend-unification investigation established that a native *audio*
+engine is unproblematic on Windows specifically: `tauri-plugin-libmpv` lists
+Windows as its **fully tested** platform (in contrast to Linux, where embedding
+is broken — but that is a *video surface* problem, which audio does not have).
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Decoding and playing the audio stream | Rust | Playback is domain logic; every other platform already decodes in Rust or a native player. The webview shim is the anomaly. |
+| Applying `AudioSettings` (EQ/normalize/gapless) | Rust | Same `AudioSettings` contract as MPV/ExoPlayer; band layout and presets stay canonical in `settings.rs`. |
+| Position/state reporting | Rust | Restores the project's core principle — the player is the authoritative source of state. Today Windows inverts this: the DOM element is authoritative and Rust mirrors it. |
+| Volume | Rust | Currently broken precisely because it is split across the boundary. |
+| Rendering the player UI | Frontend | Unchanged. |
+
+The strongest argument for this change is the third row. CLAUDE.md states
+playback state is one-directional with the player authoritative; on Windows that
+is currently false, and the `player_report_*` round-trip exists to paper over it.
+
+## Design
+
+### Engine choice
+
+Two viable options; **libmpv is recommended** for consistency with the Linux
+audio backend.
+
+| | libmpv | GStreamer |
+|---|---|---|
+| Windows status | ✅ `tauri-plugin-libmpv` reports fully tested | ✅ works, but… |
+| Rust bindings | `libmpv2` 6.0.0, active | `gstreamer-rs` 0.25.x, excellent |
+| Cross-MSVC from Linux | ⚠️ needs prebuilt DLL + import lib | ❌ `gstreamer-sys` uses pkg-config, fights `cargo-xwin` |
+| Code reuse | ✅ `MpvBackend` logic is directly reusable | ❌ a second engine to learn |
+| Crossfade capable | ❌ single-stream chain | ✅ `audiomixer` |
+
+libmpv wins on reuse: `MpvBackend`'s `set_audio_settings` — the `af` lavfi graph
+built by `build_af_filter`, `eq_filter_entries`, `normalize_filter_entry` — is
+platform-independent and would apply unchanged.
+
+The one reason to prefer GStreamer is crossfade (UR-031), which mpv structurally
+cannot do. If crossfade becomes a priority, revisit; it would then argue for
+GStreamer on *both* Linux and Windows, which is a much larger change.
+
+### Structure
+
+Rename the cfg gate so `MpvBackend` is no longer Linux-only:
+
+```rust
+// src-tauri/src/player/mod.rs
+#[cfg(any(target_os = "linux", target_os = "windows"))]
+pub mod mpv_backend;
+```
+
+`MpvBackend::new` needs one platform-specific branch: `detect_audio_system()`
+currently probes `pactl`/`pw-cli`/`/proc/asound/cards` to pick an `ao`. On
+Windows the equivalent is `wasapi` (mpv's default), so the detection is a
+`#[cfg]` returning `"wasapi"` — no probing needed.
+
+Everything else — the event loop, the 250ms position thread, the seek-suppression
+window, the `af` filter graph — is unchanged.
+
+`WebviewAudioBackend` stays for other targets (macOS and anything else hitting
+the `not(any(...))` arm) and as the fallback if libmpv fails to initialize. The
+existing `emit_backend_init_failed` path already handles that gracefully.
+
+### Build
+
+`libmpv2-sys` is well-suited to cross-compilation: no pkg-config, vendored
+headers, pregenerated bindings (no libclang). It emits `cargo:rustc-link-lib=mpv`
+unconditionally, so the build must supply a linkable import library for
+`x86_64-pc-windows-msvc`.
+
+Keep the `build_libmpv` feature **off** — its Unix path shells out to mpv-build
+and explicitly rejects cross-compilation.
+
+🔴 Per CLAUDE.md, the prebuilt libmpv **must be added to the builder image**
+(`Dockerfile.builder` → rebuild + push via `scripts/build-builder-image.sh`), not
+installed at CI job time. `libmpv-2.dll` must also be bundled into the NSIS
+installer via `tauri.conf.json`'s resources.
+
+### Verified build mechanics
+
+The cross-compile path was tested hands-on from Linux (July 2026), not inferred:
+
+- Neither shinchiro nor zhongfly ships an `mpv.def` or MSVC `mpv.lib` — only a
+  MinGW `libmpv.dll.a`. (Several online sources claim otherwise; they are wrong.)
+- An MSVC-style import lib can be generated locally with LLVM tools only:
+  `llvm-readobj --coff-exports libmpv-2.dll` → synthesize `mpv.def` →
+  `llvm-dlltool -m i386:x86-64 -d mpv.def -l mpv.lib`. `llvm-lib /def:` produces a
+  byte-identical result.
+- A real `lld-link` link against that import lib **succeeds**, and the resulting
+  import table resolves `mpv_client_api_version` from `libmpv-2.dll`. `lld-link`
+  is the linker `cargo-xwin` uses, so this is the load-bearing step.
+- Linking directly against the shipped MinGW `libmpv.dll.a` **also** succeeds, so
+  def-generation may be skippable — but that relies on lld's GNU-archive
+  tolerance rather than a documented contract. Keep `llvm-dlltool` as the
+  fallback.
+- MinGW origin is not an ABI problem: libmpv exports a pure C ABI, and the x86-64
+  Windows calling convention is platform-defined. The upstream note that MSVC
+  cannot *build* mpv is frequently misread as "MSVC cannot *link* libmpv" — that
+  is not what it says.
+- 🔴 Never free/realloc across the DLL boundary — use `mpv_free`.
+
+Build wiring is ordinary: `cargo:rustc-link-lib=dylib=mpv` plus
+`cargo:rustc-link-search`. Nothing about libmpv conflicts with `cargo-xwin`.
+
+### Size and shipping
+
+Measured uncompressed: **93 MiB** (zhongfly `mpv-dev-lgpl-x86_64`) vs **112 MiB**
+(shinchiro, full GPL build); ~26–30 MB compressed in the `.7z`.
+
+**Ship the zhongfly LGPL build** — smaller, and there is no reason to pull the
+GPL variant in for an audio-only use.
+
+Import-table inspection confirms **no companion DLLs are needed**: every
+dependency is a system DLL (`KERNEL32`, `USER32`, `d2d1`, `DWrite`, `OPENGL32`,
+`vulkan-1`, UCRT `api-ms-win-*`). One file to bundle.
+
+93 MiB is still substantial against a Tauri app's usual few MB. Since we use mpv
+audio-only, investigate whether a pruned build (no video decoders, no libplacebo)
+is worth producing for the builder image — but treat that as an optimization,
+not a blocker.
+
+## Out of scope
+
+- Windows *video*. Stays in WebView2 + hls.js — it works and has ABR.
+- Crossfade (UR-031/DR-034) — not implemented anywhere; needs its own spec.
+- Replacing `WebviewAudioBackend` for macOS.
+- MPRIS/SMTC media-key integration — worth a follow-up, not this spec.
+
+## Acceptance criteria
+
+- [ ] Windows build produces a `MpvBackend`-backed player; `backend-init-failed` is emitted (not a crash) if libmpv is unavailable.
+- [ ] Volume control works from the UI — the current hard gap.
+- [ ] EQ, normalization, and gapless audibly take effect on Windows.
+- [ ] Position/state originate in Rust; the `<audio>` element is no longer in the audio path.
+- [ ] Seek, next/previous, and queue advance work; sleep timer stops playback.
+- [ ] `libmpv-2.dll` ships in the NSIS installer and the app runs on a clean Windows VM with no mpv installed.
+- [ ] Builder image carries the Windows libmpv artefacts; **no toolchain install added to any CI step**.
+- [ ] `bun run check`, `bun run test`, `bun run check:boundary` pass.
+- [ ] `cargo fmt` clean, `cargo clippy` clean, `bun run test:rust` passes.
+- [ ] New requirement-implementing code carries `// TRACES:` comments.
+
+## Testing
+
+**Rust**: the existing `mpv_backend_test.rs` and the `build_af_filter` /
+`normalize_filter_entry` / `eq_filter_entries` unit tests already cover the
+filter-graph logic and are platform-independent — they should pass unchanged
+under a Windows `cargo check`/test. Add a test asserting `detect_audio_system()`
+returns `wasapi` under `cfg(windows)`.
+
+**Manual, on Windows**: volume, EQ preset change, normalization toggle, gapless
+between two tracks, seek, queue advance, sleep timer. Then the packaging test —
+install the NSIS output on a clean VM and confirm it launches and plays.
+
+Per CLAUDE.md, the volume gap is a *bug fix*: write a failing test for
+"`set_volume` reaches the backend" before implementing.
+
+## TRACES
+
+- Windows `MpvBackend` construction in `create_player_backend` → `// TRACES: UR-003 | IR-030`
+- `detect_audio_system` Windows branch → `IR-030`
+- Existing `set_audio_settings` gains Windows coverage → `UR-027, UR-032, UR-033 | DR-030, DR-035, DR-036`
+- Allocate **IR-030** in `requirements.md` ("libmpv integration for Windows audio playback").
+
+## Notes for the implementer
+
+- Do this **after** [libmpv2-migration.md](libmpv2-migration.md) — porting the
+  current dead `libmpv` git pin to a second platform would double the migration
+  work.
+- `libmpv2` has broken its API in every major release (4.0 removed command
+  helpers, 5.0 removed `mpv_node`, 6.0 changed `RenderContext` ownership). Pin an
+  exact version.
+- Only the `render`-feature parts of `libmpv2` concern video; audio-only use does
+  not need it, and disabling the default `render` feature may shrink the build.
+- A parallel Claude session may be active — `git diff` first.

docs/traceability-ci.md

@@ -43,14 +43,26 @@ Extracts all TRACES comments from:
 
 ### 2. Coverage Thresholds
 The workflow checks:
-- **Minimum overall coverage:** 50% (57+ requirements traced)
-- **Requirements by type:**
-  - UR (User): 23+ of 39
-  - IR (Integration): 5+ of 24
-  - DR (Development): 28+ of 48
-  - JA (Jellyfin API): 0+ of 3
+- **Minimum overall coverage:** 50%
 
-If coverage drops below threshold, the workflow **fails** and blocks merge.
+Denominators are **derived from `docs/requirements.md` at run time** — they are
+never hardcoded here or in the workflow. Run `bun run traces:coverage` for the
+current per-type breakdown; any number written into this document is a snapshot
+that will drift.
+
+> **Why this matters.** The workflow used to divide by frozen literals
+> (UR/39, IR/24, DR/48, JA/3, total 114) while `requirements.md` had grown past
+> 200. It reported **158%** coverage, so the 50% threshold was unreachable and
+> the job could not fail regardless of how far coverage dropped. See
+> [specs/traceability-gate-repair.md](specs/traceability-gate-repair.md).
+
+Coverage is the *intersection* of traced and defined IDs: an ID that appears in
+a `TRACES:` comment but is not defined in `requirements.md` is reported as
+**orphaned** and does not count toward coverage. UT/IT test identifiers are a
+separate taxonomy and are excluded entirely.
+
+The workflow **fails** and blocks merge if coverage drops below 50% — or if it
+computes above 100%, which can only mean the gate is miscounting.
 
 ### 3. Modified File Checking
 On pull requests, the workflow:
@@ -153,11 +165,13 @@ cat docs/traceability.md
 ## Coverage Goals
 
 ### Current Status
-- Overall: 51% (56/114)
-- UR: 59% (23/39)
-- IR: 21% (5/24)
-- DR: 58% (28/48)
-- JA: 0% (0/3)
+
+Run `bun run traces:coverage` — it prints the live figure and exits non-zero
+below threshold. Numbers are deliberately not pinned here; the previous snapshot
+in this section (51%, 56/114) was stale by roughly 100 requirements and was what
+made the broken CI arithmetic look plausible for so long.
+
+As of July 2026 overall coverage is ~86% (182/212).
 
 ### Targets
 - **Short term** (Sprint): Maintain ≥50% overall

docs/ux-flows.md

@@ -346,10 +346,12 @@ flowchart TB
 **User Interaction:**
 - **Tap screen:** Controls reappear for 3 seconds
 - **Double tap left side:** Rewind 10 seconds (shows animated feedback with "-10" indicator)
-- **Double tap right side:** Forward 10 seconds (shows animated feedback with "+10" indicator)
+- **Double tap right side:** Forward 30 seconds (shows animated feedback with "+30" indicator)
+- **Single tap play/pause is deferred** by the 300 ms double-tap window, so a double tap
+  skips without also toggling pause (UR-061)
 - **Swipe up/down on left side:** Adjust brightness (0.3-1.7x, shows brightness indicator with progress bar)
 - **Swipe up/down on right side:** Adjust volume (0-100%, shows volume indicator with progress bar)
-- **Keyboard arrows:** ← rewind 10s, → forward 10s (desktop/external keyboard)
+- **Keyboard arrows:** ← rewind 10s, → forward 30s (desktop/external keyboard)
 - **Keyboard space/K:** Toggle play/pause
 - **Keyboard F:** Toggle fullscreen
 - **Pinch:** Zoom (planned)
@@ -609,9 +611,12 @@ flowchart TB
 ```
 
 An episode is **never** browsed as a bare `Episode` item page. Clicking an
-episode anywhere navigates to `/library/<seriesId>?episode=<episodeId>`, so the
-episode is always shown in the context of its series and the series' full
-episode list is already loaded.
+episode anywhere — a series' season list, a Home carousel (§5B.5), etc. —
+navigates to `/library/<seriesId>?episode=<episodeId>`, so the episode is always
+shown in the context of its series and the series' full episode list is already
+loaded. Should an episode ever arrive without a `seriesId` (deep link, stale
+cache), the bare Episode page renders as a fallback and links back to its parent
+series and season by title so the user is never stranded.
 
 ### 5B.2 Episode Focus View — section order
 
@@ -629,7 +634,7 @@ episode strip.
 │   │  S2E4 • 48m • ★8.1                        │ │
 │   │  Overview…                                │ │
 │   │  ▓▓▓▓▓░░░░░  32m left                     │ │
-│   │  [▶ Play]                                 │ │
+│   │  [▶ Play]  [⬇]  [♡]                       │ │
 │   └───────────────────────────────────────────┘ │
 │                                                 │
 │  More Episodes                                  │  ← 2. EPISODE STRIP
@@ -683,10 +688,10 @@ A movie has no continuation set, so cast follows the hero directly.
 ### 5B.4 Series detail — section order
 
 ```
-Hero (poster, title, metadata, Play / Download)
+Hero (poster, title, metadata, Resume SxEy / Download / Favorite / Clear history)
   → Crew links
   → Genre tags
-  → Seasons + episodes (per-season sections)
+  → Seasons (collapsible; only the current season expanded)
   → Cast
   → More Like This
 ```
@@ -695,6 +700,185 @@ The same principle as §5B.2: **episodes come before cast and similar shows.**
 The reason a user opens a series page is to pick an episode; discovery content
 is secondary and sits underneath.
 
+**Rules for the seasons block** *(UR-062, UR-064)*:
+
+- **The page opens where the viewer is.** The backend resolves the current
+  episode — in progress, else Next Up, else first unwatched, else the premiere —
+  and the page scrolls it into view with an `Up next` badge and a highlight ring.
+  Never season 1 by default, unless season 1 *is* where the viewer is.
+- **Seasons collapse; only the current one is expanded.** A ten-season show
+  otherwise renders hundreds of rows and buries the episode the viewer came for.
+  A collapsed season still names its episode count and watched count, so
+  progress is readable without expanding it.
+- **The hero button opens, it does not play.** It reads `Resume S2E4` /
+  `Play S1E1` — naming its target — and navigates to that episode's Focus View,
+  where Play commits. Play on a *container* is navigation (§5B.5); Play on a
+  *leaf* is the commitment.
+- **A season is never its own page.** `/library/<seasonId>` redirects to
+  `/library/<seriesId>#season-N`. Every affordance that names a season — the
+  episode breadcrumb, a season card in a grid, a Downloads drill-in — lands on
+  the series with that season in view, so the episodes of all seasons stay one
+  browsable list.
+- **Watch history is erasable** per series (hero) and per season (season
+  header). It confirms first, cannot be undone, and needs the server. Clearing a
+  whole series returns it to S1E1 by the same path a never-watched show takes.
+
+### 5B.5 Home-card interaction — tap opens, long-press plays
+
+Cards on the Home screen carousels (Next Movie, Next Episode, Continue
+Watching, Recently Added, …) **do not play on tap.** A plain tap opens the
+item; playback is the deliberate, second gesture.
+
+| Card kind | Tap (short) | Long-press (~500 ms hold) |
+|-----------|-------------|---------------------------|
+| Movie | Movie detail page (`/library/<id>`) | Confirm → play now (`/player/<id>`) |
+| Episode | Series Episode Focus View (`/library/<seriesId>?episode=<id>`, per §5B.1) | Confirm → play now (`/player/<id>`) |
+| Series / Season / Album / Artist / Playlist / Folder | Detail page (`/library/<id>`) | Same as tap (no single "play now" target) |
+| Channel / live leaf | Player (`/player/<id>`) — no detail page exists | Confirm → play now |
+
+Rationale and rules:
+
+- **Tap is navigation, not commitment.** Previously a tap on a movie/episode
+  jumped straight into the player, which made it easy to lose your place in a
+  half-watched item or start a stream you only meant to inspect. Tap now lands
+  on the detail/focus page, where Play is an explicit button.
+- **Long-press is the shortcut for "just play it."** It surfaces a native
+  confirm (`Play "<name>" now?`) before starting playback, so an accidental
+  hold never blows away a resume position silently.
+- **The long-press must not fight the carousel.** Detection cancels if the
+  pointer moves more than ~10 px (a horizontal scroll of the row), so holding
+  to scroll never triggers play.
+- **Episodes still obey §5B.1** — a home tap on an episode opens the series
+  Focus View, never a bare Episode page, so the series context loads.
+
+This behavior lives in `MediaCard` (`onLongPress` prop + pointer-based
+detection) so any surface can opt in; today the Home carousels are the only
+opt-in. Grids and other surfaces keep tap-to-open with no long-press.
+
+---
+
+## 5C. Favourites
+
+Favouriting is a two-sided promise: the heart takes the input, and the app must
+be able to give it back. This section covers both sides — where you can mark a
+favourite, and where marked favourites resurface.
+
+See [specs/favorites-browsing.md](specs/favorites-browsing.md) for the layer
+assignment and wire shapes.
+
+### 5C.1 The heart appears wherever an item does
+
+A favourite is a property of an *item*, so the affordance follows the item
+rather than living on one privileged screen. Any surface that shows a whole
+item shows its heart.
+
+| Surface | Heart position | Notes |
+|---------|----------------|-------|
+| Movie / Series detail hero | In the button row, after Play and Download | §5B.3, §5B.4 |
+| Episode Focus View hero | Same row as Play / Download | §5B.2 |
+| Album, Artist, Playlist detail | In the header button row | §5.2 |
+| Media card (any grid or carousel) | Top-right overlay on the artwork | Hidden on server-only (greyed) cards |
+| Mini player | Right of the track metadata | Existing behaviour, unchanged |
+| Full player | Secondary controls row | §3.2 — **not yet built**, see §5C.5 |
+
+Rules:
+
+- **The heart never competes with the card.** On a media card it is its own
+  button and swallows the tap, so hearting an item never also opens or plays
+  it, and never triggers the §5B.5 long-press.
+- **State is shown, not guessed.** A filled heart means the *server* considers
+  the item a favourite (or you just tapped it). An item favourited in Jellyfin
+  Web, on another device, or by another client renders filled here without
+  being touched in JellyTau.
+- **Feedback is immediate.** The heart fills on tap and a toast confirms;
+  neither waits for the server round-trip.
+
+### 5C.2 Three ways back to what you favourited
+
+Favourites are not one destination — they are a lens, and the right surface
+depends on whether the user is *browsing*, *deciding*, or *hunting*.
+
+```mermaid
+flowchart TB
+    User[User wants their favourites] --> How{Intent}
+
+    How -->|Passive: show me something| Home[Home carousels<br/>Favourite Movies / Shows / Music]
+    How -->|Deliberate: my whole collection| Page[Favourites page<br/>/library/favorites]
+    How -->|Narrowing: within this library| Filter[Favourites filter<br/>on a library page]
+
+    Home -->|See all| Page
+    Page --> Detail[Item detail page]
+    Filter --> Detail
+```
+
+**Home carousels.** Rows for favourite movies, shows and music sit below
+*Recently Added*. A row with nothing in it **does not render** — a fresh install
+shows no empty favourite rows. Each row ends with *See all*, landing on the
+matching tab of the Favourites page.
+
+**The Favourites page** (`/library/favorites`) is the complete collection,
+scoped by tabs:
+
+```
+┌─────────────────────────────────────────────────┐
+│ [←]  Favourites                                 │
+│  ┌─────┬────────┬───────┬───────┐               │
+│  │ All │ Movies │ Shows │ Music │  ← scope tabs │
+│  └─────┴────────┴───────┴───────┘               │
+│                                                 │
+│  ┌────┐┌────┐┌────┐┌────┐┌────┐                 │
+│  │ ♥  ││ ♥  ││ ♥  ││ ♥  ││ ♥  │   grid/list    │
+│  └────┘└────┘└────┘└────┘└────┘   per §5A      │
+└─────────────────────────────────────────────────┘
+```
+
+- Cards obey §5A in full — shape follows the media, so a mixed *All* tab reads
+  as posters, squares and thumbnails side by side rather than one forced shape.
+- Reached from a card on the library overview (`/library`) and from *See all*
+  on any home favourites row.
+- Sorted by name. Jellyfin does not record *when* an item was favourited, so
+  "recently favourited" is not offerable — see §5C.5.
+- Empty state, per tab: *"Nothing favourited yet — tap the heart on anything
+  you like."*
+
+**The in-library filter** is for narrowing where the user already is: a
+favourites toggle in the header of the Movies, TV and Music browse pages,
+filtering the current list in place. It is **session-scoped and not persisted** —
+a sticky filter that silently hides most of a library reads as data loss on the
+next launch.
+
+### 5C.3 Removing a favourite removes it everywhere, at once
+
+Un-hearting an item on the Favourites page removes its card from the grid
+immediately; the same item disappears from the home rows and shows an empty
+heart on its detail page without a manual refresh. The reverse holds for
+favouriting. There is no confirmation prompt — the action is one tap to undo.
+
+### 5C.4 Offline
+
+- **Marking works offline.** The heart fills, the toast confirms, and the change
+  is held locally.
+- **It reaches the server on reconnect**, without the user returning to the
+  screen where they made it.
+- **Browsing offline shows favourites among media on the device**, subject to
+  the same "Show all server media" gate as every other browse surface (§7.2) —
+  with the gate off, an empty Favourites tab means *nothing favourited is
+  downloaded*, and the page does not quietly fall back to the server catalog.
+
+### 5C.5 Known deviations
+
+- **The full player has no heart.** §3.2 and §3.3 list a Favorite button among
+  the full player's secondary controls; it was never built, and this pass does
+  not add it. The mini player heart above it is the only in-player affordance.
+  *(UR-067)*
+- **No "recently favourited" sort.** Jellyfin's API does not expose a favourite
+  timestamp, so favourites can only be ordered by name. Recording the
+  timestamp locally at toggle time would order *this device's* favourites only,
+  which is worse than a consistent name sort.
+- **Music is one tab, not three.** The Music scope mixes albums, artists and
+  tracks in a single grid rather than offering sub-tabs. Acceptable while
+  favourite counts are small; revisit if the tab becomes unscannable.
+
 ---
 
 ## 6. Search Flow

package.json

@@ -1,6 +1,6 @@
 {
   "name": "jellytau",
-  "version": "0.0.16",
+  "version": "0.6.0",
   "description": "",
   "type": "module",
   "packageManager": "bun@1.3.5",
@@ -20,16 +20,25 @@
     "check:boundary": "bash scripts/check-frontend-boundary.sh",
     "android:build": "./scripts/build-android.sh",
     "android:build:release": "./scripts/build-android.sh release",
+    "android:build:device": "./scripts/build-android.sh --device",
+    "android:build:release:device": "./scripts/build-android.sh release --device",
     "android:build:clean": "rm -rf node_modules/.vite dist .svelte-kit .next build target src-tauri/target && bun install && bun run build",
     "android:deploy": "./scripts/deploy-android.sh",
     "android:dev": "./scripts/build-and-deploy.sh",
     "android:check": "./scripts/check-android.sh",
     "android:logs": "./scripts/logcat.sh",
+    "desktop:build:linux": "./scripts/build-desktop-linux.sh",
+    "desktop:build:arch": "./scripts/build-arch.sh",
+    "desktop:build:windows": "./scripts/build-windows-cross.sh",
+    "docker:build:linux": "docker compose run --rm desktop-linux-build",
+    "docker:build:arch": "docker compose run --rm arch-build",
+    "docker:build:windows": "docker compose run --rm windows-cross",
     "clean": "./scripts/clean.sh",
     "tauri": "tauri",
     "traces": "bun run scripts/extract-traces.ts",
     "traces:json": "bun run scripts/extract-traces.ts --format json",
     "traces:markdown": "bun run scripts/extract-traces.ts --format markdown > docs/traceability.md",
+    "traces:coverage": "bun run scripts/extract-traces.ts --format coverage",
     "release:notes": "bun run scripts/release-notes.ts"
   },
   "license": "MIT",

packaging/arch/PKGBUILD

@@ -0,0 +1,52 @@
+# Maintainer: Duncan Tourolle <duncan@tourolle.paris>
+#
+# JellyTau — a cross-platform Jellyfin client (Tauri + SvelteKit).
+#
+# This PKGBUILD builds from the local source tree by default (see the `dev`
+# convenience below), which is what scripts/build-arch.sh uses inside the Arch
+# Docker stage. For AUR distribution, replace the `source=()` line with a release
+# tarball/VCS URL and drop the local-copy prepare() step.
+
+pkgname=jellytau
+pkgver=0.0.18
+pkgrel=1
+pkgdesc="A cross-platform Jellyfin client"
+arch=('x86_64')
+url="https://gitea.tourolle.paris/dtourolle/jellytau"
+license=('MIT')
+# Runtime: libmpv for audio, webkit2gtk for the webview + HTML5 transcoded video.
+depends=('webkit2gtk-4.1' 'mpv' 'gtk3' 'libayatana-appindicator')
+makedepends=('rust' 'cargo' 'bun' 'nodejs' 'pkgconf' 'libsoup3')
+options=('!strip' '!lto')
+
+# Populated from the working tree by scripts/build-arch.sh (SRC env var).
+_srcdir="${JELLYTAU_SRC:-$startdir/../..}"
+
+build() {
+  cd "$_srcdir"
+  export CARGO_HOME="${CARGO_HOME:-$srcdir/cargo-home}"
+  bun install --frozen-lockfile || bun install
+  bun run build
+  # Only the raw binary is needed; packaging is done in package() below so we
+  # control the Arch filesystem layout ourselves rather than via tauri-bundler.
+  (cd src-tauri && cargo build --release --locked)
+}
+
+package() {
+  cd "$_srcdir"
+
+  install -Dm755 "src-tauri/target/release/jellytau" \
+    "$pkgdir/usr/bin/jellytau"
+
+  # Desktop entry
+  install -Dm644 "packaging/arch/jellytau.desktop" \
+    "$pkgdir/usr/share/applications/jellytau.desktop"
+
+  # Icons (hicolor)
+  install -Dm644 "src-tauri/icons/32x32.png" \
+    "$pkgdir/usr/share/icons/hicolor/32x32/apps/jellytau.png"
+  install -Dm644 "src-tauri/icons/128x128.png" \
+    "$pkgdir/usr/share/icons/hicolor/128x128/apps/jellytau.png"
+  install -Dm644 "src-tauri/icons/128x128@2x.png" \
+    "$pkgdir/usr/share/icons/hicolor/256x256/apps/jellytau.png"
+}

packaging/arch/jellytau.desktop

@@ -0,0 +1,9 @@
+[Desktop Entry]
+Type=Application
+Name=JellyTau
+Comment=A cross-platform Jellyfin client
+Exec=jellytau
+Icon=jellytau
+Terminal=false
+Categories=AudioVideo;Player;Audio;Video;
+StartupWMClass=jellytau

scripts/README.md

@@ -69,13 +69,29 @@ Extract requirement IDs (TRACES) from source code and generate a traceability ma
 bun run traces                    # Generate markdown report
 bun run traces:json              # Generate JSON report
 bun run traces:markdown           # Save to docs/traceability.md
+bun run traces:coverage           # Coverage gate — exits non-zero below 50%
 ```
 
-The script scans all TypeScript, Svelte, and Rust files looking for `TRACES:` comments and generates a comprehensive mapping of:
+The script scans all TypeScript, Svelte, and Rust files (plus `scripts/`)
+looking for `TRACES:` comments and generates a comprehensive mapping of:
 - Which code files implement which requirements
 - Line numbers and code context
 - Coverage summary by requirement type (UR, IR, DR, JA)
 
+**`bun run traces:coverage` is the supported way to check requirement coverage
+locally** — it runs the same computation CI does. Coverage denominators are
+derived from `docs/requirements.md` at run time; they are never hardcoded. An ID
+that appears in a `TRACES:` comment but is not defined in `requirements.md` is
+reported as *orphaned* and does not count toward coverage (see DR-093).
+
+> **Removed:** `check-req-coverage.sh`, `check-test-coverage.sh`, and
+> `find-req-implementations.sh` were deleted in July 2026. They read an
+> undocumented `@req:` tag convention parallel to `TRACES:`, grepped `src-tauri/`
+> unscoped (hanging on ~40 GB of `target/` artifacts), and in one case reported
+> "all requirements implemented" from an empty result set. `extract-traces.ts` is
+> the single source of truth for requirement coverage. See
+> [docs/specs/req-coverage-script-removal.md](../docs/specs/req-coverage-script-removal.md).
+
 Example TRACES comment in code:
 ```typescript
 // TRACES: UR-005, UR-026 | DR-029

scripts/build-and-deploy.sh

@@ -11,11 +11,13 @@ echo ""
 
 echo ""
 
-# Deploy APK — extract build type (default debug), ignoring flags like --clean.
-BUILD_TYPE="debug"
+# Deploy APK — forward the build type and the side-by-side flag (which decides
+# which package to launch), ignoring build-only flags like --clean and --device.
+DEPLOY_ARGS=("debug")
 for arg in "$@"; do
     case "$arg" in
-        debug|release) BUILD_TYPE="$arg" ;;
+        debug|release) DEPLOY_ARGS[0]="$arg" ;;
+        --debug|--side-by-side) DEPLOY_ARGS+=("--side-by-side") ;;
     esac
 done
-./scripts/deploy-android.sh "$BUILD_TYPE"
+./scripts/deploy-android.sh "${DEPLOY_ARGS[@]}"

scripts/build-android.sh

@@ -18,15 +18,66 @@ echo ""
 # Parse args: build type (debug/release) and optional --clean flag.
 # By default the build is INCREMENTAL — Cargo and Vite reuse their caches.
 # Pass --clean (or CLEAN=1) to wipe all caches for a from-scratch build.
+#
+# ABI selection: by default Tauri builds all four ABIs (arm64/arm/x86/x86_64),
+# which is what a distributable universal APK needs — but for an on-device test
+# it means three wasted Rust compiles. Pass --device (or ABI=aarch64) to build
+# only the connected device's architecture; --abi <t> targets one explicitly.
+#
+# Side-by-side: the `debug` build type always installs as
+# com.dtourolle.jellytau.debug ("JellyTau Debug"), so it never collides with a
+# real install. `release --debug` puts a *release* build — R8-minified, exactly
+# what ships — into that same slot, signed with the local debug keystore. That
+# is how you validate minification (R8 stripping JNI-loaded classes has broken
+# release APKs here before) without the real signing key and without
+# uninstalling the app you actually use.
 BUILD_TYPE="debug"
 CLEAN="${CLEAN:-0}"
+ABI="${ABI:-}"
+SIDE_BY_SIDE="${SIDE_BY_SIDE:-0}"
+next_is_abi=0
 for arg in "$@"; do
+    if [ "$next_is_abi" = "1" ]; then
+        ABI="$arg"
+        next_is_abi=0
+        continue
+    fi
     case "$arg" in
         --clean) CLEAN=1 ;;
+        --abi) next_is_abi=1 ;;
+        --device) ABI="device" ;;
+        --debug|--side-by-side) SIDE_BY_SIDE=1 ;;
         debug|release) BUILD_TYPE="$arg" ;;
     esac
 done
 
+# The debug build type is side-by-side unconditionally; the flag only means
+# something for a release build.
+if [ "$BUILD_TYPE" = "debug" ]; then
+    SIDE_BY_SIDE=1
+fi
+
+# Resolve --device to the attached device's Rust target triple.
+if [ "$ABI" = "device" ]; then
+    device_abi="$(adb shell getprop ro.product.cpu.abi 2>/dev/null | tr -d '\r\n')"
+    case "$device_abi" in
+        arm64-v8a)   ABI="aarch64" ;;
+        armeabi-v7a) ABI="armv7" ;;
+        x86_64)      ABI="x86_64" ;;
+        x86)         ABI="i686" ;;
+        *)
+            echo "⚠️  Could not detect device ABI (got '${device_abi:-none}') — building all targets."
+            ABI=""
+            ;;
+    esac
+    [ -n "$ABI" ] && echo "🎯 Device ABI $device_abi → building only '$ABI'"
+fi
+
+TARGET_ARGS=()
+if [ -n "$ABI" ]; then
+    TARGET_ARGS=(--target "$ABI")
+fi
+
 # Step 0: Optionally clear build caches for a fully fresh build.
 if [ "$CLEAN" = "1" ]; then
     echo "🧹 Clearing build caches (clean build)..."
@@ -43,15 +94,22 @@ echo "🎨 Building frontend..."
 bun run build
 
 # Step 2: Build Android APK
-if [ "$BUILD_TYPE" = "release" ]; then
+if [ "$BUILD_TYPE" = "release" ] && [ "$SIDE_BY_SIDE" = "1" ]; then
+    # A release build in the debug slot: R8 still runs, but the applicationId is
+    # suffixed and the debug keystore signs it (read by build.gradle.kts from
+    # JT_SIDE_BY_SIDE), so the real key is not needed and it replaces any other
+    # .debug install cleanly. Deliberately does NOT write keystore.properties.
+    echo "📦 Building side-by-side release APK (com.dtourolle.jellytau.debug)..."
+    JT_SIDE_BY_SIDE=1 bun run tauri android build --apk true "${TARGET_ARGS[@]}"
+elif [ "$BUILD_TYPE" = "release" ]; then
     # Configure release signing from .env (single source of truth). Must run
     # after sync-android-sources.sh, since gen/android is (re)generated there.
     ./scripts/write-keystore-properties.sh
     echo "📦 Building release APK..."
-    bun run tauri android build --apk true
+    bun run tauri android build --apk true "${TARGET_ARGS[@]}"
 else
     echo "📦 Building debug APK..."
-    bun run tauri android build --apk true --debug
+    bun run tauri android build --apk true --debug "${TARGET_ARGS[@]}"
 fi
 
 echo ""

scripts/build-arch.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+# Build an Arch Linux package (.pkg.tar.zst) for JellyTau via makepkg.
+#
+# Tauri's bundler has no pacman target (as of tauri-cli 2.9.x), so we ship a
+# hand-written PKGBUILD in packaging/arch/ and build it with makepkg. This must
+# run on an Arch host / the `arch-build` Docker stage — makepkg is Arch-specific
+# and refuses to run as root, so run it as a non-root user with sudo for deps.
+#
+# Usage (typically inside the arch-build Docker stage as a non-root user):
+#   scripts/build-arch.sh
+#   OUTPUT_DIR=/app/dist scripts/build-arch.sh
+set -euo pipefail
+
+REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+cd "$REPO_ROOT/packaging/arch"
+
+echo "🏛️  Building JellyTau Arch package"
+echo "=================================="
+
+# Point the PKGBUILD at the working tree and give cargo/bun a writable home.
+export JELLYTAU_SRC="$REPO_ROOT"
+export CARGO_HOME="${CARGO_HOME:-$REPO_ROOT/.cargo-arch}"
+
+# -s installs missing deps (needs sudo/root privileges for pacman), -f overwrites.
+makepkg -sf --noconfirm
+
+echo ""
+echo "✅ Built Arch package(s):"
+ls -1 ./*.pkg.tar.zst
+
+if [[ -n "${OUTPUT_DIR:-}" ]]; then
+  mkdir -p "$OUTPUT_DIR"
+  cp -v ./*.pkg.tar.zst "$OUTPUT_DIR/"
+  echo ""
+  echo "📦 Copied Arch package(s) to $OUTPUT_DIR"
+fi

scripts/build-desktop-linux.sh

@@ -0,0 +1,44 @@
+#!/bin/bash
+# Build Linux desktop packages (deb + rpm) for JellyTau.
+#
+# Produces bundles under src-tauri/target/release/bundle/{deb,rpm}.
+# Runs on the existing Ubuntu builder image. NOTE: Tauri has no pacman bundle
+# target — the Arch package is built separately with makepkg (scripts/build-arch.sh
+# / Dockerfile.arch). `appimage` is also available if you want a portable bundle.
+#
+# Usage:
+#   scripts/build-desktop-linux.sh                 # deb + rpm
+#   BUNDLES="deb,appimage" scripts/build-desktop-linux.sh # subset / add appimage
+#   OUTPUT_DIR=/app/dist scripts/build-desktop-linux.sh   # copy bundles out
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+BUNDLES="${BUNDLES:-deb,rpm}"
+
+echo "🐧 Building JellyTau Linux desktop packages"
+echo "==========================================="
+echo "Bundles: $BUNDLES"
+echo ""
+
+bun install --frozen-lockfile 2>/dev/null || bun install
+bun run build
+
+# --bundles overrides tauri.conf.json bundle.targets so this script controls
+# exactly which Linux formats are produced (never NSIS here).
+bun run tauri build --bundles "$BUNDLES"
+
+BUNDLE_ROOT="src-tauri/target/release/bundle"
+echo ""
+echo "✅ Built packages:"
+find "$BUNDLE_ROOT" -maxdepth 2 -type f \
+  \( -name '*.deb' -o -name '*.rpm' -o -name '*.AppImage' \) -print
+
+if [[ -n "${OUTPUT_DIR:-}" ]]; then
+  mkdir -p "$OUTPUT_DIR"
+  find "$BUNDLE_ROOT" -maxdepth 2 -type f \
+    \( -name '*.deb' -o -name '*.rpm' -o -name '*.AppImage' \) \
+    -exec cp -v {} "$OUTPUT_DIR/" \;
+  echo ""
+  echo "📦 Copied bundles to $OUTPUT_DIR"
+fi

scripts/build-windows-cross.sh

@@ -0,0 +1,69 @@
+#!/bin/bash
+# Cross-compile JellyTau for Windows from Linux, producing an NSIS installer.
+#
+# Uses the OFFICIAL Tauri cross-compile path (https://v2.tauri.app/distribute/
+# windows-installer/): the MSVC target driven by cargo-xwin, which downloads the
+# MSVC CRT/Windows SDK headers and links with lld. This is the target Tauri
+# officially supports for Windows (the mingw/GNU target is not), and unlike GNU
+# it can bundle the NSIS installer from a Linux host.
+#
+# Playback on Windows: video renders via WebView2 and audio via the webview
+# <audio> backend (WebviewAudioBackend) — see docs/build-windows.md.
+#
+# Requirements (present in the Docker windows-cross target / unified builder):
+#   - rustup target x86_64-pc-windows-msvc
+#   - cargo-xwin           (cargo install --locked cargo-xwin)
+#   - lld, llvm            (linker + llvm-lib used by cargo-xwin)
+#   - nsis (makensis)      (installer generator)
+#
+# Usage:
+#   scripts/build-windows-cross.sh                    # exe + NSIS installer
+#   WIN_BUNDLES=none scripts/build-windows-cross.sh   # exe only, skip bundling
+#   OUTPUT_DIR=/app/dist scripts/build-windows-cross.sh
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+TARGET="x86_64-pc-windows-msvc"
+WIN_BUNDLES="${WIN_BUNDLES:-nsis}"
+
+echo "🪟 Cross-compiling JellyTau for Windows ($TARGET, via cargo-xwin)"
+echo "================================================================"
+echo "Video plays via WebView2; audio via the webview <audio> backend."
+echo "Bundles: $WIN_BUNDLES"
+echo ""
+
+bun install --frozen-lockfile 2>/dev/null || bun install
+bun run build
+
+# --runner cargo-xwin + the MSVC target is what makes the Tauri CLI treat this as
+# a real Windows build and enable the nsis/msi bundlers on a Linux host.
+#
+# IMPORTANT: do NOT pass `--bundles nsis` here. tauri-cli 2.9.x validates the
+# `--bundles` flag against a static clap enum gated by the HOST OS (Linux allows
+# only deb/rpm/appimage) *before* it considers --target/--runner, so `--bundles
+# nsis` is rejected at arg-parse time. Instead the Windows bundle targets come
+# from tauri.conf.json (bundle.targets includes "nsis"), which is not subject to
+# that CLI validation — the bundler then picks nsis once it knows the target is
+# Windows.
+if [[ "$WIN_BUNDLES" == "none" ]]; then
+  bun run tauri build --runner cargo-xwin --target "$TARGET" --no-bundle
+else
+  bun run tauri build --runner cargo-xwin --target "$TARGET"
+fi
+
+BIN_DIR="src-tauri/target/$TARGET/release"
+echo ""
+echo "✅ Built Windows artifacts:"
+find "$BIN_DIR" -maxdepth 1 -name '*.exe' -print
+find "$BIN_DIR/bundle" -type f \( -name '*.exe' -o -name '*.msi' \) -print 2>/dev/null || true
+
+if [[ -n "${OUTPUT_DIR:-}" ]]; then
+  mkdir -p "$OUTPUT_DIR"
+  find "$BIN_DIR" -maxdepth 1 -name 'jellytau.exe' -exec cp -v {} "$OUTPUT_DIR/" \;
+  # NSIS setup installers land in bundle/nsis/*-setup.exe; MSI in bundle/msi/*.msi.
+  find "$BIN_DIR/bundle" -type f \( -name '*-setup.exe' -o -name '*.msi' \) \
+    -exec cp -v {} "$OUTPUT_DIR/" \; 2>/dev/null || true
+  echo ""
+  echo "📦 Copied Windows artifacts to $OUTPUT_DIR"
+fi

scripts/check-frontend-boundary.sh

@@ -1,6 +1,8 @@
 #!/usr/bin/env bash
 # Boundary tripwire: flag domain-taxonomy leaks in the Svelte frontend.
 #
+# Implements DR-094 (see docs/requirements.md).
+#
 # The project rule (CLAUDE.md, docs/architecture/02-svelte-frontend.md) is that
 # the frontend is presentation-only and the Rust backend owns domain logic —
 # including Jellyfin's item-type *taxonomy* (what the category "Music" means as a
@@ -9,18 +11,29 @@
 #
 # ⚠️ This is a TRIPWIRE, NOT A PROOF. A grep cannot distinguish taxonomy-as-policy
 # (a leak) from taxonomy-as-display (legitimate: "is this a music card?"). It
-# targets the one machine-detectable signature of the leak class — a *query* that
-# names a multi-type category — and defers everything subtler to the human
-# spec-review checklist (docs/specs/SPEC-REVIEW-CHECKLIST.md). A clean run here
-# does not mean the boundary is respected; it means the crudest violation isn't
-# present.
+# targets the machine-detectable signature of the leak class and defers
+# everything subtler to the human spec-review checklist
+# (docs/specs/SPEC-REVIEW-CHECKLIST.md). A clean run here does not mean the
+# boundary is respected; it means the crudest violation isn't present.
 #
-# What it flags: an `includeItemTypes: [ ... , ... ]` array literal with two or
-# more types — i.e. the frontend deciding that a *category* maps to a *set* of
-# Jellyfin types, which is domain knowledge the backend should own. Single-type
-# query arrays (`includeItemTypes: ["Movie"]`) are a page saying "I show movies"
-# and are allowed. Type *inspection* (`item.type === "Audio"`) is display logic
-# and is not matched.
+# What it flags: an array literal naming two or more Jellyfin item types,
+# ANYWHERE in src/ — i.e. the frontend deciding that a *category* maps to a *set*
+# of Jellyfin types, which is domain knowledge the backend should own.
+# Single-type arrays (`includeItemTypes: ["Movie"]`) are a page saying "I show
+# movies" and are allowed. Type *inspection* (`item.type === "Audio"`) is display
+# logic and is not matched.
+#
+# 🔴 What it still CANNOT see (do not read a green run as proof):
+#   - a type set built at run time:      [...musicTypes, "Playlist"]
+#   - types split across variables:      const A = "Audio"; [A, B]
+#   - taxonomy as control flow:          switch (t) { case "Audio": … }
+#                                        t === "Audio" || t === "MusicAlbum"
+#   - an item type absent from ITEM_TYPES below (false negative by design)
+#
+# This check was hardened in July 2026 after the audit found it passing on the
+# very leak it was written for: the original pattern was anchored to
+# `includeItemTypes:` at the query site, so assigning the same array to a named
+# const evaded it entirely. See docs/specs/boundary-tripwire-hardening.md (DR-094).
 #
 # Escaping a genuine exception: add the file+reason to the ALLOWLIST below.
 
@@ -28,7 +41,7 @@ set -euo pipefail
 
 cd "$(dirname "$0")/.."
 
-# Files permitted to contain a multi-type includeItemTypes query, with the reason.
+# Files permitted to contain a multi-type item-type array, with the reason.
 # Keep this SHORT. A growing allowlist means the boundary is eroding — that is a
 # signal to push taxonomy into Rust, not to keep appending here.
 ALLOWLIST=(
@@ -36,8 +49,31 @@ ALLOWLIST=(
   # two-type filmography query with no category-configuration behind it. Tracked
   # as acceptable pending any person-scope work; revisit if it grows.
   "src/lib/components/library/PersonDetailView.svelte"
+
+  # Grid styling predicate over `config.itemType`, a value the page already
+  # declares about itself. Selects a *look*, issues no query, and would only
+  # change if the UI were redesigned — presentation, not taxonomy-as-policy.
+  "src/lib/components/library/GenericMediaListPage.svelte"
+
+  # "Is this item a container?" predicate for downloads browsing.
+  # BORDERLINE — leans domain: the container set grows when Jellyfin adds a
+  # container type. TODO: replace with a backend-supplied `MediaItem.isContainer`
+  # flag and remove this entry. Tracked in
+  # docs/specs/boundary-tripwire-hardening.md §Out of scope.
+  "src/lib/components/downloads/DownloadedBrowse.svelte"
 )
 
+# Hard cap so erosion is caught mechanically rather than by whoever notices.
+# Deliberately just above the current count: the next exception forces a
+# conversation instead of a one-line append.
+MAX_ALLOWLIST=4
+
+if [[ "${#ALLOWLIST[@]}" -gt "$MAX_ALLOWLIST" ]]; then
+  echo "❌ Allowlist has ${#ALLOWLIST[@]} entries (max $MAX_ALLOWLIST)."
+  echo "   Push taxonomy into Rust instead of appending here."
+  exit 1
+fi
+
 is_allowed() {
   local file="$1"
   for allowed in "${ALLOWLIST[@]}"; do
@@ -46,11 +82,28 @@ is_allowed() {
   return 1
 }
 
-# Multi-element includeItemTypes array: `includeItemTypes: [ <x> , <y> ... ]`.
-# The comma inside the brackets is what makes it multi-type.
-PATTERN='includeItemTypes:[[:space:]]*\[[^]]*,[^]]*\]'
+# Two or more adjacent Jellyfin item-type string literals inside a bracket.
+#
+# NOT anchored to `includeItemTypes:` — that was the original rule, and it missed
+# the real leak: `searchScope.ts` assigned the same array to a named const and
+# dereferenced it one indirection away from the query, so the grep never saw it
+# while CI stayed green. Matching the array literal itself catches a const, a
+# Record value, a function return, and an inline query alike.
+#
+# Deliberate limits:
+#   - requires TWO adjacent types, so single-type presentation
+#     (`itemType: "Movie"`) stays legal — the rule targets *category* taxonomy;
+#   - requires string literals, so `item.type === "Audio"` (display inspection)
+#     does not match;
+#   - uses an explicit type list rather than a generic capitalised-word pattern,
+#     so unrelated string arrays (`["High","Low"]`) produce no noise.
+#
+# An item type missing from this list is a false *negative*, never a false
+# positive — the check degrades safely as Jellyfin adds types.
+ITEM_TYPES='Movie|Series|Episode|Audio|MusicAlbum|MusicArtist|MusicVideo|Season|BoxSet|Playlist|Book|AudioBook|Video|Person|Folder|CollectionFolder|TvChannel|LiveTvChannel'
+PATTERN="\[[[:space:]]*\"($ITEM_TYPES)\"[[:space:]]*,[[:space:]]*\"($ITEM_TYPES)\""
 
-echo "🔎 Checking frontend for domain-taxonomy leaks (multi-type query arrays)…"
+echo "🔎 Checking frontend for domain-taxonomy leaks (item-type array literals)…"
 
 # Collect hits, excluding tests and the allowlist.
 violations=""
@@ -69,9 +122,11 @@ done < <(grep -rInE "$PATTERN" src/ 2>/dev/null || true)
 
 if [[ -n "$violations" ]]; then
   echo ""
-  echo "❌ Frontend boundary violation: a multi-type includeItemTypes query defines"
-  echo "   a category in the presentation layer. That taxonomy belongs in Rust —"
-  echo "   send an opaque scope and let the backend expand it to item types."
+  echo "❌ Frontend boundary violation: an item-type array literal defines a"
+  echo "   category in the presentation layer. That taxonomy belongs in Rust —"
+  echo "   send an opaque scope/enum and let the backend expand it to item types"
+  echo "   (see SearchScope::item_types() in src-tauri/src/repository/types.rs)."
+  echo "   Assigning the array to a const does not make it presentation."
   echo "   See docs/specs/scoped-search-boundary.md and CLAUDE.md."
   echo ""
   echo "$violations" | sed 's/^/   /'

scripts/check-req-coverage.sh

@@ -1,84 +0,0 @@
-#!/bin/bash
-#
-# Requirements Coverage Checker
-# Extracts @req tags from codebase and compares with README.md
-#
-
-set -e
-
-REQUIREMENTS_FILE="README.md"
-SOURCE_DIRS="src-tauri/ src/"
-
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-echo "  Requirements Coverage Report"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-echo ""
-
-# Extract requirement IDs from README.md (UR-, IR-, DR-, JA-)
-echo "📊 Scanning requirements from $REQUIREMENTS_FILE..."
-requirements=$(grep -E "^\| (UR|IR|DR|JA)-[0-9]+" "$REQUIREMENTS_FILE" | \
-               sed -E 's/^\| ([A-Z]+-[0-9]+).*/\1/' | \
-               sort -u)
-
-total_reqs=$(echo "$requirements" | wc -l)
-implemented=0
-partial=0
-planned=0
-missing=0
-
-echo ""
-echo "Category Breakdown:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-
-for category in UR IR DR JA; do
-    cat_count=$(echo "$requirements" | grep "^$category-" | wc -l)
-    printf "%-4s %3d requirements\n" "$category:" "$cat_count"
-done
-
-echo ""
-echo "Implementation Status:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-
-for req in $requirements; do
-    # Count full implementations
-    full_count=$(grep -r "@req: $req" $SOURCE_DIRS 2>/dev/null | grep -v "@req-partial" | grep -v "@req-planned" | wc -l)
-
-    # Count partial implementations
-    partial_count=$(grep -r "@req-partial: $req" $SOURCE_DIRS 2>/dev/null | wc -l)
-
-    # Count planned
-    planned_count=$(grep -r "@req-planned: $req" $SOURCE_DIRS 2>/dev/null | wc -l)
-
-    if [ "$full_count" -gt 0 ]; then
-        echo "✅ $req: $full_count implementation(s)"
-        ((implemented++))
-    elif [ "$partial_count" -gt 0 ]; then
-        echo "🔶 $req: $partial_count partial implementation(s)"
-        ((partial++))
-    elif [ "$planned_count" -gt 0 ]; then
-        echo "📋 $req: Planned (not yet implemented)"
-        ((planned++))
-    else
-        echo "❌ $req: No implementation found"
-        ((missing++))
-    fi
-done
-
-echo ""
-echo "Summary:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-printf "Total Requirements:     %3d\n" "$total_reqs"
-printf "✅ Fully Implemented:   %3d (%.0f%%)\n" "$implemented" "$(echo "scale=0; $implemented * 100 / $total_reqs" | bc)"
-printf "🔶 Partially Implemented: %3d (%.0f%%)\n" "$partial" "$(echo "scale=0; $partial * 100 / $total_reqs" | bc)"
-printf "📋 Planned:             %3d (%.0f%%)\n" "$planned" "$(echo "scale=0; $planned * 100 / $total_reqs" | bc)"
-printf "❌ Missing:             %3d (%.0f%%)\n" "$missing" "$(echo "scale=0; $missing * 100 / $total_reqs" | bc)"
-echo ""
-
-# Exit code based on missing critical requirements
-if [ "$missing" -gt 0 ]; then
-    echo "⚠️  Warning: $missing requirements have no implementation"
-    exit 1
-else
-    echo "✨ All requirements have implementations!"
-    exit 0
-fi

scripts/check-test-coverage.sh

@@ -1,40 +0,0 @@
-#!/bin/bash
-#
-# Test Coverage Report
-# Links test requirements to implementations
-#
-
-echo "Test Coverage Report"
-echo "===================="
-echo ""
-
-test_reqs=$(grep -rh "@req-test:" src-tauri/ 2>/dev/null | \
-            sed 's/.*@req-test: \([A-Z][A-Z]-[0-9]*\).*/\1/' | \
-            sort -u)
-
-total_tests=0
-covered=0
-uncovered=0
-
-for req in $test_reqs; do
-    test_count=$(grep -r "@req-test: $req" src-tauri/ 2>/dev/null | wc -l)
-    impl_count=$(grep -r "@req: $req" src-tauri/ src/ 2>/dev/null | wc -l)
-
-    ((total_tests++))
-
-    if [ "$test_count" -gt 0 ] && [ "$impl_count" -gt 0 ]; then
-        echo "✅ $req: $test_count test(s), $impl_count implementation(s)"
-        ((covered++))
-    elif [ "$impl_count" -eq 0 ]; then
-        echo "⚠️  $req: $test_count test(s) but no implementation"
-        ((uncovered++))
-    fi
-done
-
-echo ""
-echo "Summary:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-printf "Total Test Requirements: %3d\n" "$total_tests"
-printf "✅ With Implementation:  %3d (%.0f%%)\n" "$covered" "$(echo "scale=0; $covered * 100 / $total_tests" | bc)"
-printf "⚠️  No Implementation:   %3d (%.0f%%)\n" "$uncovered" "$(echo "scale=0; $uncovered * 100 / $total_tests" | bc)"
-echo ""

scripts/deploy-android.sh

@@ -13,25 +13,60 @@ if ! adb devices | grep -q "device$"; then
     exit 1
 fi
 
-# Build type: debug or release (default: debug)
-BUILD_TYPE="${1:-debug}"
+# Build type: debug or release (default: debug). `--debug` alongside `release`
+# means the side-by-side release build — same APK path, but it was packaged
+# under the .debug applicationId, so the package to launch differs.
+BUILD_TYPE="debug"
+SIDE_BY_SIDE=0
+for arg in "$@"; do
+    case "$arg" in
+        --debug|--side-by-side) SIDE_BY_SIDE=1 ;;
+        debug|release) BUILD_TYPE="$arg" ;;
+    esac
+done
+[ "$BUILD_TYPE" = "debug" ] && SIDE_BY_SIDE=1
 
+# The .debug applicationId (see src-tauri/android/app/build.gradle.kts) is a
+# separate package, so it installs alongside a real release build — no
+# uninstall dance needed.
 if [ "$BUILD_TYPE" = "release" ]; then
     APK_PATH="src-tauri/gen/android/app/build/outputs/apk/universal/release/app-universal-release.apk"
 else
     APK_PATH="src-tauri/gen/android/app/build/outputs/apk/universal/debug/app-universal-debug.apk"
 fi
 
+if [ "$SIDE_BY_SIDE" = "1" ]; then
+    APP_PACKAGE="com.dtourolle.jellytau.debug"
+else
+    APP_PACKAGE="com.dtourolle.jellytau"
+fi
+
 # Check if APK exists
 if [ ! -f "$APK_PATH" ]; then
     echo "❌ APK not found at: $APK_PATH"
-    echo "Run './scripts/build-android.sh $BUILD_TYPE' first"
+    if [ "$BUILD_TYPE" = "release" ] && [ "$SIDE_BY_SIDE" = "1" ]; then
+        echo "Run './scripts/build-android.sh release --debug' first"
+    else
+        echo "Run './scripts/build-android.sh $BUILD_TYPE' first"
+    fi
     exit 1
 fi
 
 echo "📦 Installing APK: $APK_PATH"
-adb install -r "$APK_PATH"
+echo "📛 Package: $APP_PACKAGE"
+
+if ! adb install -r "$APK_PATH"; then
+    echo ""
+    echo "❌ Install failed."
+    echo "   If it says INSTALL_FAILED_UPDATE_INCOMPATIBLE, an older build of"
+    echo "   '$APP_PACKAGE' signed with a different key is still installed."
+    echo "   Uninstall just that one and retry:"
+    echo "     adb uninstall $APP_PACKAGE"
+    exit 1
+fi
 
 echo ""
 echo "✅ Deployment complete!"
-echo "🚀 Launch the app on your device"
+echo "🚀 Launching..."
+adb shell monkey -p "$APP_PACKAGE" -c android.intent.category.LAUNCHER 1 > /dev/null 2>&1 \
+    || echo "   (auto-launch failed — start it from the launcher)"

scripts/extract-traces.test.ts

@@ -0,0 +1,182 @@
+/**
+ * Tests for the traceability coverage computation.
+ *
+ * These run over fixture strings rather than the live docs/requirements.md, so
+ * their meaning does not drift as requirements are added.
+ *
+ * Background: the CI gate divided traced-requirement counts by hardcoded
+ * denominators (UR/39, IR/24, DR/48, JA/3, total 114) that had fallen out of
+ * date, reporting 158% coverage and making the 50% threshold unreachable. These
+ * tests pin the parsing and arithmetic that replace those literals.
+ *
+ * @req-test: UT-089 - Requirement definitions parsed from requirements.md
+ * @req-test: UT-090 - Coverage is the intersection of traced and defined IDs
+ */
+
+import { describe, it, expect } from "vitest";
+import { countDefinedRequirements, computeCoverage } from "./extract-traces";
+
+describe("countDefinedRequirements", () => {
+  it("counts a well-formed table row as a defined requirement", () => {
+    const md = `
+| ID | Requirement | Priority | Status |
+|----|-------------|----------|--------|
+| UR-001 | Run the app on multiple platforms | High | In Progress |
+| UR-002 | Access media when online or offline | High | Done |
+`;
+    const defined = countDefinedRequirements(md);
+    expect(defined.UR).toBe(2);
+    expect(defined.DR).toBe(0);
+  });
+
+  it("does not count IDs that appear only in the Traces To column", () => {
+    // The bug this rule avoids: a naive grep for /DR-\d{3}/ over the whole file
+    // counts DR-001 here as "defined", inflating the denominator with IDs that
+    // are merely referenced.
+    const md = `
+| DR-001 | Player state machine | Player | UR-005 | Done |
+| DR-002 | MediaItem struct | Player | UR-003, UR-004 | Done |
+`;
+    const defined = countDefinedRequirements(md);
+    expect(defined.DR).toBe(2);
+    // UR-005/UR-003/UR-004 are referenced, never defined here.
+    expect(defined.UR).toBe(0);
+  });
+
+  it("does not count IDs mentioned in prose", () => {
+    const md = `
+Some prose explaining that UR-005 relates to DR-001 and JA-002.
+
+| UR-005 | Control media playback | High | Done |
+`;
+    const defined = countDefinedRequirements(md);
+    expect(defined.UR).toBe(1);
+    expect(defined.DR).toBe(0);
+    expect(defined.JA).toBe(0);
+  });
+
+  it("deduplicates an ID listed in both the spec table and the traceability matrix", () => {
+    // requirements.md lists every UR twice: once in §1 (definition) and again in
+    // §3 (traceability matrix), both as a leading table cell. Counting rows
+    // instead of unique IDs double-counts the UR denominator (121 vs 61).
+    const md = `
+| UR-005 | Control media playback | High | Done |
+| UR-006 | Browse the library | High | Done |
+
+### Traceability Matrix
+
+| UR-005 | - | DR-001, DR-005, DR-009 |
+| UR-006 | - | DR-012 |
+`;
+    const defined = countDefinedRequirements(md);
+    expect(defined.UR).toBe(2);
+  });
+
+  it("collects the defined ID set, not just counts", () => {
+    const md = `
+| UR-001 | A | High | Done |
+| DR-050 | B | Player | UR-001 | Done |
+`;
+    const defined = countDefinedRequirements(md);
+    expect(defined.ids.has("UR-001")).toBe(true);
+    expect(defined.ids.has("DR-050")).toBe(true);
+    expect(defined.ids.has("UR-999")).toBe(false);
+  });
+});
+
+describe("computeCoverage", () => {
+  const defined = {
+    UR: 2,
+    IR: 0,
+    DR: 2,
+    JA: 0,
+    total: 4,
+    ids: new Set(["UR-001", "UR-002", "DR-001", "DR-002"]),
+  };
+
+  it("computes coverage as traced ∩ defined over defined", () => {
+    const traced = ["UR-001", "DR-001"];
+    const cov = computeCoverage(traced, defined);
+    expect(cov.covered).toBe(2);
+    expect(cov.total).toBe(4);
+    expect(cov.percent).toBe(50);
+  });
+
+  it("does not let a traced-but-undefined ID inflate the numerator", () => {
+    // This is how a ratio exceeds 100%: a TRACES comment naming a typo'd or
+    // deleted requirement counted as covered.
+    const traced = ["UR-001", "DR-001", "DR-097"];
+    const cov = computeCoverage(traced, defined);
+    expect(cov.covered).toBe(2);
+    expect(cov.percent).toBe(50);
+  });
+
+  it("reports traced-but-undefined IDs as orphaned so they get fixed", () => {
+    const traced = ["UR-001", "DR-097", "JA-404"];
+    const cov = computeCoverage(traced, defined);
+    expect(cov.orphaned).toEqual(["DR-097", "JA-404"]);
+  });
+
+  it("has no orphans when every traced ID is defined", () => {
+    const cov = computeCoverage(["UR-001", "UR-002"], defined);
+    expect(cov.orphaned).toEqual([]);
+  });
+
+  it("ignores UT/IT test IDs entirely — they are a separate taxonomy", () => {
+    // UT/IT are defined in §4 of requirements.md, not among the four
+    // requirement types. Treating them as orphans buries real typos in ~60
+    // lines of noise, and counting them would corrupt the ratio.
+    const cov = computeCoverage(["UR-001", "UT-088", "IT-017"], defined);
+    expect(cov.orphaned).toEqual([]);
+    expect(cov.covered).toBe(1);
+  });
+
+  it("reports 0% rather than dividing by zero for an empty trace set", () => {
+    const cov = computeCoverage([], defined);
+    expect(cov.covered).toBe(0);
+    expect(cov.percent).toBe(0);
+  });
+
+  it("reports 0% rather than NaN when nothing is defined", () => {
+    const empty = { UR: 0, IR: 0, DR: 0, JA: 0, total: 0, ids: new Set<string>() };
+    const cov = computeCoverage([], empty);
+    expect(cov.percent).toBe(0);
+    expect(Number.isNaN(cov.percent)).toBe(false);
+  });
+
+  it("reports exactly 100% when all defined requirements are traced, never above", () => {
+    const traced = ["UR-001", "UR-002", "DR-001", "DR-002"];
+    const cov = computeCoverage(traced, defined);
+    expect(cov.percent).toBe(100);
+  });
+
+  it("ignores duplicate traced IDs", () => {
+    const traced = ["UR-001", "UR-001", "UR-001"];
+    const cov = computeCoverage(traced, defined);
+    expect(cov.covered).toBe(1);
+  });
+});
+
+describe("live requirements.md", () => {
+  it("parses the real file to the counts the CI gate must use", () => {
+    // Guards the specific regression: CI hardcoded UR/39, IR/24, DR/48, JA/3
+    // (total 114) while the real file had grown to 211. Update these numbers
+    // deliberately when requirements are added — that edit is the signal the
+    // denominator is live rather than frozen.
+    const fs = require("fs");
+    const path = require("path");
+    // import.meta.dir is Bun-only; derive from import.meta.url under vitest.
+    const here = path.dirname(new URL(import.meta.url).pathname);
+    const md = fs.readFileSync(
+      path.resolve(here, "../docs/requirements.md"),
+      "utf-8"
+    );
+    const defined = countDefinedRequirements(md);
+
+    expect(defined.UR).toBe(75);
+    expect(defined.IR).toBe(32);
+    expect(defined.DR).toBe(185);
+    expect(defined.JA).toBe(35);
+    expect(defined.total).toBe(327);
+  });
+});

scripts/extract-traces.ts

@@ -34,11 +34,20 @@ interface TracesData {
     DR: string[];
     JA: string[];
   };
+  /** Requirements *defined* in requirements.md — the coverage denominators. */
+  defined?: { UR: number; IR: number; DR: number; JA: number; total: number };
+  coverage?: CoverageResult;
 }
 
 // Repo root, derived from this script's location (scripts/ -> repo root).
 // Must NOT be hardcoded to a developer's machine, or CI checkouts see no files.
-const BASE_DIR = path.resolve(import.meta.dir, "..");
+//
+// `import.meta.dir` is a Bun extension and is undefined when this module is
+// imported by vitest (which runs it as an ordinary ESM module), so fall back to
+// import.meta.url — this file must stay importable for extract-traces.test.ts.
+const SCRIPT_DIR =
+  import.meta.dir ?? path.dirname(new URL(import.meta.url).pathname);
+const BASE_DIR = path.resolve(SCRIPT_DIR, "..");
 
 const TRACES_PATTERN = /TRACES:\s*([^\n]+)/gi;
 const REQ_ID_PATTERN = /([A-Z]{2})-(\d{3})/g;
@@ -50,7 +59,10 @@ function extractRequirementIds(tracesString: string): string[] {
 
 function getAllSourceFiles(): string[] {
   const baseDir = BASE_DIR;
-  const patterns = ["src", "src-tauri/src"];
+  // `scripts` is scanned too: build tooling implements requirements (e.g.
+  // DR-093, the coverage engine itself) and would otherwise be invisible to the
+  // very matrix it generates.
+  const patterns = ["src", "src-tauri/src", "scripts"];
   const files: string[] = [];
 
   function walkDir(dir: string) {
@@ -192,6 +204,109 @@ function extractTraces(): TracesData {
   };
 }
 
+// ---------------------------------------------------------------------------
+// Coverage: how many *defined* requirements are actually traced.
+//
+// The denominators MUST be derived from requirements.md, never hardcoded. The
+// CI gate previously divided by frozen literals (UR/39, IR/24, DR/48, JA/3,
+// total 114) while the real file had grown to 211 requirements, so it reported
+// 158% coverage and the 50% threshold became unreachable — the gate could not
+// fail. See docs/specs/traceability-gate-repair.md.
+//
+// TRACES: | DR-093
+// ---------------------------------------------------------------------------
+
+export interface DefinedRequirements {
+  UR: number;
+  IR: number;
+  DR: number;
+  JA: number;
+  total: number;
+  ids: Set<string>;
+}
+
+export interface CoverageResult {
+  covered: number;
+  total: number;
+  percent: number;
+  /** Traced in code but not defined in requirements.md (typo, or deleted req). */
+  orphaned: string[];
+}
+
+/**
+ * A requirement is *defined* only where its ID is the leading cell of a markdown
+ * table row: `| DR-001 | … |`.
+ *
+ * This deliberately ignores IDs in the "Traces To" column and in prose — a
+ * naive scan for /DR-\d{3}/ counts those as definitions and inflates the
+ * denominator. IDs are deduplicated because requirements.md lists each UR twice
+ * (once in §1 as a definition, again in §3's traceability matrix), which would
+ * otherwise double the UR count from 61 to 121.
+ *
+ * TRACES: | DR-093
+ */
+export function countDefinedRequirements(markdown: string): DefinedRequirements {
+  const ids = new Set<string>();
+  const ROW_ID = /^\|\s*(UR|IR|DR|JA)-(\d{3})\s*\|/;
+
+  for (const line of markdown.split("\n")) {
+    const match = line.match(ROW_ID);
+    if (match) ids.add(`${match[1]}-${match[2]}`);
+  }
+
+  const countOf = (type: string) =>
+    [...ids].filter((id) => id.startsWith(`${type}-`)).length;
+
+  return {
+    UR: countOf("UR"),
+    IR: countOf("IR"),
+    DR: countOf("DR"),
+    JA: countOf("JA"),
+    total: ids.size,
+    ids,
+  };
+}
+
+/**
+ * Coverage is the *intersection* of traced and defined IDs over defined IDs.
+ *
+ * Using the raw traced count as the numerator is what lets a ratio exceed 100%:
+ * a TRACES comment naming a requirement that no longer exists would count as
+ * covered. Those IDs are reported as `orphaned` so they get fixed rather than
+ * silently counted or silently dropped.
+ *
+ * TRACES: | DR-093
+ */
+export function computeCoverage(
+  tracedIds: string[],
+  defined: DefinedRequirements
+): CoverageResult {
+  // Only the four *requirement* types participate in coverage. UT/IT are test
+  // identifiers defined in §4 of requirements.md — a different taxonomy, and
+  // flagging them as orphans would bury real typos in ~60 lines of noise.
+  const isRequirement = (id: string) => /^(UR|IR|DR|JA)-\d{3}$/.test(id);
+
+  const traced = new Set(tracedIds.filter(isRequirement));
+  const covered = [...traced].filter((id) => defined.ids.has(id));
+  const orphaned = [...traced].filter((id) => !defined.ids.has(id)).sort();
+
+  return {
+    covered: covered.length,
+    total: defined.total,
+    percent:
+      defined.total === 0
+        ? 0
+        : Math.round((covered.length / defined.total) * 100),
+    orphaned,
+  };
+}
+
+/** Read requirements.md from the repo and count what it defines. */
+export function readDefinedRequirements(): DefinedRequirements {
+  const reqPath = path.join(BASE_DIR, "docs", "requirements.md");
+  return countDefinedRequirements(fs.readFileSync(reqPath, "utf-8"));
+}
+
 function generateMarkdown(data: TracesData): string {
   let md = `# Code Traceability Matrix
 
@@ -265,21 +380,83 @@ function generateJson(data: TracesData): string {
   return JSON.stringify(data, null, 2);
 }
 
-// Main
-const args = Bun.argv.slice(2);
-const format = args.includes("--format")
-  ? args[args.indexOf("--format") + 1]
-  : "markdown";
+/**
+ * Human-readable coverage report; exits non-zero below the threshold so this is
+ * runnable as a local gate (`bun run traces:coverage`), not just in CI.
+ *
+ * TRACES: | DR-093
+ */
+function reportCoverage(data: TracesData, minThreshold: number): number {
+  const defined = data.defined!;
+  const cov = data.coverage!;
 
-console.error("🔍 Extracting TRACES from codebase...");
-const data = extractTraces();
+  const definedIds = readDefinedRequirements().ids;
 
-if (format === "json") {
-  console.log(generateJson(data));
-} else {
-  console.log(generateMarkdown(data));
+  console.log("📋 Requirement coverage (traced / defined):");
+  for (const type of ["UR", "IR", "DR", "JA"] as const) {
+    const traced = data.byType[type].filter((id) => definedIds.has(id)).length;
+    console.log(`   ${type}: ${traced} / ${defined[type]}`);
+  }
+  console.log("");
+  console.log(`📈 Overall: ${cov.covered} / ${cov.total} (${cov.percent}%)`);
+
+  if (cov.orphaned.length > 0) {
+    console.log("");
+    console.log(
+      `⚠️  Traced but not defined in requirements.md: ${cov.orphaned.join(", ")}`
+    );
+    console.log("   Fix the TRACES comment or add the requirement.");
+  }
+
+  // A ratio above 100% means the computation is broken (the condition that hid
+  // the stale-denominator bug for so long). Fail loudly rather than report it.
+  if (cov.percent > 100) {
+    console.log("");
+    console.log(`❌ Coverage > 100% — the gate is miscomputing.`);
+    return 1;
+  }
+
+  if (cov.percent < minThreshold) {
+    console.log("");
+    console.log(`❌ Coverage (${cov.percent}%) is below minimum (${minThreshold}%)`);
+    return 1;
+  }
+
+  console.log("");
+  console.log(`✅ Coverage is acceptable (${cov.percent}% >= ${minThreshold}%)`);
+  return 0;
 }
 
-console.error(
-  `\n✅ Complete! Found ${data.totalTraces} TRACES across ${data.totalFiles} files`
-);
+// Main — guarded so this module stays importable from extract-traces.test.ts.
+if (import.meta.main) {
+  const args = process.argv.slice(2);
+  const format = args.includes("--format")
+    ? args[args.indexOf("--format") + 1]
+    : "markdown";
+
+  console.error("🔍 Extracting TRACES from codebase...");
+  const data = extractTraces();
+
+  const defined = readDefinedRequirements();
+  const allTraced = Object.keys(data.requirements);
+  data.defined = {
+    UR: defined.UR,
+    IR: defined.IR,
+    DR: defined.DR,
+    JA: defined.JA,
+    total: defined.total,
+  };
+  data.coverage = computeCoverage(allTraced, defined);
+
+  if (format === "json") {
+    console.log(generateJson(data));
+  } else if (format === "coverage") {
+    process.exit(reportCoverage(data, 50));
+  } else {
+    console.log(generateMarkdown(data));
+  }
+
+  console.error(
+    `\n✅ Complete! Found ${data.totalTraces} TRACES across ${data.totalFiles} files`
+  );
+}

scripts/find-req-implementations.sh

@@ -1,56 +0,0 @@
-#!/bin/bash
-#
-# Find all files implementing a specific requirement
-#
-# Usage: ./find-req-implementations.sh UR-004
-#
-
-if [ $# -eq 0 ]; then
-    echo "Usage: $0 <REQUIREMENT_ID>"
-    echo "Example: $0 UR-004"
-    exit 1
-fi
-
-REQ_ID=$1
-
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-echo "  Implementations of $REQ_ID"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-echo ""
-
-# Full implementations
-echo "Full Implementations:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-grep -rn "@req: $REQ_ID" src-tauri/ src/ 2>/dev/null | \
-    grep -v "@req-partial" | \
-    grep -v "@req-planned" | \
-    sed 's/src-tauri\/src\///' | \
-    sed 's/src\///' || echo "  (none)"
-
-echo ""
-
-# Partial implementations
-echo "Partial Implementations:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-grep -rn "@req-partial: $REQ_ID" src-tauri/ src/ 2>/dev/null | \
-    sed 's/src-tauri\/src\///' | \
-    sed 's/src\///' || echo "  (none)"
-
-echo ""
-
-# Planned
-echo "Planned Implementations:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-grep -rn "@req-planned: $REQ_ID" src-tauri/ src/ 2>/dev/null | \
-    sed 's/src-tauri\/src\///' | \
-    sed 's/src\///' || echo "  (none)"
-
-echo ""
-
-# Tests
-echo "Test Cases:"
-echo "━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━"
-grep -rn "@req-test: $REQ_ID" src-tauri/ 2>/dev/null | \
-    sed 's/src-tauri\/src\///' || echo "  (none)"
-
-echo ""

scripts/logcat.sh

@@ -1,13 +1,34 @@
 #!/bin/bash
-# View Android logcat output filtered for the app
+# View Android logcat output filtered for the app.
+#
+# Usage: ./scripts/logcat.sh [debug|release]   (default: debug)
+#
+# The debug build has applicationIdSuffix ".debug" so it can be installed
+# alongside a release build; pick the package to follow accordingly.
 
 set -e
 
-APP_PACKAGE="com.jellytau.app"
+BUILD_TYPE="${1:-debug}"
+
+if [ "$BUILD_TYPE" = "release" ]; then
+    APP_PACKAGE="com.dtourolle.jellytau"
+else
+    APP_PACKAGE="com.dtourolle.jellytau.debug"
+fi
 
 echo "📱 Showing logcat for $APP_PACKAGE"
 echo "Press Ctrl+C to stop"
 echo ""
 
-# Filter logcat for the app's package name
-adb logcat | grep -i "$APP_PACKAGE\|tauri\|rust"
+# Prefer PID-scoped output when the app is running — it drops the noise that a
+# text grep can't. Fall back to the old keyword filter when it isn't (so you can
+# start the script first and then launch the app).
+PID="$(adb shell pidof "$APP_PACKAGE" 2>/dev/null | tr -d '\r\n' | awk '{print $1}')"
+
+if [ -n "$PID" ]; then
+    echo "   (attached to pid $PID)"
+    adb logcat --pid="$PID"
+else
+    echo "   (app not running — falling back to keyword filter)"
+    adb logcat | grep -i "$APP_PACKAGE\|jellytau\|tauri\|rust"
+fi

scripts/set-version.sh

@@ -0,0 +1,119 @@
+#!/bin/bash
+# Stamp the release version into every file that carries it.
+#
+# The git tag is the single source of truth for a release version. The versions
+# committed in package.json / tauri.conf.json / Cargo.toml are a placeholder for
+# dev builds; a tagged build overwrites all of them from the tag so they cannot
+# disagree with each other or with the tag.
+#
+# Usage:
+#   ./scripts/set-version.sh 0.5.0        # explicit
+#   JELLYTAU_VERSION=0.5.0 ./scripts/set-version.sh
+#   ./scripts/set-version.sh              # derive from git describe (dev builds)
+#
+# Accepts the version with or without a leading "v".
+#
+# Why a script and not four sed lines in CI: the version lived in four files and
+# CI only ever rewrote one of them (tauri.conf.json), so a tagged release shipped
+# a matching installer name and mismatched package metadata. Keeping the write in
+# one place is what makes "the tag is authoritative" actually true.
+
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+VERSION="${1:-${JELLYTAU_VERSION:-}}"
+
+# CI passes "${GITHUB_REF#refs/tags/}" unconditionally, which on an untagged
+# build is still a full ref ("refs/heads/master"). Treat anything that is not a
+# bare version as "no version given" and fall through to git describe, so a
+# branch build gets a sane dev version instead of failing the job.
+case "$VERSION" in
+  refs/*) VERSION="" ;;
+esac
+
+if [ -z "$VERSION" ]; then
+  # No explicit version: derive from the most recent tag. Dev builds land on
+  # something like 0.5.0 (exact tag) or 0.5.0-3-gabc1234 (ahead of the tag).
+  VERSION="$(git describe --tags --always --match 'v*' 2>/dev/null || echo "0.0.0")"
+fi
+
+# Tags are written v0.5.0; the files carry a bare semver.
+VERSION="${VERSION#v}"
+
+# Validate before writing anything — a malformed version silently propagated
+# into four files is far worse than a failed script.
+if ! echo "$VERSION" | grep -qE '^[0-9]+\.[0-9]+\.[0-9]+([-+][0-9A-Za-z.-]+)?$'; then
+  echo "❌ Not a valid semver: '$VERSION'" >&2
+  echo "   Expected MAJOR.MINOR.PATCH with an optional -prerelease/+build suffix." >&2
+  exit 1
+fi
+
+echo "📌 Setting version to $VERSION"
+
+# --- The three committed manifests -----------------------------------------
+# Anchored to the first "version" key so a dependency's version is never hit.
+
+# package.json — the top-level "version", which sits in the first few lines.
+perl -0pi -e 's/("version"\s*:\s*)"[^"]*"/$1"'"$VERSION"'"/' package.json
+
+# tauri.conf.json — likewise; this is the one the bundler reads for installer
+# names, and the one CI used to patch alone.
+perl -0pi -e 's/("version"\s*:\s*)"[^"]*"/$1"'"$VERSION"'"/' src-tauri/tauri.conf.json
+
+# Cargo.toml — only the [package] version, never a dependency's. Restricted to
+# the first occurrence of a line-anchored `version = "..."`.
+perl -0pi -e 's/^(version\s*=\s*)"[^"]*"/$1"'"$VERSION"'"/m' src-tauri/Cargo.toml
+
+# Cargo.lock — the jellytau entry. Left alone if the lock has not been generated
+# yet; the next cargo invocation writes it. Cargo would otherwise rewrite the
+# lock mid-build and dirty the tree.
+if [ -f src-tauri/Cargo.lock ]; then
+  perl -0pi -e 's/(name = "jellytau"\nversion = )"[^"]*"/$1"'"$VERSION"'"/' src-tauri/Cargo.lock
+fi
+
+# --- Android versionCode ----------------------------------------------------
+# Only when the generated Android project exists (i.e. after `tauri android
+# init`); on Linux/Windows jobs there is nothing to stamp.
+#
+# `tauri android init` derives a versionCode from the semver (0.0.15 -> 15).
+# That is both tiny and NOT monotonic across our history: earlier local/dev
+# builds shipped versionCode 1000 (from a 0.1.0 config), so a plain 15 is a
+# *downgrade* and Android refuses the update.
+#
+# The floor has to clear the highest code actually in the field, which is not the
+# same as the highest this formula has produced. v0.5.2 shipped versionCode
+# **5002** under an earlier `minor*1000` scheme; the `minor*100` formula that
+# replaced it yields only 1502 for that same version, and 1503 for 0.5.3 — so
+# every 0.5.x release built from it was an un-installable downgrade for anyone
+# already on v0.5.2, which is exactly the failure this block exists to prevent.
+# The multipliers are widened and the floor raised past 5002 accordingly.
+#
+#   code = 10000 + major*1000000 + minor*1000 + patch
+# e.g. 0.0.14 -> 10014, 0.1.0 -> 11000, 0.5.3 -> 15003, 1.0.0 -> 1010000.
+PROPS="src-tauri/gen/android/app/tauri.properties"
+if [ -f "$PROPS" ]; then
+  # Strip any -rc1/+build suffix first: it is not numeric, and feeding it to
+  # $(( )) would abort the script under `set -e`.
+  CORE="${VERSION%%-*}"
+  CORE="${CORE%%+*}"
+  MAJ=$(echo "$CORE" | cut -d. -f1)
+  MIN=$(echo "$CORE" | cut -d. -f2)
+  PAT=$(echo "$CORE" | cut -d. -f3)
+  : "${MAJ:=0}" "${MIN:=0}" "${PAT:=0}"
+  CODE=$(( 10000 + MAJ*1000000 + MIN*1000 + PAT ))
+  echo "   versionCode=$CODE (from $CORE)"
+  if grep -q '^tauri.android.versionCode=' "$PROPS"; then
+    sed -i "s/^tauri.android.versionCode=.*/tauri.android.versionCode=$CODE/" "$PROPS"
+  else
+    echo "tauri.android.versionCode=$CODE" >> "$PROPS"
+  fi
+fi
+
+# --- Report -----------------------------------------------------------------
+echo "✅ Version stamped:"
+grep -m1 '"version"' package.json | sed 's/^/   package.json:      /'
+grep -m1 '"version"' src-tauri/tauri.conf.json | sed 's/^/   tauri.conf.json:   /'
+grep -m1 '^version' src-tauri/Cargo.toml | sed 's/^/   Cargo.toml:        /'
+[ -f "$PROPS" ] && grep '^tauri.android.versionCode=' "$PROPS" | sed 's/^/   tauri.properties:  /'
+exit 0

scripts/set-version.test.ts

@@ -0,0 +1,181 @@
+/**
+ * Guards for scripts/set-version.sh — the release version stamper.
+ *
+ * TRACES: DR-153 | UT-150
+ *
+ * These run the real script against a throwaway copy of the manifests, because
+ * the failure modes are all in the shell, not in any TS logic: a regex that also
+ * matches a dependency's version, arithmetic that aborts on a `-rc1` suffix, or
+ * a CI ref reaching the validator verbatim.
+ *
+ * The versionCode formula matters most. Android refuses an update whose code is
+ * lower than the installed one, and builds already in the field shipped code
+ * 1000 — so any formula that can emit a smaller number for a *newer* release
+ * bricks updates for those users, silently and irreversibly.
+ */
+
+import { describe, expect, it, beforeEach, afterEach } from "vitest";
+import { execFileSync } from "child_process";
+import * as fs from "fs";
+import * as path from "path";
+import * as os from "os";
+
+const repoRoot = path.resolve(path.dirname(new URL(import.meta.url).pathname), "..");
+const script = path.join(repoRoot, "scripts", "set-version.sh");
+
+let tmp: string;
+
+/** A minimal repo skeleton: just the files the script rewrites. */
+function seed(dir: string) {
+  fs.mkdirSync(path.join(dir, "src-tauri", "gen", "android", "app"), { recursive: true });
+  fs.mkdirSync(path.join(dir, "scripts"), { recursive: true });
+  fs.copyFileSync(script, path.join(dir, "scripts", "set-version.sh"));
+  fs.chmodSync(path.join(dir, "scripts", "set-version.sh"), 0o755);
+
+  fs.writeFileSync(
+    path.join(dir, "package.json"),
+    JSON.stringify({ name: "jellytau", version: "0.0.1", dependencies: { hls: "1.2.3" } }, null, 2)
+  );
+  fs.writeFileSync(
+    path.join(dir, "src-tauri", "tauri.conf.json"),
+    JSON.stringify({ productName: "jellytau", version: "0.0.1" }, null, 2)
+  );
+  // A dependency carrying its own `version =` is the trap: a greedy regex
+  // rewrites it too and the build then resolves the wrong crate.
+  fs.writeFileSync(
+    path.join(dir, "src-tauri", "Cargo.toml"),
+    ['[package]', 'name = "jellytau"', 'version = "0.0.1"', '', '[dependencies]', 'serde = { version = "1.0.100" }', ''].join("\n")
+  );
+  fs.writeFileSync(
+    path.join(dir, "src-tauri", "Cargo.lock"),
+    ['[[package]]', 'name = "serde"', 'version = "1.0.100"', '', '[[package]]', 'name = "jellytau"', 'version = "0.0.1"', ''].join("\n")
+  );
+  fs.writeFileSync(
+    path.join(dir, "src-tauri", "gen", "android", "app", "tauri.properties"),
+    "tauri.android.versionCode=1\n"
+  );
+}
+
+function run(version: string, dir = tmp) {
+  return execFileSync("bash", [path.join(dir, "scripts", "set-version.sh"), version], {
+    cwd: dir,
+    encoding: "utf-8",
+  });
+}
+
+function read(rel: string): string {
+  return fs.readFileSync(path.join(tmp, rel), "utf-8");
+}
+
+function versionCode(): number {
+  const m = read("src-tauri/gen/android/app/tauri.properties").match(
+    /^tauri\.android\.versionCode=(\d+)$/m
+  );
+  return m ? Number(m[1]) : NaN;
+}
+
+beforeEach(() => {
+  tmp = fs.mkdtempSync(path.join(os.tmpdir(), "setversion-"));
+  seed(tmp);
+});
+
+afterEach(() => {
+  fs.rmSync(tmp, { recursive: true, force: true });
+});
+
+describe("set-version.sh", () => {
+  it("stamps the version into all four manifests", () => {
+    run("0.5.0");
+    expect(JSON.parse(read("package.json")).version).toBe("0.5.0");
+    expect(JSON.parse(read("src-tauri/tauri.conf.json")).version).toBe("0.5.0");
+    expect(read("src-tauri/Cargo.toml")).toContain('version = "0.5.0"');
+    expect(read("src-tauri/Cargo.lock")).toMatch(/name = "jellytau"\nversion = "0\.5\.0"/);
+  });
+
+  it("accepts a leading v, as git tags are written", () => {
+    run("v0.5.0");
+    expect(JSON.parse(read("package.json")).version).toBe("0.5.0");
+  });
+
+  // The regression that motivates anchoring the patterns.
+  it("does not rewrite dependency versions", () => {
+    run("0.5.0");
+    expect(read("src-tauri/Cargo.toml")).toContain('serde = { version = "1.0.100" }');
+    expect(read("src-tauri/Cargo.lock")).toMatch(/name = "serde"\nversion = "1\.0\.100"/);
+    expect(JSON.parse(read("package.json")).dependencies.hls).toBe("1.2.3");
+  });
+
+  describe("Android versionCode", () => {
+    // A newer release must never produce a smaller number than an older one, or
+    // Android refuses the update. The floor tracks the highest code actually in
+    // the field, which is NOT the same as the highest this formula has produced:
+    // v0.5.2 shipped versionCode 5002 from an earlier `minor*1000` scheme, while
+    // the `minor*100` formula that replaced it yields only 1502 for that same
+    // version — so every 0.5.x release built from it was an un-installable
+    // downgrade for anyone already on v0.5.2. The floor is raised to clear it.
+    it("clears the highest code shipped by earlier builds", () => {
+      run("0.0.1");
+      // v0.5.2 shipped 5002; anything at or below that cannot install over it.
+      expect(versionCode()).toBeGreaterThan(5002);
+    });
+
+    it("keeps 0.5.3 installable over the 5002 that shipped as v0.5.2", () => {
+      run("0.5.3");
+      expect(versionCode()).toBeGreaterThan(5002);
+    });
+
+    it("uses 10000 + major*1000000 + minor*1000 + patch", () => {
+      const cases: Array<[string, number]> = [
+        ["0.0.14", 10014],
+        ["0.0.15", 10015],
+        ["0.1.0", 11000],
+        ["0.4.8", 14008],
+        ["0.5.0", 15000],
+        ["0.5.3", 15003],
+        ["1.0.0", 1010000],
+      ];
+      for (const [version, code] of cases) {
+        seed(tmp);
+        run(version);
+        expect(versionCode(), `versionCode for ${version}`).toBe(code);
+      }
+    });
+
+    it("increases monotonically across an upgrade sequence", () => {
+      const ordered = ["0.0.14", "0.0.15", "0.1.0", "0.4.8", "0.5.0", "1.0.0"];
+      const codes = ordered.map((v) => {
+        seed(tmp);
+        run(v);
+        return versionCode();
+      });
+      const sorted = [...codes].sort((a, b) => a - b);
+      expect(codes).toEqual(sorted);
+      expect(new Set(codes).size).toBe(codes.length);
+    });
+
+    // `$(( 0-rc1 ))` aborts the script under `set -e`, so the suffix has to be
+    // stripped before the arithmetic.
+    it("derives the code from the numeric core of a prerelease", () => {
+      run("0.6.0-rc1");
+      expect(versionCode()).toBe(16000);
+      expect(JSON.parse(read("package.json")).version).toBe("0.6.0-rc1");
+    });
+  });
+
+  describe("input validation", () => {
+    it("rejects a malformed version without writing anything", () => {
+      expect(() => run("not-a-version")).toThrow();
+      // The manifests must be untouched, not half-written.
+      expect(JSON.parse(read("package.json")).version).toBe("0.0.1");
+      expect(JSON.parse(read("src-tauri/tauri.conf.json")).version).toBe("0.0.1");
+    });
+
+    // CI passes "${GITHUB_REF#refs/tags/}" unconditionally; on a branch build
+    // that is still a full ref, and must not fail the job.
+    it("falls back to a dev version when handed a non-tag ref", () => {
+      const out = run("refs/heads/master");
+      expect(out).not.toMatch(/refs\/heads/);
+      expect(JSON.parse(read("package.json")).version).not.toBe("0.0.1");
+    });
+  });
+});

scripts/sync-android-sources.sh

@@ -87,6 +87,16 @@ if [ -d "$RES_SRC" ]; then
         cp "$RES_SRC"/values/*.xml "$RES_DST/values/"
         echo "  Copied res: values"
     fi
+
+    # xml/ (network_security_config.xml): referenced from the manifest, so a
+    # missing copy fails the resource link rather than degrading quietly.
+    # Merged into Tauri's generated xml/ (which holds file_paths.xml) rather
+    # than replacing it.
+    if [ -d "$RES_SRC/xml" ]; then
+        mkdir -p "$RES_DST/xml"
+        cp "$RES_SRC"/xml/*.xml "$RES_DST/xml/"
+        echo "  Copied res: xml"
+    fi
     # We ship only the color adaptive icon (background + foreground). Drop any
     # monochrome layer Tauri may generate: the themed-icon monochrome doesn't
     # render well, and our adaptive-icon xml no longer references it, so a stray
@@ -108,4 +118,26 @@ if [ -d "$RES_SRC" ]; then
           "$RES_DST"/drawable*/ic_launcher_background.xml
 fi
 
+# Gradle wrapper distribution. `tauri android init` regenerates the wrapper
+# pointing at services.gradle.org, so each build downloads ~130MB of Gradle —
+# slow, and a hard failure when the CDN drops the connection mid-transfer
+# ("Unexpected end of file from server"), which is what broke the release APK
+# job. The builder image ships the matching distribution under /opt/gradle/dist,
+# so when it's present repoint the wrapper at that local zip and build offline.
+# Outside the image (dev machines) the properties file is left untouched and the
+# wrapper downloads as usual.
+WRAPPER_PROPS="$PROJECT_ROOT/src-tauri/gen/android/gradle/wrapper/gradle-wrapper.properties"
+if [ -f "$WRAPPER_PROPS" ]; then
+    WANTED_VERSION="$(sed -n 's#.*/gradle-\([0-9.]*\)-\(bin\|all\)\.zip.*#\1#p' "$WRAPPER_PROPS")"
+    LOCAL_DIST="/opt/gradle/dist/gradle-${WANTED_VERSION}-bin.zip"
+    if [ -n "$WANTED_VERSION" ] && [ -f "$LOCAL_DIST" ]; then
+        # distributionUrl is a java.util.Properties value: ':' must stay escaped.
+        sed -i "s#^distributionUrl=.*#distributionUrl=file\\\\:///opt/gradle/dist/gradle-${WANTED_VERSION}-bin.zip#" \
+            "$WRAPPER_PROPS"
+        echo "  Gradle wrapper -> local distribution ($WANTED_VERSION, offline)"
+    elif [ -n "$WANTED_VERSION" ]; then
+        echo "  Gradle wrapper: $WANTED_VERSION not in image, will download"
+    fi
+fi
+
 echo "✓ Android sources synced successfully"

scripts/test-all.sh

@@ -7,7 +7,7 @@ echo "🧪 Running all tests..."
 echo ""
 
 echo "📦 Running frontend tests..."
-bun run test
+bun run test --run
 
 echo ""
 echo "🦀 Running Rust tests..."
@@ -15,5 +15,12 @@ cd src-tauri
 cargo test
 cd ..
 
+echo ""
+echo "🚧 Checking architectural gates..."
+# Boundary tripwire (DR-094): no Jellyfin taxonomy in the presentation layer.
+bun run check:boundary
+# Traceability coverage (DR-093): fails below 50%, or above 100% (miscount).
+bun run traces:coverage
+
 echo ""
 echo "✅ All tests passed!"

src-tauri/Cargo.lock

@@ -150,6 +150,12 @@ version = "1.0.100"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a23eb6b1614318a8071c9b2521f36b424b2c83db5eb3a0fead4a6c0809af6e61"
 
+[[package]]
+name = "ascii"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d92bec98840b8f03a5ff5413de5293bfcd8bf96467cf5452609f939ec6f5de16"
+
 [[package]]
 name = "async-broadcast"
 version = "0.7.2"
@@ -552,6 +558,12 @@ dependencies = [
  "windows-link 0.2.1",
 ]
 
+[[package]]
+name = "chunked_transfer"
+version = "1.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e4de3bc4ea267985becf712dc6d9eed8b04c953b3fcfb339ebc87acd9804901"
+
 [[package]]
 name = "cipher"
 version = "0.4.4"
@@ -1671,12 +1683,24 @@ dependencies = [
  "pin-project-lite",
 ]
 
+[[package]]
+name = "http-range"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "21dec9db110f5f872ed9699c3ecf50cf16f423502706ba5c72462e28d3157573"
+
 [[package]]
 name = "httparse"
 version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
 
+[[package]]
+name = "httpdate"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df3b46402a9d5adb4c86a0cf463f42e19994e3ee891101b1841f30a545cb49a9"
+
 [[package]]
 name = "hyper"
 version = "1.8.1"
@@ -1994,7 +2018,7 @@ dependencies = [
 
 [[package]]
 name = "jellytau"
-version = "0.1.0"
+version = "0.6.0"
 dependencies = [
  "aes-gcm",
  "async-trait",
@@ -2025,6 +2049,7 @@ dependencies = [
  "tauri-plugin-os",
  "tauri-specta",
  "tempfile",
+ "tiny_http",
  "tokio",
  "tokio-rusqlite",
  "tokio-util",
@@ -4192,6 +4217,7 @@ dependencies = [
  "gtk",
  "heck 0.5.0",
  "http",
+ "http-range",
  "jni",
  "libc",
  "log",
@@ -4571,6 +4597,18 @@ dependencies = [
  "time-core",
 ]
 
+[[package]]
+name = "tiny_http"
+version = "0.12.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "389915df6413a2e74fb181895f933386023c71110878cd0825588928e64cdc82"
+dependencies = [
+ "ascii",
+ "chunked_transfer",
+ "httpdate",
+ "log",
+]
+
 [[package]]
 name = "tinystr"
 version = "0.8.2"

src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "jellytau"
-version = "0.1.0"
+version = "0.6.0"
 description = "A Tauri App"
 authors = ["you"]
 edition = "2021"
@@ -23,7 +23,12 @@ debug = "line-tables-only"
 tauri-build = { version = "2", features = [] }
 
 [dependencies]
-tauri = { version = "2", features = [] }
+# protocol-asset serves downloaded media and cached thumbnails to the webview
+# over http://asset.localhost; without it convertFileSrc yields a URL nothing
+# answers. Paired with app.security.assetProtocol in tauri.conf.json, which
+# scopes it to $APPDATA/**.
+# TRACES: UR-071 | DR-134
+tauri = { version = "2", features = ["protocol-asset"] }
 tauri-plugin-opener = "2"
 tauri-plugin-os = "2"
 serde = { version = "1", features = ["derive"] }
@@ -54,6 +59,7 @@ env_logger = "0.11"
 tauri-specta = { version = "=2.0.0-rc.21", features = ["derive", "typescript"] }
 specta-typescript = "=0.0.9"
 specta = { version = "=2.0.0-rc.22", features = ["chrono", "derive"] }
+tiny_http = { version = "0.12.0", default-features = false }
 
 # Linux-specific dependencies
 [target.'cfg(target_os = "linux")'.dependencies]

src-tauri/android/README_ANDROID_BUILD.md

@@ -41,6 +41,57 @@ When you need to modify Android/Kotlin files:
 - If you only edit `src-tauri/android/`, your changes won't be in the build
 - **You must edit both** (or edit source and copy to generated)
 
+### Debug and release install side by side
+
+The **debug** build type sets `applicationIdSuffix = ".debug"` in
+`app/build.gradle.kts`, so a debug build is a genuinely separate Android app:
+
+| build | applicationId | launcher name | versionName | signed with |
+|---|---|---|---|---|
+| `release` | `com.dtourolle.jellytau` | jellytau | `0.5.5` | real key (`.env`) |
+| `release --debug` | `com.dtourolle.jellytau.debug` | JellyTau Debug | `0.5.5-debug-release` | debug keystore |
+| `debug` | `com.dtourolle.jellytau.debug` | JellyTau Debug | `0.5.5-debug` | debug keystore |
+
+`release --debug` is the **side-by-side release**: fully R8-minified, exactly
+what ships, but packaged into the debug slot and signed with the local debug
+keystore. It exists because R8 has broken release APKs here before (stripping
+JNI-loaded player/security classes), and reproducing that previously meant
+building with the real key and clobbering your working install. It shares the
+applicationId *and* signature with the plain debug build, so the two replace
+each other cleanly; only the versionName suffix tells you which is installed.
+
+```bash
+./scripts/build-and-deploy.sh release --device --debug   # build + install it
+```
+
+The flag is plumbed through as `JT_SIDE_BY_SIDE=1`, read by `build.gradle.kts`.
+CI never sets it, so distributable release builds are untouched.
+
+That means:
+
+- **No uninstall step.** Debug builds are signed with the local auto-generated
+  `~/.android/debug.keystore`, release builds with the real key. Two different
+  keys on the *same* package is `INSTALL_FAILED_UPDATE_INCOMPATIBLE`; two
+  different packages is just two apps.
+- Each has **its own data directory** — separate settings, credentials,
+  downloads and offline cache. A debug experiment cannot corrupt the state of
+  the build you actually use. This is not optional and cannot be shared:
+  Android gives each applicationId its own UID and enforces the boundary in the
+  kernel. (`sharedUserId` is deprecated since API 29 and cannot be added to an
+  already-installed app anyway.) You log in again in the debug app, once.
+- Only the *application* id changes. Kotlin classes stay in the `namespace`
+  package `com.dtourolle.jellytau`, so the JNI class lookups in
+  `src-tauri/src/player/android/mod.rs`, the manifest `<service>` entry and the
+  R8 keep rules in `proguard-jellytau.pro` are all unaffected. The FileProvider
+  authority is `${applicationId}.fileprovider`, so it follows the suffix
+  automatically.
+- The launcher labels come from the `appLabel` / `activityLabel`
+  manifestPlaceholders (`AndroidManifest.xml` uses `${appLabel}`), *not* from
+  `resValue`, which would collide with Tauri's generated `strings.xml`.
+
+Follow the right log stream with `./scripts/logcat.sh [debug|release]`
+(defaults to debug).
+
 ### Key Files
 
 Player-related Kotlin files:

src-tauri/android/app/build.gradle.kts

@@ -22,11 +22,25 @@ val keystoreProperties = Properties().apply {
     }
 }
 
+// Side-by-side release: set by `scripts/build-android.sh release --debug`, which
+// exports JT_SIDE_BY_SIDE=1. It puts a fully R8-minified release build into the
+// debug applicationId slot, signed with the local debug keystore — so you can
+// test what minification actually produces (R8 stripping JNI-loaded classes has
+// broken release APKs here before) without the real signing key and without
+// uninstalling your working install. Unset in CI, so distributable release
+// builds are untouched.
+val sideBySideRelease = System.getenv("JT_SIDE_BY_SIDE").let { it == "1" || it == "true" }
+
 android {
     compileSdk = 36
     namespace = "com.dtourolle.jellytau"
     defaultConfig {
         manifestPlaceholders["usesCleartextTraffic"] = "false"
+        // Launcher/app names come from placeholders so the debug build can
+        // rename itself without touching the generated strings.xml (a
+        // resValue() override there would collide with Tauri's own entries).
+        manifestPlaceholders["appLabel"] = "@string/app_name"
+        manifestPlaceholders["activityLabel"] = "@string/main_activity_title"
         applicationId = "com.dtourolle.jellytau"
         minSdk = 24
         targetSdk = 36
@@ -45,6 +59,21 @@ android {
     }
     buildTypes {
         getByName("debug") {
+            // Distinct applicationId so the debug build installs SIDE BY SIDE
+            // with a release/store install instead of demanding an uninstall
+            // (different signing keys on the same package = INSTALL_FAILED_
+            // UPDATE_INCOMPATIBLE). It gets its own data dir, its own settings
+            // and its own offline cache — the two are fully independent apps.
+            //
+            // This changes only the *application* id. The Kotlin/JNI classes
+            // stay in the `namespace` package (com.dtourolle.jellytau), so the
+            // fully-qualified class names Rust looks up over JNI, the manifest
+            // <service> entry and the R8 keep rules are all unaffected. The
+            // FileProvider authority is already ${applicationId}-relative.
+            applicationIdSuffix = ".debug"
+            versionNameSuffix = "-debug"
+            manifestPlaceholders["appLabel"] = "JellyTau Debug"
+            manifestPlaceholders["activityLabel"] = "JellyTau Debug"
             manifestPlaceholders["usesCleartextTraffic"] = "true"
             isDebuggable = true
             isJniDebuggable = true
@@ -56,7 +85,18 @@ android {
             }
         }
         getByName("release") {
-            if (keystoreProperties.getProperty("storeFile") != null) {
+            if (sideBySideRelease) {
+                // Same slot, name and version scheme as the debug build type,
+                // plus "-release" so you can tell from Settings > Apps which of
+                // the two is currently sitting there. Signed with the debug
+                // keystore: it shares a signature with the debug build, so the
+                // two replace each other cleanly instead of colliding.
+                applicationIdSuffix = ".debug"
+                versionNameSuffix = "-debug-release"
+                manifestPlaceholders["appLabel"] = "JellyTau Debug"
+                manifestPlaceholders["activityLabel"] = "JellyTau Debug"
+                signingConfig = signingConfigs.getByName("debug")
+            } else if (keystoreProperties.getProperty("storeFile") != null) {
                 signingConfig = signingConfigs.getByName("release")
             }
             isMinifyEnabled = true

src-tauri/android/app/proguard-jellytau.pro

@@ -15,6 +15,7 @@
 # breaking the PiP button in release builds only.
 -keep class com.dtourolle.jellytau.PictureInPictureManager { *; }
 -keep class com.dtourolle.jellytau.VideoOverlayManager { *; }
+-keep class com.dtourolle.jellytau.WindowInsetsBridge { *; }
 -keepclassmembers class * {
     @android.webkit.JavascriptInterface <methods>;
 }

src-tauri/android/src/main/AndroidManifest.xml

@@ -11,6 +11,12 @@
   (An earlier version of this file was a partial <application> fragment on the
   assumption that Tauri merged it. It did not: the hardwareAccelerated flag it
   declared never reached any built APK. It is folded in properly below.)
+
+  ${appLabel} / ${activityLabel} are manifestPlaceholders set in
+  app/build.gradle.kts: they resolve to @string/app_name and
+  @string/main_activity_title for release, and to "JellyTau Debug" for the
+  debug build type (which also carries applicationIdSuffix ".debug" so it
+  installs alongside a release build).
 -->
 <manifest xmlns:android="http://schemas.android.com/apk/res/android">
     <uses-permission android:name="android.permission.INTERNET" />
@@ -26,14 +32,15 @@
 
     <application
         android:icon="@mipmap/ic_launcher"
-        android:label="@string/app_name"
+        android:label="${appLabel}"
         android:theme="@style/Theme.jellytau"
         android:hardwareAccelerated="true"
+        android:networkSecurityConfig="@xml/network_security_config"
         android:usesCleartextTraffic="${usesCleartextTraffic}">
         <activity
             android:configChanges="orientation|keyboardHidden|keyboard|screenSize|locale|smallestScreenSize|screenLayout|uiMode|density"
             android:launchMode="singleTask"
-            android:label="@string/main_activity_title"
+            android:label="${activityLabel}"
             android:name=".MainActivity"
             android:exported="true"
             android:supportsPictureInPicture="true"

src-tauri/android/src/main/java/com/dtourolle/jellytau/ImmersiveModeBridge.kt

@@ -0,0 +1,67 @@
+package com.dtourolle.jellytau
+
+import android.app.Activity
+import androidx.core.view.WindowCompat
+import androidx.core.view.WindowInsetsCompat
+import androidx.core.view.WindowInsetsControllerCompat
+
+/**
+ * Hides and restores the Android system bars for full-screen video.
+ *
+ * TRACES: UR-066 | DR-157
+ *
+ * ## Why the web layer cannot do this
+ *
+ * `document.documentElement.requestFullscreen()` is the only fullscreen control
+ * the frontend has, and inside an Android WebView it does nothing to the
+ * *Activity*: it expands the fullscreen element within the web viewport and
+ * leaves the window exactly as it was. Combined with `enableEdgeToEdge()` — which
+ * MainActivity must call, and which SDK 36 makes non-optional — the WebView
+ * already spans the whole window, so "fullscreen" was a no-op that changed
+ * nothing on screen while the status bar and navigation/gesture bar stayed
+ * painted over the video.
+ *
+ * Hiding them requires `WindowInsetsControllerCompat` on the Activity's window,
+ * which is reachable only from native code. Hence this bridge.
+ *
+ * ## Behaviour
+ *
+ * [enter] hides both bars and selects `BEHAVIOR_SHOW_TRANSIENT_BARS_BY_SWIPE`, so
+ * a swipe from either edge brings them back *transiently* — over the video,
+ * auto-hiding again — rather than permanently resizing the window mid-playback.
+ * That is the standard behaviour for immersive video and keeps the system's own
+ * back/home gestures reachable.
+ *
+ * [exit] restores them. It must be called when leaving fullscreen **and** when
+ * the player is torn down, or the bars stay hidden on the library screens behind
+ * it.
+ *
+ * Both must run on the main thread; the callers in MainActivity post them there,
+ * since `@JavascriptInterface` methods arrive on a WebView binder thread.
+ *
+ * Note the `--jt-inset-*` custom properties follow automatically: hiding the bars
+ * fires the decor view's inset listener with zeroes, so [WindowInsetsBridge]
+ * republishes them and the player's control layer stops reserving space it no
+ * longer needs.
+ */
+object ImmersiveModeBridge {
+
+    private fun controller(activity: Activity): WindowInsetsControllerCompat =
+        WindowCompat.getInsetsController(activity.window, activity.window.decorView)
+
+    /** Hide the status and navigation bars, swipe-to-reveal transiently. */
+    fun enter(activity: Activity) {
+        controller(activity).apply {
+            systemBarsBehavior =
+                WindowInsetsControllerCompat.BEHAVIOR_SHOW_TRANSIENT_BARS_BY_SWIPE
+            hide(WindowInsetsCompat.Type.systemBars())
+        }
+        android.util.Log.d("ImmersiveMode", "system bars hidden")
+    }
+
+    /** Restore the system bars. Safe to call when they are already showing. */
+    fun exit(activity: Activity) {
+        controller(activity).show(WindowInsetsCompat.Type.systemBars())
+        android.util.Log.d("ImmersiveMode", "system bars restored")
+    }
+}

src-tauri/android/src/main/java/com/dtourolle/jellytau/MainActivity.kt

@@ -1,10 +1,5 @@
 package com.dtourolle.jellytau
 
-import android.content.Context
-import android.media.AudioAttributes
-import android.media.AudioFocusRequest
-import android.media.AudioManager
-import android.os.Build
 import android.os.Bundle
 import android.os.Handler
 import android.os.Looper
@@ -15,12 +10,11 @@ import android.webkit.WebView
 import android.view.View
 import androidx.activity.enableEdgeToEdge
 
+
 class MainActivity : TauriActivity() {
   private val handler = Handler(Looper.getMainLooper())
   private var configAttempts = 0
   private val maxConfigAttempts = 10
-  private var audioFocusRequest: AudioFocusRequest? = null
-  private val audioManager by lazy { getSystemService(Context.AUDIO_SERVICE) as AudioManager }
 
   /**
    * Coarse override for whether backgrounding the app should auto-enter PiP.
@@ -50,10 +44,77 @@ class MainActivity : TauriActivity() {
    */
   private var mediaWebView: WebView? = null
 
+  /**
+   * The WebView the @JavascriptInterface bridges have been injected into.
+   *
+   * addJavascriptInterface must run once per WebView instance: re-injecting
+   * over an already-loaded page hands JS a stale proxy whose methods are gone.
+   * Compared by identity so a genuinely new WebView still gets its bridges.
+   */
+  private var bridgesInstalledOn: WebView? = null
+
+  /**
+   * wry hands us the WebView here, and this is the only point at which the
+   * bridges can be installed *deterministically*.
+   *
+   * WebView binds an injected object into JS at **page-load time**: an
+   * addJavascriptInterface call that lands after the page has loaded does not
+   * appear to that page at all. The bridges used to be installed from
+   * [configureWebViewForMedia], which finds the WebView by walking the view
+   * tree 500 ms after onCreate — a race against Tauri's own page load, and one
+   * that is *permanent* when lost, because the identity guard then declines to
+   * re-inject on the resume passes. The whole set (`AndroidVideoSurface`,
+   * `AndroidPictureInPicture`, `AndroidBackgroundAudio`, `AndroidNetworkType`,
+   * `AndroidImmersive`, `AndroidInsets`) simply would not exist in `window`,
+   * silently: every one of them is called through an optional chain, so a
+   * missing bridge is a no-op rather than an error. That is a candidate
+   * explanation for DR-172's central piece of evidence — native video shipped
+   * with `WebView transparent = false` logged and `= true` never appearing,
+   * i.e. the enable call never reaching Kotlin.
+   *
+   * `WryActivity.setWebView()` calls this immediately before wry issues the
+   * first `loadUrl`, so a bridge installed here is bound by the time any page
+   * runs. Note this can fire during `super.onCreate()`, i.e. *before* the rest
+   * of our own onCreate — so only work that needs nothing but the WebView
+   * belongs here. Insets are deliberately left to
+   * [configureWebViewForMedia], which runs later and on every resume.
+   *
+   * TRACES: UR-003, UR-004 | DR-183
+   */
+  override fun onWebViewCreate(webView: WebView) {
+    super.onWebViewCreate(webView)
+    android.util.Log.d("MainActivity", "onWebViewCreate - installing bridges before first page load")
+    mediaWebView = webView
+    installJavascriptBridges(webView)
+    configureWebViewSettings(webView)
+  }
+
   override fun onCreate(savedInstanceState: Bundle?) {
     enableEdgeToEdge()
     super.onCreate(savedInstanceState)
 
+    // enableEdgeToEdge() puts the WebView under the status bar, the navigation/
+    // gesture bar and the display cutout — and targeting SDK 36 makes that
+    // non-optional anyway. Android WebView never surfaces the *system bar*
+    // insets to CSS (only the display cutout), so the web layer has to be told.
+    // Without this the bottom nav renders underneath the navigation bar, badly
+    // so on devices with a tall opaque 3-button bar. (UR-066)
+    WindowInsetsBridge.install(this)
+
+    // Hand the player an Activity reference so it can attach its video
+    // SurfaceView to the content view behind the WebView.
+    //
+    // Without this, JellyTauPlayer.currentActivity stays null forever and
+    // autoAttachSurface() logs "Cannot attach surface - no Activity reference"
+    // and returns — so the SurfaceView is created, wired to ExoPlayer, and then
+    // never added to the view hierarchy. Native video decoded to a surface that
+    // was never on screen. setActivity() stores into a companion-object
+    // WeakReference, so calling it here (before Rust initializes the player over
+    // JNI) is safe and is the case it was written for.
+    //
+    // TRACES: UR-003, UR-041 | DR-151
+    com.dtourolle.jellytau.player.JellyTauPlayer.setActivity(this)
+
     // Configure WebView for media playback after Tauri initialization
     handler.postDelayed({
       configureWebViewForMedia()
@@ -130,6 +191,37 @@ class MainActivity : TauriActivity() {
     super.onDestroy()
   }
 
+  /**
+   * Rotation (and any other config change this Activity handles itself).
+   *
+   * Two things have to happen here rather than later, and both are about the
+   * *previous* video frame surviving the transition:
+   *
+   *  - The video view is hidden until a new frame arrives. The equivalent call
+   *    in `fitSurfaceToScreen` runs from the content view's layout listener,
+   *    which is after the rotation — by then the stale frame has been on screen
+   *    for the whole transition.
+   *  - The window's rotation animation is a **cross-fade of a screenshot** of
+   *    the old orientation, and that screenshot contains the old video frame at
+   *    the old size. No amount of TextureView bookkeeping can touch it, which is
+   *    why hiding on frame-arrival alone did not stop the flash. `JUMPCUT` drops
+   *    the cross-fade, so there is no old frame to fade through; it is set only
+   *    while native compositing is active (see setTransparent) so the rest of
+   *    the app keeps the normal animation.
+   *
+   * TRACES: UR-003, UR-066 | DR-194
+   */
+  override fun onConfigurationChanged(newConfig: android.content.res.Configuration) {
+    super.onConfigurationChanged(newConfig)
+    try {
+      if (com.dtourolle.jellytau.player.JellyTauPlayer.isInitialized()) {
+        com.dtourolle.jellytau.player.JellyTauPlayer.getInstance().hideUntilFreshFrame()
+      }
+    } catch (e: Exception) {
+      android.util.Log.w("MainActivity", "hideUntilFreshFrame on config change failed", e)
+    }
+  }
+
   override fun onPictureInPictureModeChanged(
     isInPictureInPictureMode: Boolean,
     newConfig: android.content.res.Configuration
@@ -141,7 +233,9 @@ class MainActivity : TauriActivity() {
 
   private fun configureWebViewForMedia() {
     try {
-      val webView = findWebView(window.decorView)
+      // onWebViewCreate normally got here first; the tree walk is the fallback
+      // for a WebView we were never handed.
+      val webView = mediaWebView ?: findWebView(window.decorView)
 
       if (webView == null) {
         android.util.Log.w("MainActivity", "WebView not found (attempt ${configAttempts + 1}/$maxConfigAttempts)")
@@ -159,19 +253,56 @@ class MainActivity : TauriActivity() {
       android.util.Log.d("MainActivity", "WebView found! Configuring settings...")
       mediaWebView = webView
 
-      // Add JavaScript interface for audio focus control
-      webView.addJavascriptInterface(object : Any() {
-        @JavascriptInterface
-        fun requestAudioFocus() {
-          handler.post { this@MainActivity.requestAudioFocus() }
-        }
+      // Re-push the safe-area insets. Unlike addJavascriptInterface this is
+      // idempotent and MUST re-run: a page load discards the inline style the
+      // last push set, so the WebView would otherwise be left with no insets.
+      WindowInsetsBridge.attachWebView(webView)
 
-        @JavascriptInterface
-        fun abandonAudioFocus() {
-          handler.post { this@MainActivity.abandonAudioFocus() }
-        }
-      }, "AndroidAudioFocus")
-      android.util.Log.d("MainActivity", "JavaScript interface 'AndroidAudioFocus' added")
+      // Normally already done by onWebViewCreate; this is the fallback path.
+      installJavascriptBridges(webView)
+      configureWebViewSettings(webView)
+
+    } catch (e: Exception) {
+      android.util.Log.e("MainActivity", "Failed to configure WebView for media", e)
+    }
+  }
+
+  /**
+   * Register the @JavascriptInterface bridges EXACTLY ONCE per WebView.
+   *
+   * This runs from [onWebViewCreate] — the only point early enough to be bound
+   * before the first page load — and from [configureWebViewForMedia] as a
+   * fallback. The latter runs from onCreate's delayed post AND from every
+   * onResume (plus each WebView re-find), so without the identity guard this
+   * re-injected every bridge repeatedly — 5 times in a 45s session. WebView
+   * binds injected objects at page-load time; re-injecting over a live page
+   * leaves JS holding a stale proxy. The object stays truthy while its methods
+   * vanish, which surfaced as a flood of "WebView: Unknown object" chromium
+   * errors and, in JS, "TypeError: setEnabled is not a function".
+   *
+   * The visible bug: the background-audio toggle turned blue but never reached
+   * native, so backgroundAudioEnabled stayed false, onStop never dispatched
+   * 'jellytau-background', and a locked screen killed audio instantly (UR-040).
+   * Audio focus and PiP broke the same way.
+   *
+   * Settings/WebChromeClient work is idempotent and must keep running on
+   * resume, so it lives in [configureWebViewSettings], not here.
+   *
+   * TRACES: UR-003, UR-004, UR-040, UR-041 | DR-183
+   */
+  private fun installJavascriptBridges(webView: WebView) {
+    try {
+      if (webView === bridgesInstalledOn) {
+        android.util.Log.d("MainActivity", "JS bridges already installed on this WebView - skipping re-injection")
+        return
+      }
+      bridgesInstalledOn = webView
+
+      // NOTE: there is deliberately no "AndroidAudioFocus" bridge. Manual focus
+      // requests from the WebView competed with Chromium's own
+      // AudioFocusDelegate and with ExoPlayer, and the resulting
+      // AUDIOFOCUS_LOSS paused playback. See the comment on the video listeners
+      // in configureWebViewSettings().
 
       // Add JavaScript interface for picture-in-picture control.
       // enterPip/canEnterPip must run on the main thread; @JavascriptInterface
@@ -199,6 +330,19 @@ class MainActivity : TauriActivity() {
         fun setAutoEnterEnabled(enabled: Boolean) {
           autoEnterPipEnabled = enabled
         }
+
+        /**
+         * Report the WebView `<video>` state.
+         *
+         * Without this PiP only ever knew about the native ExoPlayer surface,
+         * which is behind an experimental flag that defaults to off — so in the
+         * shipping configuration nothing ever satisfied canEnterPip and the
+         * button did nothing. (DR-160)
+         */
+        @JavascriptInterface
+        fun setHtml5VideoState(active: Boolean, width: Int, height: Int, playing: Boolean) {
+          PictureInPictureManager.setHtml5VideoState(active, width, height, playing)
+        }
       }, "AndroidPictureInPicture")
       android.util.Log.d("MainActivity", "JavaScript interface 'AndroidPictureInPicture' added")
 
@@ -212,10 +356,6 @@ class MainActivity : TauriActivity() {
           backgroundAudioEnabled = enabled
           android.util.Log.d("MainActivity", "backgroundAudioEnabled = $enabled")
         }
-
-        /** Whether background audio is available on this device (needs PiP-era APIs unnecessary; audio service always present on Android). */
-        @JavascriptInterface
-        fun isSupported(): Boolean = true
       }, "AndroidBackgroundAudio")
       android.util.Log.d("MainActivity", "JavaScript interface 'AndroidBackgroundAudio' added")
 
@@ -242,12 +382,120 @@ class MainActivity : TauriActivity() {
       }, "AndroidNetworkType")
       android.util.Log.d("MainActivity", "JavaScript interface 'AndroidNetworkType' added")
 
+      // Native video compositing: let the frontend make the WebView transparent
+      // so the ExoPlayer SurfaceView behind it is visible (UR-003, UR-004).
+      //
+      // Toggled rather than set once because a transparent WebView is only
+      // correct while a native video is on screen — every other screen needs its
+      // opaque background, and leaving the window transparent shows the
+      // launcher/wallpaper through the app.
+      //
+      // The CSS in app.css clears the *web* layer's backgrounds; this clears the
+      // WebView widget's own background, which CSS cannot reach. Both are
+      // required — an opaque WebView hides the surface no matter what the page
+      // paints.
+      //
+      // TRACES: UR-003, UR-004 | DR-150
+      webView.addJavascriptInterface(object : Any() {
+        /** Make the WebView background transparent (true) or opaque (false). */
+        @JavascriptInterface
+        fun setTransparent(transparent: Boolean) {
+          handler.post {
+            val color = if (transparent) {
+              android.graphics.Color.TRANSPARENT
+            } else {
+              android.graphics.Color.BLACK
+            }
+            mediaWebView?.setBackgroundColor(color)
+            // The WebView's window/surface must also stop painting opaque, or a
+            // hardware-accelerated WebView still composites its own background.
+            window.setBackgroundDrawable(
+              android.graphics.drawable.ColorDrawable(color)
+            )
+            // Drop the rotation cross-fade while a native video surface is
+            // composited behind the page. The animation fades a *screenshot* of
+            // the old orientation, which still holds the previous video frame at
+            // the old size — that is the "previous frame flashing in the black
+            // bars", and it lives in the window animation rather than in
+            // anything the TextureView owns. (DR-194)
+            val attrs = window.attributes
+            attrs.rotationAnimation = if (transparent) {
+              android.view.WindowManager.LayoutParams.ROTATION_ANIMATION_JUMPCUT
+            } else {
+              android.view.WindowManager.LayoutParams.ROTATION_ANIMATION_ROTATE
+            }
+            window.attributes = attrs
+            // `rotationAnimation` is honoured only for a **fullscreen** window —
+            // the platform says so out loud, logging
+            // "VRI[MainActivity]: setLayoutParams: not fullscreen" when the
+            // attribute is set on ours, and then animating normally regardless.
+            // Without this the JUMPCUT above is accepted and ignored, and the
+            // cross-fade keeps showing the old orientation's screenshot, stale
+            // video frame and all. FLAG_FULLSCREEN is deprecated for *hiding
+            // system bars* (immersive mode does that, on player entry), but it
+            // is still what marks the window fullscreen for this decision.
+            @Suppress("DEPRECATION")
+            if (transparent) {
+              window.addFlags(android.view.WindowManager.LayoutParams.FLAG_FULLSCREEN)
+            } else {
+              window.clearFlags(android.view.WindowManager.LayoutParams.FLAG_FULLSCREEN)
+            }
+            android.util.Log.d("MainActivity", "WebView transparent = $transparent")
+          }
+        }
+
+        /** Whether native-video compositing is available on this platform. */
+        @JavascriptInterface
+        fun isSupported(): Boolean = true
+      }, "AndroidVideoSurface")
+      android.util.Log.d("MainActivity", "JavaScript interface 'AndroidVideoSurface' added")
+
+      // Full-screen video: hide the system bars (UR-066). requestFullscreen()
+      // inside a WebView cannot touch the Activity window, so without this the
+      // status and navigation bars stayed painted over full-screen video.
+      webView.addJavascriptInterface(object : Any() {
+        /** Hide the system bars for full-screen playback. */
+        @JavascriptInterface
+        fun enter() {
+          handler.post { ImmersiveModeBridge.enter(this@MainActivity) }
+        }
+
+        /** Restore the system bars on leaving fullscreen or the player. */
+        @JavascriptInterface
+        fun exit() {
+          handler.post { ImmersiveModeBridge.exit(this@MainActivity) }
+        }
+
+        /** Whether native immersive mode exists (false on non-Android). */
+        @JavascriptInterface
+        fun isSupported(): Boolean = true
+      }, "AndroidImmersive")
+      android.util.Log.d("MainActivity", "JavaScript interface 'AndroidImmersive' added")
+
+      // Window insets (safe areas). The push path above races the page load, so
+      // the frontend pulls the current values on mount through this bridge.
+      webView.addJavascriptInterface(WindowInsetsBridge.jsInterface(), "AndroidInsets")
+      android.util.Log.d("MainActivity", "JavaScript interface 'AndroidInsets' added")
+
       // Push network changes into the WebView so a queue blocked on "waiting for
       // WiFi" resumes the moment an acceptable network appears.
       NetworkTypeMonitor.startWatching(this) {
         dispatchWebEvent("jellytau-network-changed")
       }
 
+    } catch (e: Exception) {
+      android.util.Log.e("MainActivity", "Failed to install JavaScript bridges", e)
+    }
+  }
+
+  /**
+   * WebView settings, chrome client and the video-unmute script.
+   *
+   * Split out from the bridge injection because this half is idempotent and
+   * must re-run on every resume, whereas addJavascriptInterface must not.
+   */
+  private fun configureWebViewSettings(webView: WebView) {
+    try {
       // Set WebChromeClient to handle video playback and audio focus
       webView.webChromeClient = object : WebChromeClient() {
         override fun onShowCustomView(view: View?, callback: CustomViewCallback?) {
@@ -259,6 +507,21 @@ class MainActivity : TauriActivity() {
           super.onHideCustomView()
           android.util.Log.d("MainActivity", "Video exited fullscreen")
         }
+
+        /**
+         * Forward WebView console output to logcat under the "JellyTauWeb" tag.
+         *
+         * Without this the frontend is invisible to `adb logcat`, which makes
+         * diagnosing anything that spans the JS/native boundary (the
+         * background-audio handoff in particular) guesswork.
+         */
+        override fun onConsoleMessage(msg: android.webkit.ConsoleMessage): Boolean {
+          android.util.Log.d(
+            "JellyTauWeb",
+            "${msg.message()} (${msg.sourceId()}:${msg.lineNumber()})"
+          )
+          return true
+        }
       }
       android.util.Log.d("MainActivity", "WebChromeClient configured")
 
@@ -287,29 +550,18 @@ class MainActivity : TauriActivity() {
               video.volume = 1.0;
               console.log('[Android] Video unmuted, volume:', video.volume, 'muted:', video.muted);
 
-              // Add event listeners to manage audio focus
-              video.addEventListener('play', function() {
-                console.log('[Android] Video play event - requesting audio focus');
-                if (typeof AndroidAudioFocus !== 'undefined') {
-                  AndroidAudioFocus.requestAudioFocus();
-                }
-                console.log('[Android] Video state - muted:', this.muted, 'volume:', this.volume);
-              });
-
-              video.addEventListener('pause', function() {
-                console.log('[Android] Video pause event - abandoning audio focus');
-                if (typeof AndroidAudioFocus !== 'undefined') {
-                  AndroidAudioFocus.abandonAudioFocus();
-                }
-              });
-
-              video.addEventListener('ended', function() {
-                console.log('[Android] Video ended event - abandoning audio focus');
-                if (typeof AndroidAudioFocus !== 'undefined') {
-                  AndroidAudioFocus.abandonAudioFocus();
-                }
-              });
-
+              // NOTE: deliberately no audio-focus calls here.
+              //
+              // WebView already manages audio focus for <video> through
+              // Chromium's own AudioFocusDelegate. Requesting AUDIOFOCUS_GAIN
+              // again from MainActivity made two requesters compete inside one
+              // uid: the grant was immediately followed by AUDIOFOCUS_LOSS
+              // (~45ms), whose handler paused playback - so arming background
+              // audio, or simply pressing play, paused the video in a loop.
+              //
+              // ExoPlayer is the third potential owner and stays authoritative
+              // for native playback (JellyTauPlayer manages its own focus).
+              // Leave focus to whichever engine is actually rendering.
               video.addEventListener('volumechange', function() {
                 console.log('[Android] Video volume changed - volume:', this.volume, 'muted:', this.muted);
               });
@@ -356,48 +608,4 @@ class MainActivity : TauriActivity() {
     return null
   }
 
-  private fun requestAudioFocus() {
-    android.util.Log.d("MainActivity", "Requesting audio focus for video playback")
-
-    if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
-      val audioAttributes = AudioAttributes.Builder()
-        .setUsage(AudioAttributes.USAGE_MEDIA)
-        .setContentType(AudioAttributes.CONTENT_TYPE_MOVIE)
-        .build()
-
-      audioFocusRequest = AudioFocusRequest.Builder(AudioManager.AUDIOFOCUS_GAIN)
-        .setAudioAttributes(audioAttributes)
-        .setAcceptsDelayedFocusGain(true)
-        .setOnAudioFocusChangeListener { focusChange ->
-          android.util.Log.d("MainActivity", "Audio focus changed: $focusChange")
-        }
-        .build()
-
-      val result = audioManager.requestAudioFocus(audioFocusRequest!!)
-      android.util.Log.d("MainActivity", "Audio focus request result: $result")
-    } else {
-      @Suppress("DEPRECATION")
-      val result = audioManager.requestAudioFocus(
-        { focusChange ->
-          android.util.Log.d("MainActivity", "Audio focus changed: $focusChange")
-        },
-        AudioManager.STREAM_MUSIC,
-        AudioManager.AUDIOFOCUS_GAIN
-      )
-      android.util.Log.d("MainActivity", "Audio focus request result (legacy): $result")
-    }
-  }
-
-  private fun abandonAudioFocus() {
-    android.util.Log.d("MainActivity", "Abandoning audio focus")
-
-    if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
-      audioFocusRequest?.let {
-        audioManager.abandonAudioFocusRequest(it)
-      }
-    } else {
-      @Suppress("DEPRECATION")
-      audioManager.abandonAudioFocus { }
-    }
-  }
 }

src-tauri/android/src/main/java/com/dtourolle/jellytau/PictureInPictureManager.kt

@@ -46,6 +46,62 @@ object PictureInPictureManager {
     private var receiver: BroadcastReceiver? = null
     private var hiddenWebView: WebView? = null
 
+    /**
+     * State of an HTML5 `<video>` playing inside the WebView, reported by the
+     * frontend.
+     *
+     * PiP was written for the native ExoPlayer surface only — [canEnterPip]
+     * required a SurfaceView to be attached and rendering. But native video is
+     * behind `experimentalNativeVideo`, which defaults to **off**, so in the
+     * shipping configuration video plays in the WebView's `<video>` element and
+     * every one of those conditions is false. `enterPip` therefore always bailed
+     * with "no local video playing": PiP could not work at all, however the
+     * button was pressed.
+     *
+     * On this path the WebView *is* the video, which inverts two things: the
+     * WebView must stay visible in PiP rather than be hidden, and play/pause has
+     * to reach the element rather than ExoPlayer. Both are handled below.
+     *
+     * TRACES: UR-041 | DR-160
+     */
+    @Volatile
+    private var html5VideoActive = false
+
+    @Volatile
+    private var html5VideoPlaying = false
+
+    @Volatile
+    private var html5AspectRatio: Rational? = null
+
+    /**
+     * Report the WebView `<video>` state from the frontend.
+     *
+     * @param active whether a video element is currently the playback surface
+     * @param width intrinsic video width, for the PiP window's aspect ratio
+     * @param height intrinsic video height
+     * @param playing whether it is playing right now, for the PiP play/pause action
+     */
+    fun setHtml5VideoState(active: Boolean, width: Int, height: Int, playing: Boolean) {
+        html5VideoActive = active
+        html5VideoPlaying = playing
+        html5AspectRatio = if (active && width > 0 && height > 0) {
+            clampedRatio(width.toDouble() / height.toDouble())
+        } else {
+            null
+        }
+    }
+
+    /** True when PiP would be showing the native surface rather than the WebView. */
+    private fun isNativeVideoPath(): Boolean = try {
+        val player = JellyTauPlayer.getInstance()
+        player.isPlayingVideo() &&
+            player.getSurfaceView() != null &&
+            VideoOverlayManager.isVideoSurfaceAttached()
+    } catch (e: Exception) {
+        android.util.Log.w(TAG, "native video path check failed", e)
+        false
+    }
+
     /**
      * Whether this device/OS can do PiP at all. Android 8.0 introduced the API,
      * and the user (or device manufacturer) can disable the feature per-app.
@@ -64,15 +120,10 @@ object PictureInPictureManager {
      */
     fun canEnterPip(activity: Activity): Boolean {
         if (!isPipSupported(activity)) return false
-        return try {
-            val player = JellyTauPlayer.getInstance()
-            player.isPlayingVideo() &&
-                player.getSurfaceView() != null &&
-                VideoOverlayManager.isVideoSurfaceAttached()
-        } catch (e: Exception) {
-            android.util.Log.w(TAG, "canEnterPip check failed", e)
-            false
-        }
+        // Either surface will do: the native one, or the WebView's `<video>`,
+        // which is what actually plays while experimentalNativeVideo is off.
+        // (DR-160)
+        return isNativeVideoPath() || html5VideoActive
     }
 
     /**
@@ -125,32 +176,47 @@ object PictureInPictureManager {
         val player = try {
             JellyTauPlayer.getInstance()
         } catch (e: Exception) {
-            return null
+            null
         }
 
-        val surface = player.getSurfaceView() ?: return null
         // The surface has already been letterboxed to the video's aspect ratio
         // by fitSurfaceToScreen(), so its measured bounds are the video shape.
-        val width = surface.width
-        val height = surface.height
-        if (width <= 0 || height <= 0) return null
+        val surface = player?.getSurfaceView()
+        if (surface != null && surface.width > 0 && surface.height > 0) {
+            return clampedRatio(surface.width.toDouble() / surface.height.toDouble())
+        }
 
-        val ratio = width.toDouble() / height.toDouble()
-        val minRatio = 1.0 / 2.39
-        val maxRatio = 2.39
-        val clamped = ratio.coerceIn(minRatio, maxRatio)
+        // No native surface: the WebView is the video, so use the intrinsic size
+        // the frontend reported. (DR-160)
+        return html5AspectRatio
+    }
 
-        // Scale to integers; Rational(width, height) directly can overflow for
-        // large surfaces, and the clamped value may not match the raw pixels.
+    /**
+     * Clamp a ratio to the range Android accepts and express it as a [Rational].
+     *
+     * The platform rejects ratios outside roughly 1:2.39 - 2.39:1 with an
+     * IllegalArgumentException, which would otherwise take down the Activity on
+     * unusually tall or wide content. Scaled to integers because
+     * `Rational(width, height)` can overflow for large surfaces, and the clamped
+     * value may not match the raw pixels anyway.
+     */
+    private fun clampedRatio(ratio: Double): Rational {
+        val clamped = ratio.coerceIn(1.0 / 2.39, 2.39)
         return Rational((clamped * 1000).toInt(), 1000)
     }
 
     @RequiresApi(Build.VERSION_CODES.O)
     private fun buildPlayPauseAction(activity: Activity): RemoteAction {
-        val isPlaying = try {
-            JellyTauPlayer.getInstance().getExoPlayer().isPlaying
-        } catch (e: Exception) {
-            false
+        // On the HTML5 path ExoPlayer is idle, so its `isPlaying` is always false
+        // and the button would be stuck showing "Play" mid-playback. (DR-160)
+        val isPlaying = if (isNativeVideoPath()) {
+            try {
+                JellyTauPlayer.getInstance().getExoPlayer().isPlaying
+            } catch (e: Exception) {
+                false
+            }
+        } else {
+            html5VideoPlaying
         }
 
         val (iconRes, title, controlType, requestCode) = if (isPlaying) {
@@ -222,11 +288,20 @@ object PictureInPictureManager {
      */
     fun onPipModeChanged(activity: Activity, isInPipMode: Boolean) {
         if (isInPipMode) {
-            hideWebView(activity)
+            // Hiding the WebView is correct only when the video is *behind* it on
+            // the native surface. On the HTML5 path the WebView is the video, so
+            // hiding it would leave an empty black PiP window — the frontend
+            // instead strips its own chrome when it hears the event below.
+            // (DR-160)
+            if (isNativeVideoPath()) {
+                hideWebView(activity)
+            }
             registerReceiver(activity)
+            dispatchWebEvent(activity, "jellytau-pip-entered")
         } else {
             unregisterReceiver(activity)
             showWebView()
+            dispatchWebEvent(activity, "jellytau-pip-exited")
             // The surface was laid out against the tiny PiP bounds; re-fit it to
             // the restored full-screen bounds or the video stays postage-stamp sized.
             try {
@@ -237,6 +312,23 @@ object PictureInPictureManager {
         }
     }
 
+    /**
+     * Fire a DOM event into the WebView.
+     *
+     * The HTML5 PiP path is a conversation with the frontend rather than
+     * something native can do alone: it has to be told to strip its chrome when
+     * the window shrinks, and to play/pause the element. (DR-160)
+     */
+    private fun dispatchWebEvent(activity: Activity, name: String) {
+        val webView = findWebView(activity.window.decorView) ?: return
+        webView.post {
+            webView.evaluateJavascript(
+                "window.dispatchEvent(new CustomEvent('$name'));",
+                null
+            )
+        }
+    }
+
     private fun hideWebView(activity: Activity) {
         val webView = findWebView(activity.window.decorView)
         if (webView == null) {
@@ -264,14 +356,30 @@ object PictureInPictureManager {
         val r = object : BroadcastReceiver() {
             override fun onReceive(context: Context?, intent: Intent?) {
                 if (intent?.action != ACTION_MEDIA_CONTROL) return
-                val player = try {
-                    JellyTauPlayer.getInstance()
-                } catch (e: Exception) {
-                    return
-                }
-                when (intent.getIntExtra(EXTRA_CONTROL_TYPE, 0)) {
-                    CONTROL_PLAY -> player.play()
-                    CONTROL_PAUSE -> player.pause()
+                val control = intent.getIntExtra(EXTRA_CONTROL_TYPE, 0)
+
+                if (isNativeVideoPath()) {
+                    val player = try {
+                        JellyTauPlayer.getInstance()
+                    } catch (e: Exception) {
+                        return
+                    }
+                    when (control) {
+                        CONTROL_PLAY -> player.play()
+                        CONTROL_PAUSE -> player.pause()
+                    }
+                } else {
+                    // The WebView owns playback here, so the command has to reach
+                    // the `<video>` element. Driving ExoPlayer instead would do
+                    // nothing at all, which is what a PiP button on the HTML5 path
+                    // used to do. (DR-160)
+                    val name = when (control) {
+                        CONTROL_PLAY -> "jellytau-pip-play"
+                        CONTROL_PAUSE -> "jellytau-pip-pause"
+                        else -> return
+                    }
+                    dispatchWebEvent(activity, name)
+                    html5VideoPlaying = control == CONTROL_PLAY
                 }
                 // Swap the button to reflect the new state.
                 updatePipActions(activity)

src-tauri/android/src/main/java/com/dtourolle/jellytau/VideoOverlayManager.kt

@@ -1,7 +1,7 @@
 package com.dtourolle.jellytau
 
 import android.app.Activity
-import android.view.SurfaceView
+import android.view.TextureView
 import android.view.ViewGroup
 import android.widget.FrameLayout
 import com.dtourolle.jellytau.player.JellyTauPlayer
@@ -14,15 +14,18 @@ import com.dtourolle.jellytau.player.JellyTauPlayer
  */
 object VideoOverlayManager {
 
-    private var attachedSurfaceView: SurfaceView? = null
+    private var attachedSurfaceView: TextureView? = null
     private var contentLayoutListener: android.view.View.OnLayoutChangeListener? = null
     private var listenerContentView: ViewGroup? = null
 
     /**
-     * Attach the video SurfaceView to the Activity's content view.
+     * Attach the video view to the Activity's content view.
      *
-     * The SurfaceView is added at index 0 (bottom of z-order) so it renders
-     * behind the Tauri WebView, allowing Svelte controls to overlay on top.
+     * Added at index 0 (bottom of the z-order) so it renders behind the Tauri
+     * WebView, allowing the Svelte controls to overlay on top. Since DR-192 this
+     * is a TextureView, so "behind" is ordinary view z-order within one window
+     * rather than a separate surface punched through it — which is what makes
+     * the overlay above it repaint reliably.
      *
      * @param activity The Activity to attach the surface to
      */
@@ -77,16 +80,29 @@ object VideoOverlayManager {
     }
 
     /**
-     * Detach the video SurfaceView from the Activity's view hierarchy.
+     * Detach the video SurfaceView from the view hierarchy.
      *
-     * @param activity The Activity to detach the surface from
+     * Must be called on the main thread.
+     *
+     * This had **no callers at all**, which made [attachVideoSurface] one-way:
+     * `JellyTauPlayer.clearVideoSurface()` dropped its `surfaceView` reference
+     * without removing the view, so every native video left its SurfaceView
+     * parented to the content view for the life of the process and the next one
+     * added another beneath it. The stack was invisible while the WebView was
+     * opaque, and [isVideoSurfaceAttached] — which gates
+     * `PictureInPictureManager.canEnterPip` — stayed true forever afterwards.
+     *
+     * Removes from the view's *own* parent rather than looking the content view
+     * up from an Activity, so it cannot leave a view behind when the Activity
+     * has been recreated under it.
+     *
+     * TRACES: UR-003, UR-041 | DR-184
      */
-    fun detachVideoSurface(activity: Activity) {
+    fun detachVideoSurface() {
         try {
             removeLayoutListener()
             attachedSurfaceView?.let { surfaceView ->
-                val contentView = activity.window.decorView.findViewById<ViewGroup>(android.R.id.content)
-                contentView.removeView(surfaceView)
+                (surfaceView.parent as? ViewGroup)?.removeView(surfaceView)
                 attachedSurfaceView = null
                 android.util.Log.d("VideoOverlayManager", "Video surface detached from view hierarchy")
             }

src-tauri/android/src/main/java/com/dtourolle/jellytau/WindowInsetsBridge.kt

@@ -0,0 +1,140 @@
+package com.dtourolle.jellytau
+
+import android.app.Activity
+import android.webkit.JavascriptInterface
+import android.webkit.WebView
+import androidx.core.view.ViewCompat
+import androidx.core.view.WindowInsetsCompat
+
+/**
+ * Publishes the Activity's real window insets to the WebView as CSS custom
+ * properties.
+ *
+ * TRACES: UR-066 | IR-031, DR-112
+ *
+ * ## Why this is necessary
+ *
+ * MainActivity calls `enableEdgeToEdge()`, and the app targets SDK 36 —
+ * edge-to-edge is mandatory from SDK 35 and the opt-out is ignored from SDK 36
+ * — so the Tauri WebView always spans the whole window, underneath the status
+ * bar, the navigation/gesture bar and the display cutout.
+ *
+ * The web layer cannot discover that by itself. Android WebView maps only the
+ * **display cutout** into `env(safe-area-inset-*)` (and only with
+ * `viewport-fit=cover`); the status bar and navigation bar are never reported.
+ * Unlike iOS Safari there is no CSS-visible system-bar inset. So the frontend's
+ * `env()`-based padding evaluated to 0 on every device and the bottom nav
+ * rendered underneath the navigation bar.
+ *
+ * How badly that showed depended entirely on the device's navigation mode: a
+ * thin translucent gesture pill overlaps almost harmlessly, while a tall opaque
+ * 3-button bar swallows the nav outright.
+ *
+ * ## Contract with the frontend
+ *
+ * Insets are reported in **CSS pixels** (density-independent), because that is
+ * the unit CSS will use them in — dividing by `displayMetrics.density` here is
+ * what keeps the padding correct across screen densities.
+ *
+ * - **Push**: on every inset change (rotation, navigation-mode switch, PiP
+ *   enter/exit) the four `--jt-inset-*` custom properties are written onto
+ *   `document.documentElement` and `jellytau-insets-changed` is dispatched.
+ * - **Pull**: `AndroidInsets.get()` returns the same payload as JSON. Required
+ *   because the first inset pass normally lands before the SvelteKit document
+ *   exists, and a page load discards any inline style a push had set.
+ *
+ * See `src/lib/utils/safeArea.ts` and the `--safe-*` vars in `src/app.css`.
+ */
+object WindowInsetsBridge {
+
+    /** Latest insets in CSS pixels. Written on the main thread, read from the WebView binder thread. */
+    @Volatile
+    private var top = 0
+    @Volatile
+    private var right = 0
+    @Volatile
+    private var bottom = 0
+    @Volatile
+    private var left = 0
+
+    /** Cached so a WebView found later (or re-found on resume) can be primed. */
+    private var webView: WebView? = null
+
+    /**
+     * Start listening for window insets on [activity].
+     *
+     * Call from `onCreate` right after `enableEdgeToEdge()`. The listener
+     * returns the insets **unconsumed** so the WebView still receives them for
+     * its own display-cutout handling.
+     */
+    fun install(activity: Activity) {
+        val density = activity.resources.displayMetrics.density
+
+        ViewCompat.setOnApplyWindowInsetsListener(activity.window.decorView) { _, insets ->
+            // systemBars() covers the status bar and the navigation/gesture bar;
+            // displayCutout() covers notches and punch-holes, which in landscape
+            // land on a side edge that systemBars() does not describe.
+            val i = insets.getInsets(
+                WindowInsetsCompat.Type.systemBars() or WindowInsetsCompat.Type.displayCutout()
+            )
+
+            val toCssPx = { px: Int -> if (density > 0f) Math.round(px / density) else px }
+            top = toCssPx(i.top)
+            right = toCssPx(i.right)
+            bottom = toCssPx(i.bottom)
+            left = toCssPx(i.left)
+
+            android.util.Log.d(
+                "WindowInsetsBridge",
+                "insets (css px): top=$top right=$right bottom=$bottom left=$left"
+            )
+            push()
+
+            // Do NOT return CONSUMED - other views (and the WebView's own cutout
+            // handling) still need to see these.
+            insets
+        }
+
+        // The first pass may already have happened before the listener existed.
+        ViewCompat.requestApplyInsets(activity.window.decorView)
+    }
+
+    /**
+     * Adopt the WebView carrying the UI and push the current insets into it.
+     *
+     * Safe to call repeatedly (MainActivity re-finds the WebView on every
+     * resume): this only writes CSS properties, unlike `addJavascriptInterface`,
+     * which must run exactly once per WebView.
+     */
+    fun attachWebView(view: WebView) {
+        webView = view
+        push()
+    }
+
+    /** Current insets as JSON in CSS pixels — the payload `AndroidInsets.get()` returns. */
+    fun currentJson(): String =
+        """{"top":$top,"right":$right,"bottom":$bottom,"left":$left}"""
+
+    /** The `AndroidInsets` @JavascriptInterface object for the pull path. */
+    fun jsInterface(): Any = object : Any() {
+        @JavascriptInterface
+        fun get(): String = currentJson()
+    }
+
+    /** Write the custom properties into the live document and signal the change. */
+    private fun push() {
+        val view = webView ?: return
+        val js = """
+            (function() {
+              var s = document.documentElement.style;
+              s.setProperty('--jt-inset-top', '${top}px');
+              s.setProperty('--jt-inset-right', '${right}px');
+              s.setProperty('--jt-inset-bottom', '${bottom}px');
+              s.setProperty('--jt-inset-left', '${left}px');
+              window.dispatchEvent(new CustomEvent('jellytau-insets-changed'));
+            })();
+        """.trimIndent()
+
+        view.post { view.evaluateJavascript(js, null) }
+    }
+}

src-tauri/android/src/main/java/com/dtourolle/jellytau/player/CodecDetector.kt

@@ -1,7 +1,11 @@
 package com.dtourolle.jellytau.player
 
+import android.content.Context
 import android.media.MediaCodecList
 import android.util.Log
+import androidx.media3.common.AudioAttributes
+import androidx.media3.common.util.UnstableApi
+import androidx.media3.exoplayer.audio.AudioCapabilities
 
 /**
  * Detects hardware codec capabilities using MediaCodecList.
@@ -75,6 +79,36 @@ object CodecDetector {
         )
     }
 
+    /**
+     * Report how many channels the *current audio output route* can voice.
+     *
+     * This is a different question from "can this device decode 5.1", which the
+     * codec list already answers: a phone decodes a 5.1 AC-3 track happily and
+     * still has only two channels to play it out of. Left unreported, Jellyfin
+     * is free to direct-play the multichannel track, and what the user hears is
+     * device dependent — a failed AudioSink configuration (silence) or dialogue
+     * folded into the surround channels and lost.
+     *
+     * Returns 0 when there is no answer; Rust reads that as "unknown" and falls
+     * back to stereo rather than claiming a capability we have not observed.
+     */
+    @UnstableApi
+    fun detectMaxAudioChannels(context: Context): Int {
+        return try {
+            val capabilities = AudioCapabilities.getCapabilities(
+                context,
+                AudioAttributes.DEFAULT,
+                /* routedDevice= */ null
+            )
+            val channels = capabilities.maxChannelCount
+            Log.i(TAG, "Audio route max channel count: $channels")
+            channels.coerceAtLeast(0)
+        } catch (e: Exception) {
+            Log.e(TAG, "Error querying audio capabilities", e)
+            0
+        }
+    }
+
     /**
      * Map Android MIME types to Jellyfin codec names.
      *

src-tauri/android/src/main/java/com/dtourolle/jellytau/player/JellyTauPlaybackService.kt

@@ -119,6 +119,54 @@ class JellyTauPlaybackService : MediaSessionService() {
                 nativeOnMediaCommand("seek:$positionSeconds")
             }
 
+            // media3 seeks by more routes than seekTo(long), and the ones below
+            // reach the *real* ExoPlayer if they are not overridden — bypassing
+            // Rust entirely and operating on the handoff stream's relative
+            // timeline. That is the same mechanism as the truncation bug, reached
+            // by a different door.
+            //
+            // seekToDefaultPosition is deliberately swallowed rather than
+            // forwarded. Util.handlePlayButtonAction calls it on an ended or idle
+            // player and then calls play(); on a handoff stream the seek lands at
+            // stream zero — the point the screen was locked at — which is exactly
+            // the reported jump-back. Sending "seek:0.0" instead would be worse
+            // still, restarting the whole episode. Rust already owns what "play
+            // after the stream ended" means (truncation recovery, or advancing to
+            // the next episode), and the play() that follows reaches it, so the
+            // right move here is to not move at all.
+            //
+            // TRACES: UR-040, UR-005 | DR-159
+            override fun seekToDefaultPosition() {
+                android.util.Log.d(
+                    "JellyTauPlaybackService",
+                    "Ignoring seekToDefaultPosition — Rust owns end-of-stream handling"
+                )
+            }
+
+            override fun seekToDefaultPosition(mediaItemIndex: Int) {
+                android.util.Log.d(
+                    "JellyTauPlaybackService",
+                    "Ignoring seekToDefaultPosition(index) — Rust owns end-of-stream handling"
+                )
+            }
+
+            // `currentPosition` is ExoPlayer's own, so it is relative during a
+            // handoff; the base makes the target absolute, which is what Rust
+            // expects from every command on this boundary.
+            override fun seekBack() {
+                val target =
+                    ((currentPosition + handoffBaseMs - seekBackIncrement) / 1000.0)
+                        .coerceAtLeast(0.0)
+                nativeOnMediaCommand("seek:$target")
+            }
+
+            override fun seekForward() {
+                val target =
+                    ((currentPosition + handoffBaseMs + seekForwardIncrement) / 1000.0)
+                        .coerceAtLeast(0.0)
+                nativeOnMediaCommand("seek:$target")
+            }
+
             override fun stop() {
                 nativeOnMediaCommand("stop")
             }
@@ -262,23 +310,32 @@ class JellyTauPlaybackService : MediaSessionService() {
     private var lastArtist: String = ""
     private var lastIsPlaying: Boolean = false
 
-    // Base offset (ms) added to every position reported to the lockscreen
-    // MediaSession. During a background-audio handoff the audio stream is
-    // requested with StartTimeTicks = the handoff point, so ExoPlayer reports
-    // position RELATIVE to that point (starting at 0). The metadata duration,
-    // however, is the full absolute length — so without this base the scrubber
-    // thumb sits near 0:00 on a full-length bar. Set from the known handoff
-    // position via setPositionOffset(); 0 for normal playback.
-    private var positionOffsetMs: Long = 0L
+    // The handoff base (ms): during a background-audio handoff the audio stream is
+    // requested with StartTimeTicks = the handoff point, so ExoPlayer's timeline
+    // starts at 0 *there* and every position it reports is relative to it. This
+    // is the number that converts one back to a real position on the episode.
+    //
+    // It is deliberately read, not applied, here. This used to be a display-only
+    // correction added at the two setPlaybackState calls below, which left every
+    // other consumer — progress reporting to Jellyfin, the frontend, media3's own
+    // seeks — working in the relative timeline while treating it as absolute, each
+    // crossing silently losing exactly `base` seconds. The conversion now happens
+    // once, in JellyTauPlayer's position tick, so everything downstream of it
+    // speaks the episode's timeline; applying it again here would double-count.
+    //
+    // TRACES: UR-040 | DR-159
+    @Volatile
+    var handoffBaseMs: Long = 0L
+        private set
 
     /**
-     * Set the base position offset (seconds) applied to lockscreen positions.
-     * Called by the native layer when entering/exiting a background-audio handoff.
-     * Pass 0 to clear (normal playback, where ExoPlayer's position is absolute).
+     * Set the handoff base (seconds). Called by the native layer when entering or
+     * leaving a background-audio handoff; 0 clears it for normal playback, where
+     * ExoPlayer's position is already absolute.
      */
-    fun setPositionOffset(offsetSeconds: Double) {
-        positionOffsetMs = (offsetSeconds * 1000.0).toLong().coerceAtLeast(0L)
-        android.util.Log.d("JellyTauPlaybackService", "Position offset set to ${positionOffsetMs}ms")
+    fun setHandoffBase(offsetSeconds: Double) {
+        handoffBaseMs = (offsetSeconds * 1000.0).toLong().coerceAtLeast(0L)
+        android.util.Log.d("JellyTauPlaybackService", "Handoff base set to ${handoffBaseMs}ms")
     }
 
     /**
@@ -314,8 +371,9 @@ class JellyTauPlaybackService : MediaSessionService() {
 
         session.setMetadata(metadataBuilder.build())
 
-        // Update MediaSession playback state (position made absolute via the base offset).
-        session.setPlaybackState(buildPlaybackState(isPlaying, position + positionOffsetMs))
+        // Already absolute: this call comes from Rust, whose stored position is on
+        // the episode's timeline. (DR-159)
+        session.setPlaybackState(buildPlaybackState(isPlaying, position))
 
         // While casting, re-assert the remote volume provider. Metadata pushes
         // arrive on the session poller thread and can race with (or arrive
@@ -337,15 +395,15 @@ class JellyTauPlaybackService : MediaSessionService() {
      * notification. Without this, the lockscreen scrubber freezes at the position
      * from the last play/pause and drifts out of sync with actual playback.
      *
-     * @param position Position in milliseconds
+     * @param position Absolute position in milliseconds, on the item's own
+     *   timeline — the caller has already applied [handoffBaseMs].
      * @param isPlaying Whether playback is currently active
      */
     fun updatePlaybackPosition(position: Long, isPlaying: Boolean) {
         val session = mediaSessionCompat ?: return
         val notificationStateChanged = isPlaying != lastIsPlaying
         lastIsPlaying = isPlaying
-        // Absolute position for the scrubber = relative ExoPlayer position + base offset.
-        session.setPlaybackState(buildPlaybackState(isPlaying, position + positionOffsetMs))
+        session.setPlaybackState(buildPlaybackState(isPlaying, position))
         // Only rebuild the notification when the play/pause icon actually flips.
         if (notificationStateChanged) {
             updateNotification(lastTitle, lastArtist, isPlaying)

src-tauri/android/src/main/java/com/dtourolle/jellytau/player/JellyTauPlayer.kt

@@ -8,8 +8,7 @@ import android.net.Uri
 import android.os.Build
 import android.os.Handler
 import android.os.Looper
-import android.view.SurfaceHolder
-import android.view.SurfaceView
+import android.view.TextureView
 import android.view.ViewGroup
 import android.widget.FrameLayout
 import androidx.annotation.OptIn
@@ -36,6 +35,59 @@ class JellyTauPlayer(private val appContext: Context) {
         /** Position update interval in milliseconds */
         private const val POSITION_UPDATE_INTERVAL_MS = 250L
 
+        /** AudioEffect priority. Positive = higher priority than the default. */
+        private const val EFFECT_PRIORITY = 1000
+
+        /**
+         * How long to wait for a fresh frame after a resize before revealing the
+         * view anyway. Playback may be paused, in which case no frame is coming.
+         */
+        private const val FRESH_FRAME_TIMEOUT_MS = 400L
+
+        /**
+         * Canonical 10-band ISO centre frequencies (Hz), mirroring EQ_BANDS in
+         * settings.rs. Kept in sync deliberately: Rust owns the band layout, this
+         * is only the lookup table used to map those gains onto whatever bands
+         * the device's equalizer actually has.
+         */
+        private val CANONICAL_BAND_CENTRES_HZ =
+            intArrayOf(31, 62, 125, 250, 500, 1000, 2000, 4000, 8000, 16000)
+
+        /**
+         * Map canonical band gains onto a device's band centres by nearest
+         * centre frequency.
+         *
+         * Pure function so it can be unit-tested without a device — device band
+         * counts vary (commonly 5) and getting this wrong silently mis-shapes the
+         * EQ curve rather than failing.
+         *
+         * TRACES: UR-027 | DR-030
+         */
+        @JvmStatic
+        fun resampleBands(
+            canonicalGains: FloatArray,
+            canonicalCentresHz: IntArray,
+            deviceCentresHz: IntArray
+        ): FloatArray {
+            if (canonicalGains.isEmpty() || deviceCentresHz.isEmpty()) {
+                return FloatArray(deviceCentresHz.size)
+            }
+            val usable = minOf(canonicalGains.size, canonicalCentresHz.size)
+            return FloatArray(deviceCentresHz.size) { d ->
+                val target = deviceCentresHz[d]
+                var nearest = 0
+                var bestDelta = Int.MAX_VALUE
+                for (c in 0 until usable) {
+                    val delta = kotlin.math.abs(canonicalCentresHz[c] - target)
+                    if (delta < bestDelta) {
+                        bestDelta = delta
+                        nearest = c
+                    }
+                }
+                canonicalGains[nearest]
+            }
+        }
+
         /** Singleton instance for JNI access */
         @Volatile
         private var instance: JellyTauPlayer? = null
@@ -83,7 +135,7 @@ class JellyTauPlayer(private val appContext: Context) {
                         }
 
                         // Detect and report hardware codec capabilities to Rust
-                        detectAndReportCodecs()
+                        detectAndReportCodecs(context.applicationContext)
                     }
                 }
             }
@@ -94,23 +146,31 @@ class JellyTauPlayer(private val appContext: Context) {
          * Called during player initialization.
          */
         @JvmStatic
-        fun detectAndReportCodecs() {
+        fun detectAndReportCodecs(context: Context) {
             val capabilities = CodecDetector.detectHardwareCodecs()
 
             // Convert lists to comma-separated strings for JNI transfer
             val videoCodecsStr = capabilities.videoCodecs.joinToString(",")
             val audioCodecsStr = capabilities.audioCodecs.joinToString(",")
 
-            android.util.Log.i("JellyTauPlayer", "Reporting codecs to Rust: video=$videoCodecsStr, audio=$audioCodecsStr")
+            // What the route can *voice*, which the codec list does not answer.
+            // 0 means "no answer"; Rust falls back to stereo.
+            val maxAudioChannels = CodecDetector.detectMaxAudioChannels(context)
+
+            android.util.Log.i("JellyTauPlayer", "Reporting codecs to Rust: video=$videoCodecsStr, audio=$audioCodecsStr, maxAudioChannels=$maxAudioChannels")
 
             // Call native method to store in Rust
-            nativeOnCodecsDetected(videoCodecsStr, audioCodecsStr)
+            nativeOnCodecsDetected(videoCodecsStr, audioCodecsStr, maxAudioChannels)
         }
 
         /**
          * Native method to report detected codecs to Rust.
          */
-        private external fun nativeOnCodecsDetected(videoCodecs: String, audioCodecs: String)
+        private external fun nativeOnCodecsDetected(
+            videoCodecs: String,
+            audioCodecs: String,
+            maxAudioChannels: Int
+        )
 
         /**
          * Check if the player is initialized.
@@ -135,6 +195,18 @@ class JellyTauPlayer(private val appContext: Context) {
     private val coroutineScope = CoroutineScope(Dispatchers.Main + SupervisorJob())
     private var positionUpdateJob: Job? = null
 
+    /** Graphic EQ bound to the current audio session, or null if not attached. */
+    private var equalizer: android.media.audiofx.Equalizer? = null
+
+    /** Loudness/normalization effect bound to the current audio session. */
+    private var loudnessEnhancer: android.media.audiofx.LoudnessEnhancer? = null
+
+    /**
+     * Last settings pushed from Rust, replayed when the audio session is rebuilt.
+     * Held as the raw payload so re-application needs no second parse contract.
+     */
+    private var lastAudioSettings: org.json.JSONObject? = null
+
     /** Current media ID being played */
     private var currentMediaId: String? = null
 
@@ -158,9 +230,18 @@ class JellyTauPlayer(private val appContext: Context) {
     /** Media type enum */
     enum class MediaType { AUDIO, VIDEO }
 
-    /** SurfaceView for video playback */
-    private var surfaceView: SurfaceView? = null
-    private var surfaceHolder: SurfaceHolder? = null
+    /** TextureView for video playback — see getOrCreateSurfaceView() for why. */
+    private var videoView: TextureView? = null
+
+    /** The Surface handed to ExoPlayer, owned here rather than by the player. */
+    private var videoSurface: android.view.Surface? = null
+
+    /**
+     * True while the view is hidden waiting for a new frame after a resize.
+     * See fitSurfaceToScreen (DR-194).
+     */
+    @Volatile
+    private var awaitingFreshFrame = false
     /** Last reported video frame size, used to fit the surface to the screen preserving aspect ratio */
     private var videoWidth: Int = 0
     private var videoHeight: Int = 0
@@ -173,6 +254,20 @@ class JellyTauPlayer(private val appContext: Context) {
     }
     private var audioFocusRequest: AudioFocusRequest? = null
 
+    /**
+     * Set when a video was loaded but audio focus was not granted outright.
+     *
+     * `setAcceptsDelayedFocusGain(true)` means the system may answer DELAYED and
+     * hand us focus later; until then it withholds our audio. Starting playback
+     * anyway plays the video silently, which is exactly the "video has no sound"
+     * symptom. We hold playback and start it from the AUDIOFOCUS_GAIN callback.
+     */
+    private var pendingPlayOnFocusGain = false
+
+    /** Whether we currently hold audio focus, so `play()` does not re-request
+     *  (and leak) a focus request we already own. */
+    private var hasAudioFocus = false
+
     init {
         // Configure audio attributes for music playback with audio focus handling
         val audioAttributes = AudioAttributes.Builder()
@@ -302,29 +397,61 @@ class JellyTauPlayer(private val appContext: Context) {
                     android.util.Log.d("JellyTauPlayer", "    Video group: ${group.length} tracks, selected=${group.isSelected}")
                 }
 
-                // CRITICAL FIX: Auto-select first audio track if none is selected
-                // This fixes the issue where some videos play without audio
+                // TRACES: UR-004 | DR-146
+                // Auto-select an audio track if none is selected — some videos
+                // otherwise play with no sound. Pick a track the renderer says it
+                // *supports*: when nothing was selected because track 0 failed to
+                // initialize, forcing track 0 again just reinstates the silence.
                 if (!hasSelectedAudio && audioTracks.isNotEmpty() && currentMediaType == MediaType.VIDEO) {
-                    android.util.Log.w("JellyTauPlayer", "⚠️  NO AUDIO TRACK SELECTED! Auto-selecting first audio track...")
+                    android.util.Log.w("JellyTauPlayer", "⚠️  NO AUDIO TRACK SELECTED! Looking for a supported audio track...")
 
                     val trackSelector = exoPlayer.trackSelector
                     if (trackSelector != null) {
                         try {
-                            // Select the first audio track group
-                            val firstAudioGroup = audioTracks[0]
-                            val override = androidx.media3.common.TrackSelectionOverride(
-                                firstAudioGroup.mediaTrackGroup,
-                                0 // Select the first track in this group
-                            )
+                            var chosenGroup: androidx.media3.common.Tracks.Group? = null
+                            var chosenIndex = -1
+                            outer@ for (group in audioTracks) {
+                                for (i in 0 until group.length) {
+                                    if (group.isTrackSupported(i)) {
+                                        chosenGroup = group
+                                        chosenIndex = i
+                                        break@outer
+                                    }
+                                }
+                            }
 
-                            val parameters = trackSelector.parameters
-                                .buildUpon()
-                                .clearOverridesOfType(C.TRACK_TYPE_AUDIO)
-                                .addOverride(override)
-                                .build()
+                            if (chosenGroup == null) {
+                                // Every track is undecodable on this device. The
+                                // server should have transcoded; say so loudly
+                                // rather than leaving a silent video unexplained.
+                                android.util.Log.e(
+                                    "JellyTauPlayer",
+                                    "✗ No supported audio track in ${audioTracks.size} group(s) - " +
+                                        "device cannot decode any of them, expected a transcode"
+                                )
+                            } else {
+                                val format = chosenGroup.getTrackFormat(chosenIndex)
+                                val override = androidx.media3.common.TrackSelectionOverride(
+                                    chosenGroup.mediaTrackGroup,
+                                    chosenIndex
+                                )
 
-                            trackSelector.setParameters(parameters)
-                            android.util.Log.d("JellyTauPlayer", "✓ Auto-selected first audio track")
+                                val parameters = trackSelector.parameters
+                                    .buildUpon()
+                                    // Audio may also be off because the track type
+                                    // was disabled; an override alone would not
+                                    // bring it back.
+                                    .setTrackTypeDisabled(C.TRACK_TYPE_AUDIO, false)
+                                    .clearOverridesOfType(C.TRACK_TYPE_AUDIO)
+                                    .addOverride(override)
+                                    .build()
+
+                                trackSelector.setParameters(parameters)
+                                android.util.Log.d(
+                                    "JellyTauPlayer",
+                                    "✓ Auto-selected supported audio track $chosenIndex (${format.sampleMimeType}, ${format.channelCount}ch)"
+                                )
+                            }
                         } catch (e: Exception) {
                             android.util.Log.e("JellyTauPlayer", "Failed to auto-select audio track", e)
                         }
@@ -334,6 +461,11 @@ class JellyTauPlayer(private val appContext: Context) {
 
             override fun onAudioSessionIdChanged(audioSessionId: Int) {
                 android.util.Log.d("JellyTauPlayer", "▶▶▶ AUDIO SESSION ID CHANGED: $audioSessionId")
+                // ExoPlayer rebuilt its audio sink (e.g. on a format change), so
+                // effects bound to the old session are dead. Re-attach, or the EQ
+                // silently stops applying mid-queue.
+                releaseAudioEffects()
+                lastAudioSettings?.let { applyAudioEffects(it) }
             }
         })
     }
@@ -359,6 +491,15 @@ class JellyTauPlayer(private val appContext: Context) {
      */
     fun play() {
         mainHandler.post {
+            // Video manages focus by hand, so an explicit play after a refusal (or
+            // after a LOSS paused us) has to ask again — otherwise it resumes into
+            // a stream the system is still muting.
+            if (currentMediaType == MediaType.VIDEO && !hasAudioFocus && !requestAudioFocus()) {
+                pendingPlayOnFocusGain = true
+                android.util.Log.d("JellyTauPlayer", "play() without audio focus - holding until GAIN")
+                return@post
+            }
+            pendingPlayOnFocusGain = false
             exoPlayer.play()
         }
     }
@@ -416,6 +557,133 @@ class JellyTauPlayer(private val appContext: Context) {
         }
     }
 
+    /**
+     * Apply audio settings pushed from Rust as JSON.
+     *
+     * Rust owns *what* the values are (band layout, preset curves, normalization
+     * presets); this owns *when* the Android AudioEffect objects exist, since
+     * that needs the live audio session id and must survive a sink rebuild.
+     *
+     * Posted to the main handler rather than run inline: AudioEffect construction
+     * from a player callback can re-enter the player and deadlock.
+     *
+     * TRACES: UR-027, UR-032, UR-033 | DR-030, DR-035, DR-036
+     */
+    fun setAudioSettings(json: String) {
+        mainHandler.post {
+            try {
+                val settings = org.json.JSONObject(json)
+                lastAudioSettings = settings
+
+                // Gapless: ExoPlayer is gapless by default for compatible
+                // formats, so honouring the setting means disabling it when off.
+                exoPlayer.pauseAtEndOfMediaItems = !settings.optBoolean("gaplessPlayback", true)
+
+                applyAudioEffects(settings)
+            } catch (e: Exception) {
+                android.util.Log.e("JellyTauPlayer", "Failed to apply audio settings", e)
+            }
+        }
+    }
+
+    /** Attach/update the EQ and loudness effects for the current audio session. */
+    private fun applyAudioEffects(settings: org.json.JSONObject) {
+        val sessionId = exoPlayer.audioSessionId
+        if (sessionId == C.AUDIO_SESSION_ID_UNSET) {
+            // No sink yet; onAudioSessionIdChanged will re-drive this.
+            return
+        }
+
+        try {
+            applyEqualizer(sessionId, settings)
+        } catch (e: Exception) {
+            android.util.Log.e("JellyTauPlayer", "Equalizer unavailable on this device", e)
+        }
+
+        try {
+            applyNormalization(sessionId, settings)
+        } catch (e: Exception) {
+            android.util.Log.e("JellyTauPlayer", "LoudnessEnhancer unavailable on this device", e)
+        }
+    }
+
+    private fun applyEqualizer(sessionId: Int, settings: org.json.JSONObject) {
+        val enabled = settings.optBoolean("equalizerEnabled", false)
+
+        if (!enabled) {
+            equalizer?.enabled = false
+            return
+        }
+
+        val eq = equalizer ?: android.media.audiofx.Equalizer(EFFECT_PRIORITY, sessionId).also {
+            equalizer = it
+        }
+
+        val bandsJson = settings.optJSONArray("equalizerBands")
+        val canonicalGains = FloatArray(bandsJson?.length() ?: 0) { i ->
+            bandsJson!!.optDouble(i, 0.0).toFloat()
+        }
+        if (canonicalGains.isEmpty()) {
+            eq.enabled = false
+            return
+        }
+
+        // The device's band count/centres are device-dependent (commonly 5) and
+        // will not match our canonical 10-band ISO layout, so resample.
+        val deviceBandCount = eq.numberOfBands.toInt()
+        val deviceCentresHz = IntArray(deviceBandCount) { i ->
+            eq.getCenterFreq(i.toShort()) / 1000 // device reports milliHertz
+        }
+        val levelRange = eq.bandLevelRange // millibels, [min, max]
+
+        val resampled = resampleBands(canonicalGains, CANONICAL_BAND_CENTRES_HZ, deviceCentresHz)
+
+        for (i in 0 until deviceBandCount) {
+            val millibels = (resampled[i] * 100f)
+                .coerceIn(levelRange[0].toFloat(), levelRange[1].toFloat())
+            eq.setBandLevel(i.toShort(), millibels.toInt().toShort())
+        }
+        eq.enabled = true
+    }
+
+    private fun applyNormalization(sessionId: Int, settings: org.json.JSONObject) {
+        val enabled = settings.optBoolean("normalizeVolume", false)
+
+        if (!enabled) {
+            loudnessEnhancer?.enabled = false
+            return
+        }
+
+        val enhancer = loudnessEnhancer
+            ?: android.media.audiofx.LoudnessEnhancer(sessionId).also { loudnessEnhancer = it }
+
+        // Approximate parity with the Linux dynaudnorm path: LoudnessEnhancer is
+        // a gain stage, not a true EBU R128 normalizer, so these are relative
+        // offsets preserving the Loud > Normal > Quiet ordering.
+        val targetGainMb = when (settings.optString("volumeLevel", "normal")) {
+            "loud" -> 600
+            "quiet" -> -600
+            else -> 0
+        }
+        enhancer.setTargetGain(targetGainMb)
+        enhancer.enabled = true
+    }
+
+    private fun releaseAudioEffects() {
+        try {
+            equalizer?.release()
+        } catch (e: Exception) {
+            android.util.Log.w("JellyTauPlayer", "Equalizer release failed", e)
+        }
+        try {
+            loudnessEnhancer?.release()
+        } catch (e: Exception) {
+            android.util.Log.w("JellyTauPlayer", "LoudnessEnhancer release failed", e)
+        }
+        equalizer = null
+        loudnessEnhancer = null
+    }
+
     /**
      * Get the current playback position in seconds.
      */
@@ -620,14 +888,17 @@ class JellyTauPlayer(private val appContext: Context) {
             android.util.Log.d("JellyTauPlayer", "ExoPlayer audio session ID: ${exoPlayer.audioSessionId}")
 
             // Setup video surface if needed
+            var focusGranted = true
             if (currentMediaType == MediaType.VIDEO) {
                 getOrCreateSurfaceView()
                 android.util.Log.d("JellyTauPlayer", "Video surface created for playback")
                 // Automatically attach the surface to the Activity
                 autoAttachSurface()
 
-                // CRITICAL: Request audio focus for video playback
-                requestAudioFocus()
+                // CRITICAL: Request audio focus for video playback. Video manages
+                // focus by hand (handleAudioFocus=false above), so nothing else
+                // will hold playback back if the request is delayed or refused.
+                focusGranted = requestAudioFocus()
             } else {
                 clearVideoSurface()
                 // Abandon audio focus when switching to audio (audio uses ExoPlayer's built-in handling)
@@ -697,8 +968,13 @@ class JellyTauPlayer(private val appContext: Context) {
             }
             android.util.Log.d("JellyTauPlayer", "✓ Current volume: ${exoPlayer.volume}, deviceVolume: ${audioManager.getStreamVolume(AudioManager.STREAM_MUSIC)}/${audioManager.getStreamMaxVolume(AudioManager.STREAM_MUSIC)}")
 
-            exoPlayer.playWhenReady = true
-            android.util.Log.d("JellyTauPlayer", "playWhenReady set to TRUE. Current state: ${exoPlayer.playbackState}")
+            // Only roll if we hold audio focus. A DELAYED grant means the system
+            // is withholding our audio until it calls back with AUDIOFOCUS_GAIN;
+            // playing through it produces picture with no sound. Playback resumes
+            // from the focus listener instead.
+            pendingPlayOnFocusGain = !focusGranted
+            exoPlayer.playWhenReady = focusGranted
+            android.util.Log.d("JellyTauPlayer", "playWhenReady set to $focusGranted (pendingPlayOnFocusGain=$pendingPlayOnFocusGain). Current state: ${exoPlayer.playbackState}")
 
             // Start the foreground service for lockscreen controls
             startPlaybackService()
@@ -756,6 +1032,7 @@ class JellyTauPlayer(private val appContext: Context) {
         mainHandler.post {
             stopPositionUpdates()
             coroutineScope.cancel()
+            releaseAudioEffects()
             exoPlayer.release()
             instance = null
         }
@@ -767,16 +1044,41 @@ class JellyTauPlayer(private val appContext: Context) {
             android.util.Log.d("JellyTauPlayer", "Started position updates coroutine")
             while (isActive) {
                 if (exoPlayer.isPlaying) {
-                    val positionMs = exoPlayer.currentPosition.coerceAtLeast(0)
+                    // THE boundary between the two timelines, and the only place
+                    // the conversion happens.
+                    //
+                    // During a background-audio handoff the stream is requested
+                    // with StartTimeTicks = the handoff point, so ExoPlayer's zero
+                    // is that point and everything it reports is relative to it.
+                    // The base used to be added only where a position was *shown*
+                    // (the lockscreen scrubber), leaving progress reports to
+                    // Jellyfin, the frontend and the truncation maths all working
+                    // in the relative timeline while treating it as absolute —
+                    // each crossing losing exactly `base` seconds, which is why the
+                    // jump-back distance varied with where the screen was locked.
+                    // Shifting once, here, means every consumer downstream speaks
+                    // the episode's timeline and none of them needs to know a
+                    // handoff happened.
+                    //
+                    // The duration is shifted with it, so position and duration
+                    // stay on the same timeline — the stream's own length is only
+                    // what remains after the handoff point.
+                    //
+                    // TRACES: UR-040 | DR-159
+                    val service = JellyTauPlaybackService.getInstance()
+                    val baseMs = service?.handoffBaseMs ?: 0L
+
+                    val positionMs = exoPlayer.currentPosition.coerceAtLeast(0) + baseMs
                     val position = positionMs / 1000.0
-                    val duration = if (exoPlayer.duration > 0) exoPlayer.duration / 1000.0 else 0.0
+                    val duration =
+                        if (exoPlayer.duration > 0) (exoPlayer.duration + baseMs) / 1000.0 else 0.0
                     android.util.Log.v("JellyTauPlayer", "Position update: $position / $duration")
                     nativeOnPositionUpdate(position, duration)
 
                     // Keep the lockscreen scrubber live. Without this the
                     // MediaSession position only refreshes on play/pause, so the
                     // scrubber freezes mid-track and drifts out of sync.
-                    JellyTauPlaybackService.getInstance()?.updatePlaybackPosition(positionMs, true)
+                    service?.updatePlaybackPosition(positionMs, true)
                 }
                 delay(POSITION_UPDATE_INTERVAL_MS)
             }
@@ -790,52 +1092,114 @@ class JellyTauPlayer(private val appContext: Context) {
     }
 
     /**
-     * Get or create the SurfaceView for video playback.
-     * Returns the view ID that can be attached to the view hierarchy.
+     * Get or create the video view, and hand it to ExoPlayer.
      *
-     * Note: The surface is created but not automatically attached to the view hierarchy.
-     * Call attachSurfaceToActivity() or use VideoOverlayManager to attach it.
+     * This is a **TextureView**, not a SurfaceView, and that is the whole point.
+     *
+     * A SurfaceView renders on its own layer *outside* the app window and punches
+     * a transparent hole through the window to show it. Anything drawn above
+     * that hole — for us, the entire Svelte UI in a transparent WebView — is at
+     * the mercy of that composition path, and Android's own graphics
+     * documentation says plainly that "overlays do not currently work correctly
+     * with SurfaceView or TextureView". On device that showed up as the WebView
+     * overlay silently dropping its incremental damage: the clock text stopped
+     * advancing on screen while the DOM kept updating (slider 476 → 479 across
+     * three seconds behind a display showing neither), the control bar would not
+     * fade, and rotation lost the transport UI. Only *structural* DOM changes
+     * got through, which is why the play overlay — an `{#if}` block that is added
+     * and removed — always appeared to work while the progress bar never did.
+     *
+     * A TextureView is an ordinary view: its frames are drawn as a texture inside
+     * the window's normal rendering pass, so there is no second layer, no
+     * transparent region, and the WebView above composites like it would over any
+     * other view. This is the standard remedy for ExoPlayer overlay problems and
+     * is why media3 offers `surface_type="texture_view"` at all.
+     *
+     * The cost is real and accepted: TextureView uses more power and memory than
+     * SurfaceView and adds a frame of latency. Hardware decode through MediaCodec
+     * is unaffected — only presentation changes — so the reason native video
+     * exists survives the trade.
+     *
+     * `setVideoTextureView` installs ExoPlayer's own `SurfaceTextureListener`, so
+     * there is deliberately no listener of ours here; adding one would displace
+     * it and the video would never appear.
+     *
+     * Note: the view is created but not attached to the hierarchy. Call
+     * attachSurfaceToActivity() or use VideoOverlayManager to attach it.
+     *
+     * TRACES: UR-003, UR-004 | DR-192
      */
     fun getOrCreateSurfaceView(): Int {
-        if (surfaceView == null) {
-            surfaceView = SurfaceView(appContext).apply {
+        if (videoView == null) {
+            videoView = TextureView(appContext).apply {
                 layoutParams = FrameLayout.LayoutParams(
                     ViewGroup.LayoutParams.MATCH_PARENT,
                     ViewGroup.LayoutParams.MATCH_PARENT
                 )
-                // Render BEHIND WebView - video shows through transparent areas
-                setZOrderMediaOverlay(false)
+                // The view is opaque where video is drawn; the WebView above it
+                // is what supplies transparency, exactly as before.
+                isOpaque = true
 
-                // Set up SurfaceHolder callbacks
-                holder.addCallback(object : SurfaceHolder.Callback {
-                    override fun surfaceCreated(holder: SurfaceHolder) {
-                        android.util.Log.d("JellyTauPlayer", "Surface created")
-                        surfaceHolder = holder
-                        exoPlayer.setVideoSurfaceHolder(holder)
+                // Own the listener rather than calling `setVideoTextureView`,
+                // which installs ExoPlayer's own and leaves us blind to frame
+                // arrival. `onSurfaceTextureUpdated` is the only honest signal
+                // that a NEW frame has landed in the texture, and that is
+                // precisely what the letterbox artefact waits on — see
+                // fitSurfaceToScreen. Handing ExoPlayer the Surface directly is
+                // the same wiring `setVideoTextureView` does internally.
+                //
+                // TRACES: UR-003, UR-004 | DR-194
+                surfaceTextureListener = object : TextureView.SurfaceTextureListener {
+                    override fun onSurfaceTextureAvailable(
+                        texture: android.graphics.SurfaceTexture,
+                        width: Int,
+                        height: Int
+                    ) {
+                        videoSurface?.release()
+                        videoSurface = android.view.Surface(texture)
+                        exoPlayer.setVideoSurface(videoSurface)
                         android.util.Log.d("JellyTauPlayer", "Video surface attached to ExoPlayer")
                     }
 
-                    override fun surfaceChanged(holder: SurfaceHolder, format: Int, width: Int, height: Int) {
-                        android.util.Log.d("JellyTauPlayer", "Surface changed: ${width}x${height}")
+                    override fun onSurfaceTextureSizeChanged(
+                        texture: android.graphics.SurfaceTexture,
+                        width: Int,
+                        height: Int
+                    ) {
                     }
 
-                    override fun surfaceDestroyed(holder: SurfaceHolder) {
-                        android.util.Log.d("JellyTauPlayer", "Surface destroyed")
-                        exoPlayer.clearVideoSurfaceHolder(holder)
-                        surfaceHolder = null
+                    override fun onSurfaceTextureDestroyed(
+                        texture: android.graphics.SurfaceTexture
+                    ): Boolean {
+                        exoPlayer.setVideoSurface(null)
+                        videoSurface?.release()
+                        videoSurface = null
+                        return true
                     }
-                })
+
+                    override fun onSurfaceTextureUpdated(
+                        texture: android.graphics.SurfaceTexture
+                    ) {
+                        // A genuinely new frame is now in the texture, so
+                        // whatever was retained from before the resize is gone.
+                        if (awaitingFreshFrame) {
+                            awaitingFreshFrame = false
+                            videoView?.alpha = 1f
+                        }
+                    }
+                }
             }
+            android.util.Log.d("JellyTauPlayer", "Video TextureView created")
         }
-        return surfaceView!!.hashCode()
+        return videoView!!.hashCode()
     }
 
     /**
-     * Get the SurfaceView instance (for VideoOverlayManager).
-     * Returns null if no surface has been created yet.
+     * Get the video view instance (for VideoOverlayManager).
+     * Returns null if none has been created yet.
      */
-    fun getSurfaceView(): SurfaceView? {
-        return surfaceView
+    fun getSurfaceView(): TextureView? {
+        return videoView
     }
 
     /**
@@ -850,7 +1214,7 @@ class JellyTauPlayer(private val appContext: Context) {
      * This should be called from MainActivity when video playback is active.
      */
     fun attachSurfaceToActivity(activity: android.app.Activity) {
-        if (surfaceView != null && currentMediaType == MediaType.VIDEO) {
+        if (videoView != null && currentMediaType == MediaType.VIDEO) {
             com.dtourolle.jellytau.VideoOverlayManager.attachVideoSurface(activity)
             android.util.Log.d("JellyTauPlayer", "Surface attached to Activity")
         }
@@ -894,9 +1258,34 @@ class JellyTauPlayer(private val appContext: Context) {
      * pillarbox). A raw SurfaceView with MATCH_PARENT otherwise stretches the
      * video to the surface bounds, which crops the bottom on rotation.
      */
+    /**
+     * Hide the video view now, and keep it hidden until a genuinely new frame
+     * arrives (or the timeout fires).
+     *
+     * Called from `MainActivity.onConfigurationChanged`, i.e. at the *start* of a
+     * rotation. [fitSurfaceToScreen] is too late for this: it runs from the
+     * content view's layout listener, after the rotation has already happened,
+     * so the stale frame has been on screen for the whole transition by then.
+     *
+     * TRACES: UR-003, UR-066 | DR-194
+     */
+    fun hideUntilFreshFrame() {
+        mainHandler.post {
+            val view = videoView ?: return@post
+            awaitingFreshFrame = true
+            view.alpha = 0f
+            mainHandler.postDelayed({
+                if (awaitingFreshFrame) {
+                    awaitingFreshFrame = false
+                    videoView?.alpha = 1f
+                }
+            }, FRESH_FRAME_TIMEOUT_MS)
+        }
+    }
+
     fun fitSurfaceToScreen() {
         mainHandler.post {
-            val view = surfaceView ?: return@post
+            val view = videoView ?: return@post
             val parent = view.parent as? ViewGroup
             // Available area: prefer the parent's measured size, fall back to the screen.
             val availW = parent?.width?.takeIf { it > 0 }
@@ -928,6 +1317,44 @@ class JellyTauPlayer(private val appContext: Context) {
             if (lp is FrameLayout.LayoutParams) {
                 lp.gravity = android.view.Gravity.CENTER
             }
+
+            // Hide the view across a resize, and reveal it when a genuinely NEW
+            // video frame lands in the texture.
+            //
+            // A TextureView retains its last frame. Between a rotation and this
+            // re-fit landing, that retained frame is stretched across the OLD
+            // rect — larger than the new one along at least one axis — so the
+            // previous frame flashes in what should be the letterbox bars.
+            //
+            // Waiting a fixed number of animation frames does NOT fix it, which
+            // the first attempt at this proved on device: an animation frame is
+            // not a video frame, and at 24fps the next decoded frame can be
+            // several vsyncs away. The tell was that pausing and playing cleared
+            // the artefact by hand — that forces a fresh frame, which is the
+            // real precondition. So the reveal is driven by
+            // `onSurfaceTextureUpdated` instead.
+            //
+            // The timeout is not belt-and-braces, it is required: if playback is
+            // paused when the resize happens, no new frame is coming and the
+            // video would stay invisible forever. Revealing a stale frame after
+            // a beat is strictly better than a permanently black player.
+            //
+            // Scoped to an actual size change so steady-state playback never
+            // touches alpha.
+            //
+            // TRACES: UR-003, UR-066 | DR-194
+            val sizeChanged = lp.width != targetW || lp.height != targetH
+            if (sizeChanged) {
+                awaitingFreshFrame = true
+                view.alpha = 0f
+                mainHandler.postDelayed({
+                    if (awaitingFreshFrame) {
+                        awaitingFreshFrame = false
+                        videoView?.alpha = 1f
+                    }
+                }, FRESH_FRAME_TIMEOUT_MS)
+            }
+
             lp.width = targetW
             lp.height = targetH
             view.layoutParams = lp
@@ -940,22 +1367,38 @@ class JellyTauPlayer(private val appContext: Context) {
     }
 
     /**
-     * Clear the video surface when switching to audio playback.
+     * Clear the video surface when switching to audio playback, or on stop.
+     *
+     * Detaching is not optional bookkeeping: dropping the reference without
+     * removing the view left the SurfaceView parented to the content view for
+     * the life of the process, and the next video stacked another one under it.
+     * See VideoOverlayManager.detachVideoSurface.
+     *
+     * Always called on the main thread (every caller runs inside a
+     * `mainHandler.post`), which is what touching the view hierarchy requires.
+     *
+     * TRACES: UR-003, UR-041 | DR-184
      */
     private fun clearVideoSurface() {
-        surfaceView?.let {
+        videoView?.let {
             exoPlayer.clearVideoSurface()
-            surfaceView = null
-            surfaceHolder = null
-            android.util.Log.d("JellyTauPlayer", "Video surface cleared")
+            com.dtourolle.jellytau.VideoOverlayManager.detachVideoSurface()
+            videoView = null
+            android.util.Log.d("JellyTauPlayer", "Video surface cleared and detached")
         }
     }
 
     /**
      * Request audio focus for video playback.
      * This is critical for video to have audio on Android.
+     *
+     * TRACES: UR-004 | DR-145
+     *
+     * @return true if focus was granted outright and playback may start now.
+     *   false for a DELAYED or refused request — the caller must hold playback
+     *   and let the AUDIOFOCUS_GAIN callback start it, or the video plays mute.
      */
-    private fun requestAudioFocus() {
+    private fun requestAudioFocus(): Boolean {
         android.util.Log.d("JellyTauPlayer", "Requesting audio focus for video playback")
 
         if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
@@ -972,17 +1415,29 @@ class JellyTauPlayer(private val appContext: Context) {
                     when (focusChange) {
                         AudioManager.AUDIOFOCUS_GAIN -> {
                             android.util.Log.d("JellyTauPlayer", "✓ Audio focus GAINED - ensuring full volume")
+                            hasAudioFocus = true
                             if (exoPlayer.volume < 1.0f) {
                                 exoPlayer.volume = 1.0f
                                 android.util.Log.d("JellyTauPlayer", "  Volume restored to 1.0 from ${exoPlayer.volume}")
                             }
+                            // A delayed grant arriving: this is the point at which
+                            // the video may actually be heard, so start it now.
+                            if (pendingPlayOnFocusGain) {
+                                pendingPlayOnFocusGain = false
+                                android.util.Log.d("JellyTauPlayer", "  Delayed focus granted - starting held playback")
+                                exoPlayer.playWhenReady = true
+                            }
                         }
                         AudioManager.AUDIOFOCUS_LOSS -> {
                             android.util.Log.d("JellyTauPlayer", "Audio focus LOST - pausing")
+                            hasAudioFocus = false
+                            pendingPlayOnFocusGain = false
                             pause()
                         }
                         AudioManager.AUDIOFOCUS_LOSS_TRANSIENT -> {
                             android.util.Log.d("JellyTauPlayer", "Audio focus LOST TRANSIENT - pausing")
+                            hasAudioFocus = false
+                            pendingPlayOnFocusGain = false
                             pause()
                         }
                         AudioManager.AUDIOFOCUS_LOSS_TRANSIENT_CAN_DUCK -> {
@@ -993,16 +1448,25 @@ class JellyTauPlayer(private val appContext: Context) {
                 }
                 .build()
 
-            val result = audioManager.requestAudioFocus(audioFocusRequest!!)
-            when (result) {
+            return when (val result = audioManager.requestAudioFocus(audioFocusRequest!!)) {
                 AudioManager.AUDIOFOCUS_REQUEST_GRANTED -> {
                     android.util.Log.d("JellyTauPlayer", "✓ Audio focus GRANTED")
+                    hasAudioFocus = true
+                    true
                 }
                 AudioManager.AUDIOFOCUS_REQUEST_FAILED -> {
-                    android.util.Log.e("JellyTauPlayer", "✗ Audio focus REQUEST FAILED!")
+                    // Something holds exclusive focus (a call, say). Playing now
+                    // would be a silent video, so hold and wait for the grant.
+                    android.util.Log.e("JellyTauPlayer", "✗ Audio focus REQUEST FAILED - holding playback")
+                    false
                 }
                 AudioManager.AUDIOFOCUS_REQUEST_DELAYED -> {
-                    android.util.Log.d("JellyTauPlayer", "⏳ Audio focus DELAYED")
+                    android.util.Log.d("JellyTauPlayer", "⏳ Audio focus DELAYED - holding playback until GAIN")
+                    false
+                }
+                else -> {
+                    android.util.Log.w("JellyTauPlayer", "Unknown audio focus result: $result - holding playback")
+                    false
                 }
             }
         } else {
@@ -1014,10 +1478,15 @@ class JellyTauPlayer(private val appContext: Context) {
                 AudioManager.STREAM_MUSIC,
                 AudioManager.AUDIOFOCUS_GAIN
             )
-            if (result == AudioManager.AUDIOFOCUS_REQUEST_GRANTED) {
+            return if (result == AudioManager.AUDIOFOCUS_REQUEST_GRANTED) {
                 android.util.Log.d("JellyTauPlayer", "✓ Audio focus GRANTED (legacy)")
+                hasAudioFocus = true
+                true
             } else {
+                // Pre-O has no delayed grant and no listener to resume from, so a
+                // refusal is terminal for this attempt; the user can hit play again.
                 android.util.Log.e("JellyTauPlayer", "✗ Audio focus REQUEST FAILED (legacy)!")
+                false
             }
         }
     }
@@ -1028,6 +1497,11 @@ class JellyTauPlayer(private val appContext: Context) {
     private fun abandonAudioFocus() {
         android.util.Log.d("JellyTauPlayer", "Abandoning audio focus")
 
+        // No focus, nothing to resume: a stale flag would start playback the next
+        // time some unrelated GAIN arrives.
+        pendingPlayOnFocusGain = false
+        hasAudioFocus = false
+
         if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
             audioFocusRequest?.let {
                 val result = audioManager.abandonAudioFocusRequest(it)

src-tauri/android/src/main/res/values/themes.xml

@@ -1,13 +1,30 @@
 <resources xmlns:tools="http://schemas.android.com/tools">
-    <!-- Base application theme -->
+    <!--
+      Base application theme.
+
+      This app is EDGE-TO-EDGE: MainActivity calls enableEdgeToEdge(), and
+      targeting SDK 36 makes it mandatory anyway (enforced from SDK 35, with the
+      opt-out ignored from SDK 36). The WebView therefore spans the whole window,
+      under the status bar, the navigation/gesture bar and the display cutout.
+
+      This theme used to declare `android:fitsSystemWindows=true` with a "don't
+      draw behind system bars" comment. That was never true: enableEdgeToEdge()
+      calls setDecorFitsSystemWindows(false) at runtime and wins, and the
+      platform ignores the attribute at this target SDK regardless. Leaving it
+      in only hid the fact that nothing was insetting the content.
+
+      Insets are handled where they can actually be honoured: WindowInsetsBridge
+      reads them and hands them to CSS as jt-inset custom properties.
+      See UR-066 / DR-112.
+    -->
     <style name="Theme.jellytau" parent="Theme.MaterialComponents.DayNight.NoActionBar">
-        <!-- Status bar color -->
+        <!-- System bars are transparent; the app draws its own background behind
+             them (e.g. BottomUi's surface extends under the gesture bar). -->
         <item name="android:statusBarColor">@android:color/transparent</item>
-        <!-- Make status bar icons dark or light based on background -->
+        <item name="android:navigationBarColor">@android:color/transparent</item>
+        <!-- Light icons on our dark background, both bars. -->
         <item name="android:windowLightStatusBar" tools:targetApi="m">false</item>
-        <!-- Don't draw behind status bar -->
+        <item name="android:windowLightNavigationBar" tools:targetApi="o_mr1">false</item>
         <item name="android:windowDrawsSystemBarBackgrounds">true</item>
-        <!-- Ensure content doesn't extend into system bars -->
-        <item name="android:fitsSystemWindows">true</item>
     </style>
 </resources>

src-tauri/android/src/main/res/xml/network_security_config.xml

@@ -0,0 +1,23 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!--
+  Cleartext HTTP stays blocked everywhere except loopback.
+
+  Downloaded media is served to the webview by a local HTTP server on
+  127.0.0.1 (see media_server.rs / DR-137). Release builds set
+  usesCleartextTraffic="false", so Android's network security policy rejected
+  those requests before any I/O happened and offline video failed instantly with
+  NETWORK_NO_SOURCE.
+
+  Only 127.0.0.1 is exempted. The base config keeps the release default, so a
+  remote server still has to be HTTPS — this must not become a blanket
+  cleartext opt-in.
+
+  TRACES: UR-071 | DR-138
+-->
+<network-security-config>
+    <base-config cleartextTrafficPermitted="false" />
+
+    <domain-config cleartextTrafficPermitted="true">
+        <domain includeSubdomains="false">127.0.0.1</domain>
+    </domain-config>
+</network-security-config>

src-tauri/src/auth/mod.rs

@@ -129,6 +129,18 @@ impl AuthManager {
         Ok(normalized)
     }
 
+    /// Normalize a username before it goes to the server.
+    ///
+    /// Only surrounding whitespace is stripped — interior spaces are legal in
+    /// Jellyfin usernames. Without this, a trailing space from a soft keyboard's
+    /// autocorrect makes the server report an unknown user, which surfaces as a
+    /// 401 that looks exactly like a wrong password.
+    ///
+    /// TRACES: UR-042 | DR-054
+    pub fn normalize_username(username: &str) -> String {
+        username.trim().to_string()
+    }
+
     /// Connect to server and get server info
     pub async fn connect_to_server(&self, server_url: &str) -> Result<ServerInfo, String> {
         let normalized_url = Self::normalize_url(server_url)?;
@@ -185,6 +197,7 @@ impl AuthManager {
     ) -> Result<AuthResult, String> {
         let url = Self::normalize_url(server_url)?;
         let endpoint = format!("{}/Users/AuthenticateByName", url);
+        let username = Self::normalize_username(username);
 
         log::info!("[AuthManager] Authenticating user: {}", username);
 
@@ -443,6 +456,26 @@ mod tests {
         );
     }
 
+    /// Usernames must be trimmed before they reach the server: the Android soft
+    /// keyboard appends a trailing space after autocorrect, and Jellyfin then
+    /// reports an unknown user — a 401 indistinguishable from a wrong password.
+    #[test]
+    fn test_normalize_username_trims_whitespace() {
+        assert_eq!(AuthManager::normalize_username("duncan "), "duncan");
+        assert_eq!(AuthManager::normalize_username(" duncan"), "duncan");
+        assert_eq!(AuthManager::normalize_username("  duncan  "), "duncan");
+        assert_eq!(AuthManager::normalize_username("duncan\n"), "duncan");
+    }
+
+    /// Interior spaces are legal in Jellyfin usernames and must survive.
+    #[test]
+    fn test_normalize_username_preserves_interior_spaces() {
+        assert_eq!(
+            AuthManager::normalize_username("  duncan tourolle  "),
+            "duncan tourolle"
+        );
+    }
+
     /// Test URL normalization - real world case
     #[test]
     fn test_normalize_url_real_world_case() {

src-tauri/src/commands/catalog.rs

@@ -14,10 +14,12 @@
 //!   that were queued offline (they have `stream_url IS NULL`), mirroring the
 //!   heal-and-pump pattern in `player_preload_upcoming`.
 
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::Arc;
+use std::time::Duration;
 
 use log::{info, warn};
-use tauri::State;
+use tauri::{Emitter, Manager, State};
 
 use crate::commands::download::{pump_download_queue, DownloadManagerWrapper};
 use crate::commands::repository::RepositoryManagerWrapper;
@@ -29,18 +31,89 @@ use crate::storage::db_service::{DatabaseService, Query, QueryParam};
 /// full-catalog sync.
 const LAST_CATALOG_SYNC_KEY: &str = "last_catalog_sync";
 
+/// How long an index stays fresh before a re-index is due.
+///
+/// This lives in Rust rather than being a frontend constant because it decides
+/// *whether the local cache is authoritative* — the same class of decision as
+/// `include_catalog_browse`, and squarely the "sync policy" the spec review
+/// checklist keeps out of the presentation layer. If it later becomes
+/// user-configurable it stays a Rust-owned setting edited through a command.
+const CATALOG_INDEX_TTL: Duration = Duration::from_secs(6 * 60 * 60);
+
+/// How often the scheduler wakes to *check* staleness. Far shorter than the TTL
+/// because a tick is nearly free — one indexed `app_settings` lookup — and it is
+/// what makes the indexer responsive to events it cannot subscribe to: signing
+/// in, and coming back online. The TTL, not the tick, decides whether a crawl
+/// actually happens.
+const CATALOG_INDEX_TICK: Duration = Duration::from_secs(5 * 60);
+
+/// Delay before the first staleness check, to let sign-in complete and the
+/// repository be registered. Without it the first check runs against an empty
+/// repository manager and a fresh install would sit unindexed until the next
+/// tick.
+const CATALOG_INDEX_FIRST_CHECK: Duration = Duration::from_secs(15);
+
+/// Kebab-case, per the project's event convention.
+pub const CATALOG_INDEX_EVENT: &str = "catalog-index-event";
+
+/// Guards against two passes running at once. Replaces the frontend's
+/// `syncInProgress` boolean in `offlineCatalog.ts`, which could not see a pass
+/// started by the scheduler.
+static INDEX_IN_PROGRESS: AtomicBool = AtomicBool::new(false);
+
+/// Clears [`INDEX_IN_PROGRESS`] however the pass leaves — including on the `?`
+/// early return when `get_libraries` fails, which a plain store at the end of
+/// the function would leak.
+struct IndexPassGuard;
+
+impl Drop for IndexPassGuard {
+    fn drop(&mut self) {
+        INDEX_IN_PROGRESS.store(false, Ordering::SeqCst);
+    }
+}
+
+/// Progress of a background index pass, for the staleness hint in the UI.
+#[derive(specta::Type, Debug, Clone, serde::Serialize, serde::Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct CatalogIndexEvent {
+    /// `started` | `finished` | `failed`
+    pub state: String,
+    pub items_cached: usize,
+    pub items_pruned: usize,
+    pub libraries_failed: usize,
+    /// Present on `failed`.
+    pub error: Option<String>,
+}
+
 /// Item types worth caching for offline browsing: containers the library
 /// landing pages render plus the playable leaves users queue for download.
+/// `MusicArtist` and `Playlist` are here because search groups results by them
+/// (UR-060's Artists group). Without them in the crawl, the local index can
+/// never answer an artist query and those groups can only ever be filled by the
+/// server leg. Keep this in step with what `prune_stale_catalog` is allowed to
+/// sweep — the crawl is only authoritative for the types it asks for.
+///
+/// TRACES: UR-065, UR-060 | DR-111
 const CATALOG_ITEM_TYPES: &[&str] = &[
     "MusicAlbum",
+    "MusicArtist",
     "Movie",
     "Series",
     "Season",
     "Episode",
     "Audio",
     "BoxSet",
+    "Playlist",
 ];
 
+/// Jellyfin item types whose download is a *video* stream rather than an audio
+/// one. The download queue stores an opaque `media_type` ('audio'/'video'); this
+/// is where the taxonomy that produces it lives, so the frontend never has to
+/// know which item types are video.
+///
+/// TRACES: UR-071 | DR-135
+const VIDEO_ITEM_TYPES: &[&str] = &["Movie", "Episode", "Video", "MusicVideo"];
+
 #[derive(specta::Type, Debug, Clone, serde::Serialize, serde::Deserialize)]
 #[serde(rename_all = "camelCase")]
 pub struct CatalogSyncResult {
@@ -48,6 +121,9 @@ pub struct CatalogSyncResult {
     pub items_cached: usize,
     /// Libraries that failed to sync (e.g. server hiccup); best-effort.
     pub libraries_failed: usize,
+    /// Entries removed because the server no longer has them. Always 0 when any
+    /// library failed, since a partial crawl cannot prove an item is gone.
+    pub items_pruned: usize,
 }
 
 #[derive(specta::Type, Debug, Clone, serde::Serialize, serde::Deserialize)]
@@ -71,8 +147,6 @@ pub async fn sync_full_catalog(
     db: State<'_, DatabaseWrapper>,
     handle: String,
 ) -> Result<CatalogSyncResult, String> {
-    use crate::repository::MediaRepository;
-
     let repo = repository.0.get(&handle).ok_or("Repository not found")?;
 
     let db_service = {
@@ -80,6 +154,27 @@ pub async fn sync_full_catalog(
         Arc::new(database.service())
     };
 
+    run_index_pass(repo, db_service).await
+}
+
+/// One full-catalog indexing pass, shared by the [`sync_full_catalog`] command
+/// and the background scheduler (DR-109) so there is exactly one implementation
+/// and one concurrency guard.
+///
+/// TRACES: UR-065 | DR-109, DR-110
+pub(crate) async fn run_index_pass(
+    repo: Arc<crate::repository::HybridRepository>,
+    db_service: Arc<crate::storage::db_service::RusqliteService>,
+) -> Result<CatalogSyncResult, String> {
+    use crate::repository::MediaRepository;
+
+    // One pass at a time. The command and the scheduler can both land here, and
+    // two concurrent crawls would double the server load and race on the sweep.
+    if INDEX_IN_PROGRESS.swap(true, Ordering::SeqCst) {
+        return Err("A catalog index pass is already running".to_string());
+    }
+    let _guard = IndexPassGuard;
+
     let libraries = repo.get_libraries().await.map_err(|e| e.to_string())?;
     info!(
         "[Catalog] Full sync starting across {} libraries",
@@ -88,6 +183,10 @@ pub async fn sync_full_catalog(
 
     let include_types: Vec<String> = CATALOG_ITEM_TYPES.iter().map(|s| s.to_string()).collect();
 
+    // Taken before the crawl: every row the crawl writes gets a `synced_at`
+    // newer than this, so anything still older afterwards is gone server-side.
+    let pass_started_at = chrono::Utc::now().to_rfc3339();
+
     let mut items_cached = 0usize;
     let mut libraries_failed = 0usize;
 
@@ -118,6 +217,35 @@ pub async fn sync_full_catalog(
         }
     }
 
+    // Propagate server-side deletions — but only after a *complete* crawl.
+    // `sync_full_catalog` is best-effort per library, and `items.parent_id` is
+    // ON DELETE CASCADE, so sweeping when a library failed to fetch could
+    // cascade a whole series away because one request timed out.
+    let mut items_pruned = 0usize;
+    if libraries_failed == 0 && !libraries.is_empty() {
+        match repo
+            .prune_stale_catalog(&pass_started_at, &include_types)
+            .await
+        {
+            Ok(removed) => {
+                items_pruned = removed;
+                if removed > 0 {
+                    info!(
+                        "[Catalog] Pruned {} entries no longer on the server",
+                        removed
+                    );
+                }
+            }
+            Err(e) => warn!("[Catalog] Prune of stale catalog entries failed: {:?}", e),
+        }
+    } else if libraries_failed > 0 {
+        info!(
+            "[Catalog] Skipping stale-entry prune: {} librar{} failed to sync, so the crawl is not authoritative",
+            libraries_failed,
+            if libraries_failed == 1 { "y" } else { "ies" }
+        );
+    }
+
     // Record the sync time so callers can skip re-syncing too eagerly.
     let now = chrono::Utc::now().to_rfc3339();
     let upsert = Query::with_params(
@@ -133,16 +261,169 @@ pub async fn sync_full_catalog(
     }
 
     info!(
-        "[Catalog] Full sync complete: {} items cached, {} libraries failed",
-        items_cached, libraries_failed
+        "[Catalog] Full sync complete: {} items cached, {} pruned, {} libraries failed",
+        items_cached, items_pruned, libraries_failed
     );
 
     Ok(CatalogSyncResult {
         items_cached,
         libraries_failed,
+        items_pruned,
     })
 }
 
+/// Whether an index pass is due, given when one last completed.
+///
+/// Pure so the policy is unit-testable without a clock, a server, or a database.
+/// `None` (never indexed) and an unparseable stored value both mean "due" — a
+/// corrupt timestamp should trigger a re-index, not silently freeze the catalog.
+///
+/// TRACES: UR-065 | DR-109 | UT-115
+pub(crate) fn index_is_due(
+    last_synced_at: Option<&str>,
+    now: chrono::DateTime<chrono::Utc>,
+    ttl: Duration,
+) -> bool {
+    let Some(raw) = last_synced_at else {
+        return true;
+    };
+    let Ok(last) = chrono::DateTime::parse_from_rfc3339(raw) else {
+        return true;
+    };
+    now.signed_duration_since(last.with_timezone(&chrono::Utc))
+        .to_std()
+        .map(|elapsed| elapsed >= ttl)
+        // Negative elapsed => the stored stamp is in the future (clock skew).
+        // Not due; a future stamp will age into due-ness on its own.
+        .unwrap_or(false)
+}
+
+/// Read the last-sync timestamp straight from `app_settings`.
+async fn read_last_sync(
+    db_service: &Arc<crate::storage::db_service::RusqliteService>,
+) -> Option<String> {
+    db_service
+        .query_optional(
+            Query::with_params(
+                "SELECT value FROM app_settings WHERE key = ?",
+                vec![QueryParam::String(LAST_CATALOG_SYNC_KEY.to_string())],
+            ),
+            |row| row.get(0),
+        )
+        .await
+        .ok()
+        .flatten()
+}
+
+/// Start the background catalog indexer.
+///
+/// Replaces the frontend's startup-only `syncCatalog()` call: index freshness is
+/// sync policy and belongs in Rust (see the layer assignment in
+/// docs/specs/catalog-index-search.md). Ticks every [`CATALOG_INDEX_TICK`] and
+/// runs a pass when a repository exists, the server is reachable, and the index
+/// is older than [`CATALOG_INDEX_TTL`].
+///
+/// TRACES: UR-065 | DR-109, IR-030
+pub fn spawn_catalog_indexer(app: tauri::AppHandle) {
+    tauri::async_runtime::spawn(async move {
+        // Check shortly after launch, then on every tick — not tick-then-check,
+        // which would leave a fresh install unindexed for a full tick.
+        tokio::time::sleep(CATALOG_INDEX_FIRST_CHECK).await;
+
+        loop {
+            if let Err(e) = maybe_run_scheduled_pass(&app).await {
+                // Never fatal — a failed pass leaves the existing index in place
+                // and we retry on the next tick.
+                warn!("[Catalog] Scheduled index pass skipped: {}", e);
+            }
+
+            tokio::time::sleep(CATALOG_INDEX_TICK).await;
+        }
+    });
+}
+
+/// One scheduler tick: check the preconditions, then index if due.
+async fn maybe_run_scheduled_pass(app: &tauri::AppHandle) -> Result<(), String> {
+    if INDEX_IN_PROGRESS.load(Ordering::SeqCst) {
+        return Ok(());
+    }
+
+    let db_service = {
+        let db = app.state::<DatabaseWrapper>();
+        let database = db.0.lock().map_err(|e| e.to_string())?;
+        Arc::new(database.service())
+    };
+
+    if !index_is_due(
+        read_last_sync(&db_service).await.as_deref(),
+        chrono::Utc::now(),
+        CATALOG_INDEX_TTL,
+    ) {
+        return Ok(());
+    }
+
+    // Offline: leave the index alone. The crawl would fail every library and,
+    // more importantly, a partial crawl must never reach the deletion sweep.
+    {
+        let monitor = app.state::<crate::commands::connectivity::ConnectivityMonitorWrapper>();
+        let monitor = monitor.0.lock().await;
+        if !monitor.get_status().await.is_server_reachable {
+            return Ok(());
+        }
+    }
+
+    let repo = {
+        let manager = app.state::<RepositoryManagerWrapper>();
+        let handles = manager.0.handles();
+        let Some(handle) = handles.first() else {
+            // Not signed in yet.
+            return Ok(());
+        };
+        manager.0.get(handle).ok_or("Repository not found")?
+    };
+
+    info!("[Catalog] Index is stale; starting a scheduled pass");
+    let _ = app.emit(
+        CATALOG_INDEX_EVENT,
+        CatalogIndexEvent {
+            state: "started".to_string(),
+            items_cached: 0,
+            items_pruned: 0,
+            libraries_failed: 0,
+            error: None,
+        },
+    );
+
+    match run_index_pass(repo, db_service).await {
+        Ok(result) => {
+            let _ = app.emit(
+                CATALOG_INDEX_EVENT,
+                CatalogIndexEvent {
+                    state: "finished".to_string(),
+                    items_cached: result.items_cached,
+                    items_pruned: result.items_pruned,
+                    libraries_failed: result.libraries_failed,
+                    error: None,
+                },
+            );
+            Ok(())
+        }
+        Err(e) => {
+            let _ = app.emit(
+                CATALOG_INDEX_EVENT,
+                CatalogIndexEvent {
+                    state: "failed".to_string(),
+                    items_cached: 0,
+                    items_pruned: 0,
+                    libraries_failed: 0,
+                    error: Some(e.clone()),
+                },
+            );
+            Err(e)
+        }
+    }
+}
+
 /// Report the last-synced timestamp so the UI can show a hint / decide whether
 /// to trigger a fresh sync.
 #[tauri::command]
@@ -190,24 +471,111 @@ pub struct ResumeQueuedResult {
     pub failed: usize,
 }
 
+/// Requeue video downloads that were fetched as audio.
+///
+/// Before [`resolve_pending_download_urls`] consulted the item's type, a row
+/// with no `media_type` — which is every row queued from a media card, since
+/// `download_item` does not record one — resolved against
+/// `get_audio_stream_url`. A movie queued that way completed with an audio-only
+/// transcode on disk, so playing it offline could only ever fail. Those rows are
+/// identifiable after the fact (no `media_type`, but a video item), so reset them
+/// to pending with no URL and let the resolver fetch the real video.
+///
+/// Rows carrying an explicit `media_type` were resolved correctly and are left
+/// alone, as are genuine audio downloads.
+///
+/// Returns the number of rows requeued.
+///
+/// TRACES: UR-071 | DR-136 | UT-126
+pub(crate) async fn requeue_mistyped_video_downloads(
+    db_service: &Arc<crate::storage::db_service::RusqliteService>,
+) -> Result<usize, String> {
+    let video_types = VIDEO_ITEM_TYPES
+        .iter()
+        .map(|t| format!("'{t}'"))
+        .collect::<Vec<_>>()
+        .join(", ");
+
+    let query = Query::new(&format!(
+        "UPDATE downloads
+         SET status = 'pending', stream_url = NULL, progress = 0,
+             bytes_downloaded = 0, started_at = NULL, completed_at = NULL
+         WHERE media_type IS NULL
+           AND status = 'completed'
+           AND item_id IN (SELECT id FROM items WHERE item_type IN ({video_types}))"
+    ));
+
+    let n = db_service.execute(query).await.map_err(|e| e.to_string())? as usize;
+
+    if n > 0 {
+        info!(
+            "[Catalog] Requeued {} video download(s) that were fetched as audio",
+            n
+        );
+    }
+    Ok(n)
+}
+
 /// Core of [`resume_queued_downloads`], factored out for testing: select every
 /// `pending`/`stream_url IS NULL` row, resolve each via `resolve` (returning
 /// `None` leaves the row pending), and heal the row so the pump can start it.
 /// The `resolve` closure receives `(item_id, media_type, quality_preset)`.
+///
+/// `only_ids` restricts the sweep to specific download rows. Reconnect passes
+/// `None` and heals everything; a bulk enqueue (an album, say) passes the rows
+/// it just created, so clicking download on one album cannot also start every
+/// unrelated row that has been sitting pending.
 pub(crate) async fn resolve_pending_download_urls<F, Fut>(
     db_service: &Arc<crate::storage::db_service::RusqliteService>,
     target_dir: &str,
+    only_ids: Option<&[i64]>,
     resolve: F,
 ) -> Result<ResumeQueuedResult, String>
 where
     F: Fn(String, String, String) -> Fut,
     Fut: std::future::Future<Output = Option<String>>,
 {
-    let rows_query = Query::new(
-        "SELECT id, item_id, COALESCE(media_type, 'audio'), COALESCE(quality_preset, 'original')
-         FROM downloads
-         WHERE status = 'pending' AND stream_url IS NULL",
-    );
+    if only_ids.is_some_and(|ids| ids.is_empty()) {
+        return Ok(ResumeQueuedResult {
+            resolved: 0,
+            failed: 0,
+        });
+    }
+    // A row's own media_type wins; otherwise the *item's* type decides. Rows
+    // queued from a media card never carry one (`download_item` does not record
+    // it), and defaulting that NULL to 'audio' resolved movies against
+    // `get_audio_stream_url` — the file on disk was an audio-only transcode, so
+    // offline video could never play. Falling back to 'audio' only when the item
+    // is unknown keeps the historical behaviour for uncached items.
+    // TRACES: UR-071, UR-052 | DR-135
+    let video_types = VIDEO_ITEM_TYPES
+        .iter()
+        .map(|t| format!("'{t}'"))
+        .collect::<Vec<_>>()
+        .join(", ");
+    let id_filter = match only_ids {
+        Some(ids) => format!(
+            " AND d.id IN ({})",
+            ids.iter()
+                .map(|id| id.to_string())
+                .collect::<Vec<_>>()
+                .join(", ")
+        ),
+        None => String::new(),
+    };
+    let rows_query = Query::new(&format!(
+        "SELECT d.id, d.item_id,
+                COALESCE(
+                    d.media_type,
+                    CASE WHEN i.item_type IN ({video_types}) THEN 'video'
+                         WHEN i.item_type IS NOT NULL THEN 'audio'
+                    END,
+                    'audio'),
+                COALESCE(d.quality_preset, 'original')
+         FROM downloads d
+         LEFT JOIN items i ON i.id = d.item_id
+         WHERE d.status = 'pending' AND d.stream_url IS NULL{id_filter}"
+    ));
     let rows: Vec<(i64, String, String, String)> = db_service
         .query_many(rows_query, |row| {
             Ok((row.get(0)?, row.get(1)?, row.get(2)?, row.get(3)?))
@@ -285,8 +653,6 @@ pub async fn resume_queued_downloads(
 ) -> Result<ResumeQueuedResult, String> {
     use crate::repository::MediaRepository;
 
-    use crate::repository::HybridRepository;
-
     let repo = repository.0.get(&handle).ok_or("Repository not found")?;
 
     // The pump needs a target_dir; use the same storage root the other download
@@ -317,22 +683,34 @@ pub async fn resume_queued_downloads(
         Err(e) => warn!("[Catalog] Failed to reset stale downloads: {}", e),
     }
 
+    // Repair rows that completed as audio because their media_type was missing;
+    // they hold an audio-only transcode where a video should be, so requeue them
+    // for the resolver below. TRACES: UR-071 | DR-136
+    if let Err(e) = requeue_mistyped_video_downloads(&db_service).await {
+        warn!(
+            "[Catalog] Failed to requeue mis-typed video downloads: {}",
+            e
+        );
+    }
+
     // Resolve each row's URL against the (now reachable) repository.
     let repo_for_resolve = Arc::clone(&repo);
     let outcome = resolve_pending_download_urls(
         &db_service,
         &target_dir,
+        None,
         move |item_id: String, media_type: String, quality: String| {
             let repo = Arc::clone(&repo_for_resolve);
             async move {
                 if media_type == "video" {
                     Some(
-                        <HybridRepository as MediaRepository>::get_video_download_url(
+                        crate::repository::resolve_video_download_url(
                             repo.as_ref(),
                             &item_id,
                             &quality,
                             None,
-                        ),
+                        )
+                        .await,
                     )
                 } else {
                     match repo.get_audio_stream_url(&item_id).await {
@@ -378,6 +756,43 @@ mod tests {
     use rusqlite::Connection;
     use std::sync::Mutex;
 
+    /// The re-index policy. Pure, so it is testable without a clock, a server or
+    /// a database — which is the reason it was factored out of the scheduler.
+    ///
+    /// TRACES: UR-065 | DR-109 | UT-115
+    #[test]
+    fn test_index_is_due() {
+        let ttl = Duration::from_secs(6 * 60 * 60);
+        let now = chrono::DateTime::parse_from_rfc3339("2026-08-04T12:00:00+00:00")
+            .unwrap()
+            .with_timezone(&chrono::Utc);
+
+        // Never indexed => due. This is the first-run case.
+        assert!(index_is_due(None, now, ttl));
+
+        // Indexed 7 hours ago => past the 6h TTL => due.
+        assert!(index_is_due(Some("2026-08-04T05:00:00+00:00"), now, ttl));
+
+        // Indexed 1 hour ago => fresh => not due. This is what stops the
+        // scheduler re-crawling every tick.
+        assert!(!index_is_due(Some("2026-08-04T11:00:00+00:00"), now, ttl));
+
+        // Exactly at the TTL boundary counts as due.
+        assert!(index_is_due(Some("2026-08-04T06:00:00+00:00"), now, ttl));
+
+        // A corrupt stored value must trigger a re-index, not freeze the
+        // catalog forever behind an unparseable timestamp.
+        assert!(index_is_due(Some("not-a-timestamp"), now, ttl));
+        assert!(index_is_due(Some(""), now, ttl));
+
+        // A timestamp in the future (clock skew, or a restored backup) is not
+        // due — it ages into due-ness rather than causing a crawl every tick.
+        assert!(!index_is_due(Some("2026-08-05T00:00:00+00:00"), now, ttl));
+
+        // Offsets other than UTC are compared as instants, not as strings.
+        assert!(!index_is_due(Some("2026-08-04T13:30:00+02:00"), now, ttl));
+    }
+
     fn test_db() -> Arc<RusqliteService> {
         let conn = Connection::open_in_memory().unwrap();
         conn.execute_batch(
@@ -389,7 +804,15 @@ mod tests {
                 stream_url TEXT,
                 target_dir TEXT,
                 media_type TEXT,
-                quality_preset TEXT
+                quality_preset TEXT,
+                progress REAL DEFAULT 0,
+                bytes_downloaded INTEGER DEFAULT 0,
+                started_at TEXT,
+                completed_at TEXT
+            );
+            CREATE TABLE items (
+                id TEXT PRIMARY KEY,
+                item_type TEXT
             );
             "#,
         )
@@ -397,6 +820,18 @@ mod tests {
         Arc::new(RusqliteService::new(Arc::new(Mutex::new(conn))))
     }
 
+    async fn insert_item(db: &Arc<RusqliteService>, item_id: &str, item_type: &str) {
+        db.execute(Query::with_params(
+            "INSERT INTO items (id, item_type) VALUES (?, ?)",
+            vec![
+                QueryParam::String(item_id.to_string()),
+                QueryParam::String(item_type.to_string()),
+            ],
+        ))
+        .await
+        .unwrap();
+    }
+
     async fn insert_download(
         db: &Arc<RusqliteService>,
         item_id: &str,
@@ -433,6 +868,12 @@ mod tests {
             .unwrap()
     }
 
+    /// IT-017: a download queued from a greyed-out offline catalog entry
+    /// (pending, `stream_url IS NULL`) persists, and on reconnect its URL is
+    /// resolved and the row is healed (URL + target dir) so the pump can start
+    /// it — while already-resolved rows are left untouched.
+    ///
+    /// TRACES: UR-052, UR-011 | IT-017
     #[tokio::test]
     async fn resolves_offline_queued_row_and_leaves_resolved_rows_untouched() {
         let db = test_db();
@@ -443,12 +884,14 @@ mod tests {
         // A completed row: irrelevant.
         insert_download(&db, "done", "completed", Some("http://done/url"), None).await;
 
-        let out =
-            resolve_pending_download_urls(&db, "/data/downloads", |item_id, _mt, _q| async move {
-                Some(format!("http://resolved/{item_id}"))
-            })
-            .await
-            .unwrap();
+        let out = resolve_pending_download_urls(
+            &db,
+            "/data/downloads",
+            None,
+            |item_id, _mt, _q| async move { Some(format!("http://resolved/{item_id}")) },
+        )
+        .await
+        .unwrap();
 
         assert_eq!(out.resolved, 1);
         assert_eq!(out.failed, 0);
@@ -464,15 +907,79 @@ mod tests {
         assert_eq!(url2.as_deref(), Some("http://existing/url"));
     }
 
+    /// A bulk enqueue resolves only the rows it just created. Downloading one
+    /// album must not also start every unrelated row that has been sitting
+    /// pending with no URL (the smart cache leaves plenty of those).
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-171
+    #[tokio::test]
+    async fn only_ids_restricts_the_sweep_to_the_given_rows() {
+        let db = test_db();
+        insert_download(&db, "mine", "pending", None, Some("audio")).await;
+        insert_download(&db, "someone-elses", "pending", None, Some("audio")).await;
+
+        let mine: i64 = db
+            .query_one(
+                Query::new("SELECT id FROM downloads WHERE item_id = 'mine'"),
+                |row| row.get(0),
+            )
+            .await
+            .unwrap();
+
+        let out = resolve_pending_download_urls(
+            &db,
+            "/data",
+            Some(&[mine]),
+            |item_id, _mt, _q| async move { Some(format!("http://resolved/{item_id}")) },
+        )
+        .await
+        .unwrap();
+
+        assert_eq!(out.resolved, 1);
+        assert_eq!(out.failed, 0);
+
+        let (_s, url, _t) = get_row(&db, "mine").await;
+        assert_eq!(url.as_deref(), Some("http://resolved/mine"));
+
+        let (status, other_url, _t) = get_row(&db, "someone-elses").await;
+        assert_eq!(status, "pending");
+        assert_eq!(
+            other_url, None,
+            "a scoped resolve must leave unrelated pending rows alone"
+        );
+    }
+
+    /// An empty id list resolves nothing — it must not fall through to "sweep
+    /// everything", which is what an unguarded `IN ()` would amount to.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-171
+    #[tokio::test]
+    async fn an_empty_id_list_resolves_nothing() {
+        let db = test_db();
+        insert_download(&db, "untouched", "pending", None, Some("audio")).await;
+
+        let out =
+            resolve_pending_download_urls(&db, "/data", Some(&[]), |item_id, _mt, _q| async move {
+                Some(format!("http://resolved/{item_id}"))
+            })
+            .await
+            .unwrap();
+
+        assert_eq!(out.resolved, 0);
+        let (_s, url, _t) = get_row(&db, "untouched").await;
+        assert_eq!(url, None);
+    }
+
     #[tokio::test]
     async fn counts_unresolvable_rows_as_failed_and_leaves_them_pending() {
         let db = test_db();
         insert_download(&db, "bad", "pending", None, None).await;
 
         // Resolver returns None (e.g. server lookup failed).
-        let out = resolve_pending_download_urls(&db, "/data", |_id, _mt, _q| async move { None })
-            .await
-            .unwrap();
+        let out =
+            resolve_pending_download_urls(&db, "/data", None, |_id, _mt, _q| async move { None })
+                .await
+                .unwrap();
 
         assert_eq!(out.resolved, 0);
         assert_eq!(out.failed, 1);
@@ -483,18 +990,153 @@ mod tests {
         assert_eq!(url, None);
     }
 
+    /// A movie queued from a media card has no `media_type` — `download_item`
+    /// never records one. Defaulting that NULL to 'audio' resolved the row
+    /// against `get_audio_stream_url`, so the "downloaded movie" on disk was an
+    /// audio-only transcode and offline video playback could never work. The
+    /// item's own type is the authority.
+    ///
+    /// TRACES: UR-071, UR-052 | DR-135 | UT-125
+    #[tokio::test]
+    async fn null_media_type_resolves_from_the_item_type_not_audio() {
+        let db = test_db();
+        insert_item(&db, "movie-1", "Movie").await;
+        insert_item(&db, "ep-1", "Episode").await;
+        insert_item(&db, "track-1", "Audio").await;
+        for id in ["movie-1", "ep-1", "track-1"] {
+            insert_download(&db, id, "pending", None, None).await;
+        }
+
+        let seen = Arc::new(Mutex::new(Vec::new()));
+        let seen_c = Arc::clone(&seen);
+        resolve_pending_download_urls(&db, "/data", None, move |item_id, media_type, _q| {
+            let seen = Arc::clone(&seen_c);
+            async move {
+                seen.lock().unwrap().push((item_id.clone(), media_type));
+                Some(format!("http://resolved/{item_id}"))
+            }
+        })
+        .await
+        .unwrap();
+
+        let seen = seen.lock().unwrap().clone();
+        let of = |id: &str| {
+            seen.iter()
+                .find(|(i, _)| i == id)
+                .map(|(_, m)| m.clone())
+                .unwrap()
+        };
+        assert_eq!(of("movie-1"), "video", "a Movie must download as video");
+        assert_eq!(of("ep-1"), "video", "an Episode must download as video");
+        assert_eq!(of("track-1"), "audio", "a track is still audio");
+    }
+
+    /// An unknown item (never cached locally) has no type to derive from, so it
+    /// keeps the historical audio default rather than failing the row.
+    ///
+    /// TRACES: UR-071 | DR-135 | UT-125
+    #[tokio::test]
+    async fn unknown_item_falls_back_to_audio() {
+        let db = test_db();
+        insert_download(&db, "ghost", "pending", None, None).await;
+
+        let seen = Arc::new(Mutex::new(String::new()));
+        let seen_c = Arc::clone(&seen);
+        resolve_pending_download_urls(&db, "/data", None, move |_id, media_type, _q| {
+            let seen = Arc::clone(&seen_c);
+            async move {
+                *seen.lock().unwrap() = media_type;
+                Some("http://x".to_string())
+            }
+        })
+        .await
+        .unwrap();
+
+        assert_eq!(*seen.lock().unwrap(), "audio");
+    }
+
+    /// An explicit `media_type` on the row always wins over the item's type.
+    ///
+    /// TRACES: UR-071 | DR-135 | UT-125
+    #[tokio::test]
+    async fn explicit_media_type_beats_the_item_type() {
+        let db = test_db();
+        insert_item(&db, "odd", "Audio").await;
+        insert_download(&db, "odd", "pending", None, Some("video")).await;
+
+        let seen = Arc::new(Mutex::new(String::new()));
+        let seen_c = Arc::clone(&seen);
+        resolve_pending_download_urls(&db, "/data", None, move |_id, media_type, _q| {
+            let seen = Arc::clone(&seen_c);
+            async move {
+                *seen.lock().unwrap() = media_type;
+                Some("http://x".to_string())
+            }
+        })
+        .await
+        .unwrap();
+
+        assert_eq!(*seen.lock().unwrap(), "video");
+    }
+
+    /// Rows already downloaded under the audio default hold an audio-only
+    /// transcode on disk, so they play as a broken video forever. They are
+    /// identifiable — no `media_type` but a video item — and are requeued so the
+    /// resolver fetches the real video. Correctly-typed rows and genuine audio
+    /// downloads must be left alone.
+    ///
+    /// TRACES: UR-071 | DR-136 | UT-126
+    #[tokio::test]
+    async fn requeues_video_downloaded_under_the_audio_default() {
+        let db = test_db();
+        insert_item(&db, "movie-1", "Movie").await;
+        insert_item(&db, "track-1", "Audio").await;
+        insert_item(&db, "movie-ok", "Movie").await;
+        // Mis-downloaded: completed, no media_type, video item.
+        insert_download(&db, "movie-1", "completed", Some("http://audio/url"), None).await;
+        // A real audio download: untouched.
+        insert_download(&db, "track-1", "completed", Some("http://audio/ok"), None).await;
+        // A correctly-typed video download: untouched.
+        insert_download(
+            &db,
+            "movie-ok",
+            "completed",
+            Some("http://video/ok"),
+            Some("video"),
+        )
+        .await;
+
+        let requeued = requeue_mistyped_video_downloads(&db).await.unwrap();
+        assert_eq!(requeued, 1);
+
+        let (status, url, _t) = get_row(&db, "movie-1").await;
+        assert_eq!(status, "pending", "the mis-typed row must download again");
+        assert_eq!(url, None, "its audio URL must be cleared so it re-resolves");
+
+        let (status, url, _t) = get_row(&db, "track-1").await;
+        assert_eq!(status, "completed", "a real audio download is untouched");
+        assert_eq!(url.as_deref(), Some("http://audio/ok"));
+
+        let (status, _u, _t) = get_row(&db, "movie-ok").await;
+        assert_eq!(status, "completed", "a correct video download is untouched");
+    }
+
     #[tokio::test]
     async fn video_rows_use_media_type_in_resolver() {
         let db = test_db();
         insert_download(&db, "vid-1", "pending", None, Some("video")).await;
 
-        let out =
-            resolve_pending_download_urls(&db, "/data", |item_id, media_type, _q| async move {
+        let out = resolve_pending_download_urls(
+            &db,
+            "/data",
+            None,
+            |item_id, media_type, _q| async move {
                 assert_eq!(media_type, "video");
                 Some(format!("http://transcode/{item_id}"))
-            })
-            .await
-            .unwrap();
+            },
+        )
+        .await
+        .unwrap();
 
         assert_eq!(out.resolved, 1);
         let (_s, url, _t) = get_row(&db, "vid-1").await;

src-tauri/src/commands/download/mod.rs

@@ -270,6 +270,19 @@ pub async fn download_item(
         if !can_download {
             warn!("Storage limit reached. Attempting to free space...");
 
+            // Reclaim expired temporary entries first: they are dead weight, so
+            // freeing them may avoid evicting cache that is still within its
+            // life. Best-effort — a failure here just means eviction does more.
+            // TRACES: UR-071 | DR-127
+            match cache_arc
+                .reclaim_expired_async(&db_service, &user_id, &chrono::Utc::now().to_rfc3339())
+                .await
+            {
+                Ok(n) if n > 0 => info!("Reclaimed {} expired cache entries", n),
+                Ok(_) => {}
+                Err(e) => warn!("Expired-entry reclaim failed: {}", e),
+            }
+
             // Try to evict LRU items to make space
             match cache_arc
                 .evict_lru_async(&db_service, &user_id, size as u64)
@@ -337,57 +350,209 @@ pub async fn download_item(
     Ok(download_id)
 }
 
-/// Queue an entire album for download
-#[tauri::command]
-#[specta::specta]
-pub async fn download_album(
-    db: State<'_, DatabaseWrapper>,
-    album_id: String,
-    user_id: String,
-    base_path: String,
-) -> Result<Vec<i64>, String> {
-    let db_service = {
-        let database = db.0.lock().map_err(|e| e.to_string())?;
-        Arc::new(database.service())
-    };
+/// One track of an album, as the album-download path queues it.
+///
+/// `artist_name` carries whatever the catalog holds for the track's artists (a
+/// JSON array, as stored on `items.artists`); it is display metadata for the
+/// downloads list, not a lookup key.
+///
+/// TRACES: UR-018, UR-055 | DR-173
+#[derive(Debug, Clone, PartialEq)]
+pub(crate) struct AlbumTrack {
+    pub id: String,
+    pub name: String,
+    pub artist_name: Option<String>,
+    pub album_name: Option<String>,
+    pub index_number: Option<i32>,
+}
 
-    // Get all tracks in the album with metadata
+impl From<&crate::repository::types::MediaItem> for AlbumTrack {
+    fn from(item: &crate::repository::types::MediaItem) -> Self {
+        Self {
+            id: item.id.clone(),
+            name: item.name.clone(),
+            artist_name: item
+                .artists
+                .as_ref()
+                .and_then(|a| serde_json::to_string(a).ok()),
+            album_name: item.album_name.clone(),
+            index_number: item.index_number,
+        }
+    }
+}
+
+/// The album's tracks as the local catalog cache knows them.
+///
+/// Only a fallback for [`download_album`]: the cache links a track to its album
+/// through `items.album_id`, which Jellyfin does not populate on every listing
+/// endpoint, so this can legitimately return fewer tracks than the album has.
+///
+/// TRACES: UR-018, UR-055 | DR-173
+pub(crate) async fn cached_album_tracks(
+    db_service: &Arc<crate::storage::db_service::RusqliteService>,
+    album_id: &str,
+) -> Result<Vec<AlbumTrack>, String> {
     let tracks_query = Query::with_params(
-        "SELECT id, name, artists, album_name FROM items
-         WHERE album_id = ? AND item_type = 'Audio'
+        "SELECT id, name, artists, album_name, index_number FROM items
+         WHERE (album_id = ? OR parent_id = ?) AND item_type = 'Audio'
          ORDER BY index_number",
-        vec![QueryParam::String(album_id)],
+        vec![
+            QueryParam::String(album_id.to_string()),
+            QueryParam::String(album_id.to_string()),
+        ],
     );
 
-    let tracks: Vec<(String, String, Option<String>, Option<String>)> = db_service
+    db_service
         .query_many(tracks_query, |row| {
-            Ok((row.get(0)?, row.get(1)?, row.get(2)?, row.get(3)?))
+            Ok(AlbumTrack {
+                id: row.get(0)?,
+                name: row.get(1)?,
+                artist_name: row.get(2)?,
+                album_name: row.get(3)?,
+                index_number: row.get(4)?,
+            })
         })
         .await
-        .map_err(|e| e.to_string())?;
+        .map_err(|e| e.to_string())
+}
 
-    let mut download_ids = Vec::new();
+/// Queue one download row per track and link every track to its album.
+///
+/// The linkage is the half that is easy to miss: offline browsing joins a track
+/// to its album on `items.album_id` (see `OfflineRepository::get_items`), so a
+/// track whose cached row lacks it stays invisible under the album even after
+/// its file is on disk. Queuing a track *is* the statement that it belongs to
+/// this album, so the link is written here rather than hoped for from whichever
+/// listing endpoint happened to cache the row.
+///
+/// Idempotent: re-queuing an album fills in what is missing and returns the same
+/// row ids, in the order the tracks were given.
+///
+/// A file name per track, unique within the album.
+///
+/// A title is not a unique name inside its own album: a deluxe edition carries
+/// the album version and a demo of the same song, and a two-disc set repeats
+/// titles across discs. Naming files after the title alone gave those tracks one
+/// path, and each download overwrote the previous one — an album that quietly
+/// ends up short by however many titles it repeats. The track number
+/// disambiguates the ordinary case; anything still colliding falls back to the
+/// item id, which is unique by construction.
+///
+/// TRACES: UR-018, UR-055 | DR-173 | UT-172
+pub(crate) fn album_file_names(tracks: &[AlbumTrack]) -> Vec<String> {
+    let mut counts: std::collections::HashMap<String, usize> = std::collections::HashMap::new();
+    for track in tracks {
+        *counts.entry(track.name.to_lowercase()).or_default() += 1;
+    }
 
-    // Queue each track with album priority (100) and metadata
-    for (track_id, track_name, artist_name, album_name) in tracks {
-        let file_path = format!("{}/{}.mp3", base_path, sanitize_filename(&track_name));
+    tracks
+        .iter()
+        .map(|track| {
+            let title = sanitize_filename(&track.name);
+            if counts.get(&track.name.to_lowercase()).copied().unwrap_or(0) <= 1 {
+                return format!("{}.mp3", title);
+            }
+            match track.index_number {
+                Some(n) => format!("{:02} - {} [{}].mp3", n, title, track.id),
+                None => format!("{} [{}].mp3", title, track.id),
+            }
+        })
+        .collect()
+}
 
+/// TRACES: UR-018, UR-055 | DR-173 | UT-170
+pub(crate) async fn queue_album_tracks(
+    db_service: &Arc<crate::storage::db_service::RusqliteService>,
+    album_id: &str,
+    tracks: &[AlbumTrack],
+    user_id: &str,
+    base_path: &str,
+) -> Result<Vec<i64>, String> {
+    let mut download_ids = Vec::with_capacity(tracks.len());
+    let file_names = album_file_names(tracks);
+
+    for (track, file_name) in tracks.iter().zip(file_names) {
+        // Cache a row for a track the catalog has never seen, borrowing the
+        // album's server. Nothing is inserted when the album itself is unknown,
+        // which also keeps the parent_id foreign key satisfiable.
+        let cache_query = Query::with_params(
+            "INSERT OR IGNORE INTO items
+                (id, server_id, parent_id, name, item_type, album_id, album_name, artists, index_number)
+             SELECT ?, a.server_id, a.id, ?, 'Audio', a.id, ?, ?, ?
+             FROM items a WHERE a.id = ?",
+            vec![
+                QueryParam::String(track.id.clone()),
+                QueryParam::String(track.name.clone()),
+                track
+                    .album_name
+                    .clone()
+                    .map(QueryParam::String)
+                    .unwrap_or(QueryParam::Null),
+                track
+                    .artist_name
+                    .clone()
+                    .map(QueryParam::String)
+                    .unwrap_or(QueryParam::Null),
+                track
+                    .index_number
+                    .map(QueryParam::Int)
+                    .unwrap_or(QueryParam::Null),
+                QueryParam::String(album_id.to_string()),
+            ],
+        );
+        db_service
+            .execute(cache_query)
+            .await
+            .map_err(|e| e.to_string())?;
+
+        // Link an already-cached track to the album. The parent_id subquery
+        // resolves to NULL when the album is not cached, so the foreign key
+        // holds either way.
+        let link_query = Query::with_params(
+            "UPDATE items
+                SET album_id = ?,
+                    parent_id = COALESCE(parent_id, (SELECT id FROM items WHERE id = ?))
+              WHERE id = ?",
+            vec![
+                QueryParam::String(album_id.to_string()),
+                QueryParam::String(album_id.to_string()),
+                QueryParam::String(track.id.clone()),
+            ],
+        );
+        db_service
+            .execute(link_query)
+            .await
+            .map_err(|e| e.to_string())?;
+
+        let file_path = format!("{}/{}", base_path, file_name);
+
+        // Queue at album priority (100). A track already downloaded stays
+        // completed — re-queuing an album must fill the gaps, not re-fetch it.
         let insert_query = Query::with_params(
-            "INSERT INTO downloads (item_id, user_id, file_path, status, priority, queued_at, item_name, artist_name, album_name)
-             VALUES (?, ?, ?, 'pending', 100, CURRENT_TIMESTAMP, ?, ?, ?)
+            "INSERT INTO downloads (item_id, user_id, file_path, status, priority, queued_at, item_name, artist_name, album_name, media_type)
+             VALUES (?, ?, ?, 'pending', 100, CURRENT_TIMESTAMP, ?, ?, ?, 'audio')
              ON CONFLICT(item_id, user_id) DO UPDATE SET
                priority = 100,
-               status = 'pending',
+               status = CASE WHEN downloads.status = 'completed' THEN 'completed' ELSE 'pending' END,
+               media_type = 'audio',
                item_name = COALESCE(excluded.item_name, downloads.item_name),
                artist_name = COALESCE(excluded.artist_name, downloads.artist_name),
                album_name = COALESCE(excluded.album_name, downloads.album_name)",
             vec![
-                QueryParam::String(track_id.clone()),
-                QueryParam::String(user_id.clone()),
+                QueryParam::String(track.id.clone()),
+                QueryParam::String(user_id.to_string()),
                 QueryParam::String(file_path),
-                QueryParam::String(track_name),
-                artist_name.map(QueryParam::String).unwrap_or(QueryParam::Null),
-                album_name.map(QueryParam::String).unwrap_or(QueryParam::Null),
+                QueryParam::String(track.name.clone()),
+                track
+                    .artist_name
+                    .clone()
+                    .map(QueryParam::String)
+                    .unwrap_or(QueryParam::Null),
+                track
+                    .album_name
+                    .clone()
+                    .map(QueryParam::String)
+                    .unwrap_or(QueryParam::Null),
             ],
         );
 
@@ -400,8 +565,8 @@ pub async fn download_album(
         let id_query = Query::with_params(
             "SELECT id FROM downloads WHERE item_id = ? AND user_id = ?",
             vec![
-                QueryParam::String(track_id),
-                QueryParam::String(user_id.clone()),
+                QueryParam::String(track.id.clone()),
+                QueryParam::String(user_id.to_string()),
             ],
         );
 
@@ -415,6 +580,129 @@ pub async fn download_album(
     Ok(download_ids)
 }
 
+/// Queue an entire album for download.
+///
+/// Owns the whole operation: the album's track list comes from the server (the
+/// only place that knows all of it), every track is queued and linked to its
+/// album, each row's stream URL is resolved here, and the queue is pumped.
+///
+/// The frontend used to do the second half — resolve one URL per track and pair
+/// it with the returned ids **by position**. That pairing had no basis: the ids
+/// came back in the backend's own order over a different set of rows, so
+/// whenever the two lists disagreed a row was handed another track's URL, and
+/// any track past the end of the shorter list was never started at all. Nothing
+/// crosses the boundary now except the album id.
+///
+/// TRACES: UR-018, UR-055 | DR-173 | UT-170
+#[tauri::command]
+#[specta::specta]
+pub async fn download_album(
+    db: State<'_, DatabaseWrapper>,
+    repository: State<'_, crate::commands::repository::RepositoryManagerWrapper>,
+    download_manager: State<'_, DownloadManagerWrapper>,
+    app: tauri::AppHandle,
+    handle: String,
+    album_id: String,
+    user_id: String,
+    base_path: String,
+) -> Result<Vec<i64>, String> {
+    let db_service = {
+        let database = db.0.lock().map_err(|e| e.to_string())?;
+        Arc::new(database.service())
+    };
+
+    let repo = repository.0.get(&handle);
+
+    // Ask the server what the album contains; the cache is only a fallback for
+    // when it cannot answer.
+    let tracks: Vec<AlbumTrack> = match &repo {
+        Some(repo) => match repo.get_album_tracks(&album_id).await {
+            Ok(items) if !items.is_empty() => items.iter().map(AlbumTrack::from).collect(),
+            Ok(_) => cached_album_tracks(&db_service, &album_id).await?,
+            Err(e) => {
+                warn!(
+                    "[download_album] Could not list album {} from the repository ({:?}); \
+                     falling back to the cached track list",
+                    album_id, e
+                );
+                cached_album_tracks(&db_service, &album_id).await?
+            }
+        },
+        None => cached_album_tracks(&db_service, &album_id).await?,
+    };
+
+    if tracks.is_empty() {
+        warn!("[download_album] No tracks found for album {}", album_id);
+        return Ok(Vec::new());
+    }
+
+    let download_ids =
+        queue_album_tracks(&db_service, &album_id, &tracks, &user_id, &base_path).await?;
+
+    info!(
+        "[download_album] Queued {} track(s) for album {}",
+        download_ids.len(),
+        album_id
+    );
+
+    // Resolve each queued row's stream URL here, then pump. Without a
+    // repository (or while offline) the rows stay pending with no URL and
+    // `resume_queued_downloads` picks them up on reconnect.
+    let Some(repo) = repo else {
+        return Ok(download_ids);
+    };
+
+    let target_dir = {
+        let database = db.0.lock().map_err(|e| e.to_string())?;
+        database
+            .path()
+            .parent()
+            .ok_or_else(|| "Database path has no parent directory".to_string())?
+            .to_string_lossy()
+            .to_string()
+    };
+
+    let repo_for_resolve = Arc::clone(&repo);
+    let outcome = crate::commands::catalog::resolve_pending_download_urls(
+        &db_service,
+        &target_dir,
+        Some(&download_ids),
+        move |item_id: String, _media_type: String, _quality: String| {
+            let repo = Arc::clone(&repo_for_resolve);
+            async move {
+                use crate::repository::MediaRepository;
+                match repo.get_audio_stream_url(&item_id).await {
+                    Ok(url) => Some(url),
+                    Err(e) => {
+                        warn!(
+                            "[download_album] Failed to resolve stream URL for {}: {:?}",
+                            item_id, e
+                        );
+                        None
+                    }
+                }
+            }
+        },
+    )
+    .await?;
+
+    if outcome.failed > 0 {
+        warn!(
+            "[download_album] {} track(s) could not be resolved and stay queued for the next \
+             reconnect",
+            outcome.failed
+        );
+    }
+
+    let active_downloads = {
+        let manager = download_manager.0.lock().map_err(|e| e.to_string())?;
+        manager.get_active_downloads()
+    };
+    pump_download_queue(app, db_service, active_downloads).await;
+
+    Ok(download_ids)
+}
+
 /// Queue a video item (movie or episode) for download with quality preset
 #[tauri::command]
 #[specta::specta]
@@ -832,7 +1120,19 @@ pub async fn get_downloads(
     Ok(DownloadsResponse { downloads, stats })
 }
 
-/// Pause a download
+/// Pause a download.
+///
+/// Writing `status = 'paused'` is only half of it, and used to be all of it: the
+/// streaming task knew nothing about the row and kept running, then overwrote it
+/// with `completed`/`failed` when it finished. The row flicked to "paused" and
+/// undid itself — the reported "pause does not work". Signalling the worker is
+/// what actually stops the bytes; it leaves the `.part` file in place so
+/// [`resume_download`] can continue from it.
+///
+/// A queued (not yet started) download has no worker to signal, and the status
+/// write alone is enough — the pump skips anything that is not `pending`.
+///
+/// TRACES: UR-055 | DR-168
 #[tauri::command]
 #[specta::specta]
 pub async fn pause_download(
@@ -845,19 +1145,34 @@ pub async fn pause_download(
     };
 
     let query = Query::with_params(
-        "UPDATE downloads SET status = 'paused' WHERE id = ? AND status = 'downloading'",
+        "UPDATE downloads SET status = 'paused' WHERE id = ? AND status IN ('downloading', 'pending')",
         vec![QueryParam::Int64(download_id)],
     );
 
     db_service.execute(query).await.map_err(|e| e.to_string())?;
+
+    let was_running = crate::download::stop::signal(download_id);
+    info!(
+        "[pause] Download {} paused (in flight: {})",
+        download_id, was_running
+    );
     Ok(())
 }
 
-/// Resume a paused download
+/// Resume a paused download.
+///
+/// Flipping the row back to `pending` is likewise not enough on its own: the
+/// pump is not a poller, it runs when something calls it, so a resumed download
+/// sat untouched until some unrelated event happened to pump the queue. That is
+/// the other half of "resume does not work".
+///
+/// TRACES: UR-055 | DR-168
 #[tauri::command]
 #[specta::specta]
 pub async fn resume_download(
+    app: tauri::AppHandle,
     db: State<'_, DatabaseWrapper>,
+    download_manager: State<'_, DownloadManagerWrapper>,
     download_id: i64,
 ) -> Result<(), String> {
     let db_service = {
@@ -866,11 +1181,22 @@ pub async fn resume_download(
     };
 
     let query = Query::with_params(
-        "UPDATE downloads SET status = 'pending' WHERE id = ? AND status = 'paused'",
+        "UPDATE downloads SET status = 'pending', error_message = NULL WHERE id = ? AND status IN ('paused', 'failed')",
         vec![QueryParam::Int64(download_id)],
     );
 
     db_service.execute(query).await.map_err(|e| e.to_string())?;
+
+    // Drop any stale stop flag before the pump can start this id again, or the
+    // resumed run would read the pause that stopped it and halt immediately.
+    crate::download::stop::clear(download_id);
+
+    let active_downloads = {
+        let manager = download_manager.0.lock().map_err(|e| e.to_string())?;
+        manager.get_active_downloads()
+    };
+    pump_download_queue(app, db_service, active_downloads).await;
+
     Ok(())
 }
 
@@ -910,6 +1236,13 @@ pub async fn cancel_download(
         .await
         .map_err(|e| e.to_string())?;
 
+    // Stop the worker if this download is actually running. Without this the
+    // task keeps streaming into a `.part` file whose `downloads` row has just
+    // been deleted — bytes with nothing pointing at them, and the file below is
+    // removed while still being written to. (DR-168)
+    crate::download::stop::signal(download_id);
+    crate::download::stop::clear(download_id);
+
     // Unregister from download manager (in case it was active)
     {
         let manager = download_manager.0.lock().map_err(|e| e.to_string())?;
@@ -921,10 +1254,12 @@ pub async fn cancel_download(
         );
     }
 
-    // Delete partial file if exists
+    // Delete the partial file, and any completed file, if present. Both go
+    // through `partial_path` so this cannot drift from what the worker writes —
+    // it did, and every cancelled download leaked its partial. (DR-169)
     if let Some(path) = file_path {
-        let partial_path = format!("{}.part", path);
-        let _ = std::fs::remove_file(&partial_path); // Ignore errors
+        let target = std::path::PathBuf::from(&path);
+        let _ = std::fs::remove_file(crate::download::worker::partial_path(&target));
     }
 
     Ok(())
@@ -1231,8 +1566,6 @@ pub async fn enqueue_video_downloads(
     download_ids: Vec<i64>,
     target_dir: String,
 ) -> Result<(), String> {
-    use crate::repository::MediaRepository;
-
     let repo = repository.0.get(&handle).ok_or("Repository not found")?;
 
     let db_service = {
@@ -1257,10 +1590,12 @@ pub async fn enqueue_video_downloads(
             }
         };
 
-        // Build the transcode URL (pure URL builder, no server round-trip).
-        let stream_url = repo
-            .as_ref()
-            .get_video_download_url(&item_id, &quality, None);
+        // Build the download URL, resolving the source's audio codec first so a
+        // track this device cannot decode is re-encoded on the way down rather
+        // than saved as a silent file (DR-167).
+        let stream_url =
+            crate::repository::resolve_video_download_url(repo.as_ref(), &item_id, &quality, None)
+                .await;
 
         let update_query = Query::with_params(
             "UPDATE downloads SET status = 'pending', stream_url = ?, target_dir = ? WHERE id = ?",
@@ -1517,7 +1852,11 @@ fn spawn_download_worker(
             let _ = progress_app.emit("download-event", event);
         };
 
-        let result = worker.download(&task, on_progress).await;
+        // Registering returns a fresh flag, so a download resumed after a pause
+        // does not inherit the stop that ended its previous run. (DR-168)
+        let stop_flag = crate::download::stop::register(download_id);
+        let result = worker.download(&task, &stop_flag, on_progress).await;
+        crate::download::stop::clear(download_id);
 
         // Free the slot before pumping so the next download can take it.
         if let Ok(mut active) = active_downloads.lock() {
@@ -1588,6 +1927,17 @@ fn spawn_download_worker(
                     Err(e) => error!("   Completed event emit failed: {:?}", e),
                 }
             }
+            // A pause or cancel is not a failure. The row already says `paused`
+            // (or the row is gone, for a cancel), and overwriting that with
+            // `failed` is what made a pause look like an error and stranded the
+            // download outside the resumable set. The `.part` file is deliberately
+            // left alone — it is what the resume continues from. (DR-168)
+            Err(e) if e.is_stopped() => {
+                info!(
+                    "[pump] Download {} stopped by request; partial file kept for resume",
+                    download_id
+                );
+            }
             Err(e) => {
                 error!("Download failed: {:?}", e);
 
@@ -1835,17 +2185,26 @@ pub async fn clear_stale_downloads(
         Arc::new(database.service())
     };
 
-    // Get file paths for stale downloads (pending/paused/failed)
+    // Ids as well as paths: a stale row may still have a worker attached (a
+    // 'downloading' row that was paused mid-flight is 'paused' here), and
+    // deleting the row without stopping the task leaves it writing to a file we
+    // are about to remove. (DR-168)
     let file_query = Query::with_params(
-        "SELECT file_path FROM downloads WHERE user_id = ? AND status IN ('pending', 'paused', 'failed')",
+        "SELECT id, file_path FROM downloads WHERE user_id = ? AND status IN ('pending', 'paused', 'failed')",
         vec![QueryParam::String(user_id.clone())],
     );
 
-    let file_paths: Vec<String> = db_service
-        .query_many(file_query, |row| row.get(0))
+    let stale: Vec<(i64, String)> = db_service
+        .query_many(file_query, |row| Ok((row.get(0)?, row.get(1)?)))
         .await
         .map_err(|e| e.to_string())?;
 
+    for (id, _) in &stale {
+        crate::download::stop::signal(*id);
+        crate::download::stop::clear(*id);
+    }
+    let file_paths: Vec<String> = stale.into_iter().map(|(_, path)| path).collect();
+
     // Delete all pending, paused, and failed downloads (but keep completed ones)
     let delete_query = Query::with_params(
         "DELETE FROM downloads WHERE user_id = ? AND status IN ('pending', 'paused', 'failed')",
@@ -1857,10 +2216,12 @@ pub async fn clear_stale_downloads(
         .await
         .map_err(|e| e.to_string())?;
 
-    // Delete any partial files
+    // Delete any partial files, via the shared helper so this cannot drift from
+    // what the worker actually writes. (DR-169)
     for path in file_paths {
-        let _ = std::fs::remove_file(&path);
-        let _ = std::fs::remove_file(format!("{}.part", path));
+        let target = std::path::PathBuf::from(&path);
+        let _ = std::fs::remove_file(&target);
+        let _ = std::fs::remove_file(crate::download::worker::partial_path(&target));
     }
 
     Ok(deleted_count as i64)
@@ -2572,4 +2933,294 @@ mod tests {
             download_source: "user".to_string(),
         }
     }
+
+    // ===== Album download: track sourcing and album linkage =====
+
+    /// A database with just the tables the album-download path touches.
+    fn album_test_db() -> Arc<crate::storage::db_service::RusqliteService> {
+        let conn = rusqlite::Connection::open_in_memory().unwrap();
+        conn.execute_batch(
+            r#"
+            CREATE TABLE items (
+                id TEXT PRIMARY KEY,
+                server_id TEXT NOT NULL,
+                parent_id TEXT,
+                name TEXT NOT NULL,
+                item_type TEXT NOT NULL,
+                album_id TEXT,
+                album_name TEXT,
+                album_artist TEXT,
+                artists TEXT,
+                index_number INTEGER
+            );
+            CREATE TABLE downloads (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                item_id TEXT NOT NULL,
+                user_id TEXT NOT NULL,
+                file_path TEXT NOT NULL,
+                status TEXT DEFAULT 'pending',
+                priority INTEGER DEFAULT 0,
+                progress REAL DEFAULT 0,
+                queued_at TEXT,
+                item_name TEXT,
+                artist_name TEXT,
+                album_name TEXT,
+                media_type TEXT,
+                stream_url TEXT,
+                target_dir TEXT,
+                UNIQUE(item_id, user_id)
+            );
+            INSERT INTO items (id, server_id, name, item_type)
+                VALUES ('album1', 'server1', 'The Golden Age', 'MusicAlbum');
+            "#,
+        )
+        .unwrap();
+        Arc::new(crate::storage::db_service::RusqliteService::new(Arc::new(
+            Mutex::new(conn),
+        )))
+    }
+
+    fn album_track(id: &str, name: &str, index: i32) -> AlbumTrack {
+        AlbumTrack {
+            id: id.to_string(),
+            name: name.to_string(),
+            artist_name: Some("Woodkid".to_string()),
+            album_name: Some("The Golden Age".to_string()),
+            index_number: Some(index),
+        }
+    }
+
+    /// The album-download regression: every track the album actually has must be
+    /// queued, and each queued track must be linked to its album.
+    ///
+    /// `download_album` used to take its track list from
+    /// `items WHERE album_id = ?`. Jellyfin does not return `AlbumId` on every
+    /// listing endpoint, so tracks cached from those endpoints sit in `items`
+    /// with a NULL `album_id` — invisible to that query. "Download album" then
+    /// silently queued only the subset that happened to carry the link, which is
+    /// the reported "only 4-5 songs downloaded". The same column is what offline
+    /// browsing joins tracks to their album on (`i.album_id = ?` in
+    /// `OfflineRepository::get_items`), so even a track that did download stayed
+    /// invisible under its album offline.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-170
+    #[tokio::test]
+    async fn test_queue_album_tracks_queues_every_track_and_links_it_to_the_album() {
+        let db = album_test_db();
+
+        // The cache holds all three tracks, but only one carries `album_id` —
+        // exactly the state the bug report's database is in.
+        for sql in [
+            "INSERT INTO items (id, server_id, name, item_type, album_id) \
+             VALUES ('t1', 'server1', 'Run Boy Run', 'Audio', 'album1')",
+            "INSERT INTO items (id, server_id, name, item_type, album_id) \
+             VALUES ('t2', 'server1', 'The Great Escape', 'Audio', NULL)",
+            "INSERT INTO items (id, server_id, name, item_type, album_id) \
+             VALUES ('t3', 'server1', 'Boat Song', 'Audio', NULL)",
+        ] {
+            db.execute(Query::new(sql)).await.unwrap();
+        }
+
+        let tracks = vec![
+            album_track("t1", "Run Boy Run", 1),
+            album_track("t2", "The Great Escape", 2),
+            album_track("t3", "Boat Song", 3),
+        ];
+
+        let ids = queue_album_tracks(&db, "album1", &tracks, "user1", "albums/album1")
+            .await
+            .unwrap();
+
+        assert_eq!(
+            ids.len(),
+            3,
+            "every track of the album must get a download row"
+        );
+
+        let queued: i64 = db
+            .query_one(
+                Query::new("SELECT COUNT(*) FROM downloads WHERE status = 'pending'"),
+                |row| row.get(0),
+            )
+            .await
+            .unwrap();
+        assert_eq!(queued, 3);
+
+        // Each track is now linked to its album, so the offline album page can
+        // find it once the download completes.
+        let linked: i64 = db
+            .query_one(
+                Query::new("SELECT COUNT(*) FROM items WHERE album_id = 'album1'"),
+                |row| row.get(0),
+            )
+            .await
+            .unwrap();
+        assert_eq!(
+            linked, 3,
+            "queued tracks must be linked to their album; offline browsing joins on album_id"
+        );
+    }
+
+    /// The returned ids must line up with the tracks that were passed in. The
+    /// frontend used to pair `downloadIds[i]` with its own `tracks[i]`, which is
+    /// only sound if both lists agree — they did not, because the backend
+    /// ordered by `index_number` over a different set of rows. Resolving URLs in
+    /// Rust removes the pairing entirely, but the order is still the contract
+    /// for anything that reads the ids back.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-170
+    #[tokio::test]
+    async fn test_queue_album_tracks_returns_ids_in_track_order() {
+        let db = album_test_db();
+        let tracks = vec![
+            album_track("t1", "Run Boy Run", 1),
+            album_track("t2", "The Great Escape", 2),
+        ];
+
+        let ids = queue_album_tracks(&db, "album1", &tracks, "user1", "albums/album1")
+            .await
+            .unwrap();
+
+        for (id, track) in ids.iter().zip(tracks.iter()) {
+            let item_id: String = db
+                .query_one(
+                    Query::with_params(
+                        "SELECT item_id FROM downloads WHERE id = ?",
+                        vec![QueryParam::Int64(*id)],
+                    ),
+                    |row| row.get(0),
+                )
+                .await
+                .unwrap();
+            assert_eq!(&item_id, &track.id, "id {} must be {}'s row", id, track.id);
+        }
+    }
+
+    /// Re-queueing an album already partly downloaded must not duplicate rows or
+    /// reset a completed track — it fills in what is missing.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-170
+    #[tokio::test]
+    async fn test_queue_album_tracks_is_idempotent() {
+        let db = album_test_db();
+        let tracks = vec![
+            album_track("t1", "Run Boy Run", 1),
+            album_track("t2", "The Great Escape", 2),
+        ];
+
+        let first = queue_album_tracks(&db, "album1", &tracks, "user1", "albums/album1")
+            .await
+            .unwrap();
+        let second = queue_album_tracks(&db, "album1", &tracks, "user1", "albums/album1")
+            .await
+            .unwrap();
+
+        assert_eq!(first, second, "the same tracks must map to the same rows");
+
+        let rows: i64 = db
+            .query_one(Query::new("SELECT COUNT(*) FROM downloads"), |row| {
+                row.get(0)
+            })
+            .await
+            .unwrap();
+        assert_eq!(rows, 2, "re-queueing must not duplicate download rows");
+    }
+
+    /// Two tracks of one album can share a title — a deluxe edition carrying the
+    /// album version and a demo of the same song, or the same song on two discs.
+    /// Naming the file after the title alone gave them one path, so the second
+    /// download overwrote the first and the album ended up short however many
+    /// duplicates it had.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-172
+    #[test]
+    fn test_album_file_names_are_unique_within_the_album() {
+        let tracks = vec![
+            album_track("t1", "Crucified Again", 5),
+            album_track("t2", "Crucified Again", 5),
+            album_track("t3", "Get Right", 7),
+        ];
+
+        let names = album_file_names(&tracks);
+
+        assert_eq!(names.len(), 3);
+        let unique: std::collections::HashSet<_> = names.iter().collect();
+        assert_eq!(
+            unique.len(),
+            3,
+            "every track of an album needs its own file: {:?}",
+            names
+        );
+        assert!(names.iter().all(|n| n.ends_with(".mp3")), "{:?}", names);
+        assert!(
+            names[2].contains("Get Right"),
+            "an unambiguous title keeps its name: {}",
+            names[2]
+        );
+    }
+
+    /// Path separators in a track title must not escape the album directory.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-172
+    #[test]
+    fn test_album_file_names_sanitize_the_title() {
+        let names = album_file_names(&[album_track("t1", "AC/DC: Live?", 1)]);
+        assert!(!names[0].contains('/'), "{}", names[0]);
+        assert!(!names[0].contains(':'), "{}", names[0]);
+    }
+
+    /// The offline fallback reads the catalog directly, not through the
+    /// availability-gated offline listing: queueing an album while the server is
+    /// unreachable is a supported flow (the rows resolve on reconnect), and
+    /// gating it on what is already downloaded would queue only the tracks the
+    /// device already has.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-170
+    #[tokio::test]
+    async fn test_cached_album_tracks_finds_tracks_by_either_album_link() {
+        let db = album_test_db();
+        for sql in [
+            "INSERT INTO items (id, server_id, name, item_type, album_id, index_number) \
+             VALUES ('t1', 'server1', 'Run Boy Run', 'Audio', 'album1', 1)",
+            // Linked by parent_id only — how a track cached from a folder
+            // listing lands in the catalog.
+            "INSERT INTO items (id, server_id, name, item_type, parent_id, index_number) \
+             VALUES ('t2', 'server1', 'The Great Escape', 'Audio', 'album1', 2)",
+            // A different album's track must not be swept in.
+            "INSERT INTO items (id, server_id, name, item_type, album_id) \
+             VALUES ('other', 'server1', 'Iron', 'Audio', 'album2')",
+        ] {
+            db.execute(Query::new(sql)).await.unwrap();
+        }
+
+        let tracks = cached_album_tracks(&db, "album1").await.unwrap();
+        let ids: Vec<_> = tracks.iter().map(|t| t.id.as_str()).collect();
+        assert_eq!(ids, vec!["t1", "t2"]);
+    }
+
+    /// Tracks the cache has never seen still get queued: the row is created and
+    /// an `items` row is written for it, so the download is both startable and
+    /// visible offline afterwards.
+    ///
+    /// TRACES: UR-018, UR-055 | DR-173 | UT-170
+    #[tokio::test]
+    async fn test_queue_album_tracks_handles_tracks_absent_from_the_cache() {
+        let db = album_test_db();
+        let tracks = vec![album_track("never-cached", "Iron", 1)];
+
+        let ids = queue_album_tracks(&db, "album1", &tracks, "user1", "albums/album1")
+            .await
+            .unwrap();
+        assert_eq!(ids.len(), 1);
+
+        let (item_type, album_id): (String, Option<String>) = db
+            .query_one(
+                Query::new("SELECT item_type, album_id FROM items WHERE id = 'never-cached'"),
+                |row| Ok((row.get(0)?, row.get(1)?)),
+            )
+            .await
+            .unwrap();
+        assert_eq!(item_type, "Audio");
+        assert_eq!(album_id.as_deref(), Some("album1"));
+    }
 }

src-tauri/src/commands/favorites.rs

@@ -0,0 +1,367 @@
+//! Pushing favourite toggles made while the server was unreachable.
+//!
+//! Favouriting works offline: `storage_toggle_favorite` writes the local
+//! `user_data` row and sets `pending_sync = 1`. Until DR-120 nothing ever
+//! cleared that flag — the offline `mark_favorite`/`unmark_favorite` are no-ops
+//! and `syncService.queueFavorite` had no callers — so an offline toggle was
+//! silently lost.
+//!
+//! The drain lives in Rust, not the frontend, because it must run whether or
+//! not any view is mounted; a drain started by a component dies with it.
+//!
+//! TRACES: UR-069 | DR-120 | UT-103
+
+use std::sync::Arc;
+
+use async_trait::async_trait;
+use log::{debug, info, warn};
+use tauri::{Emitter, Listener, Manager};
+
+use crate::repository::types::RepoError;
+use crate::repository::MediaRepository;
+use crate::storage::db_service::{DatabaseService, Query, QueryParam, RusqliteService};
+
+/// The subset of the repository the drain needs.
+///
+/// Narrow on purpose: a test double for `MediaRepository` would be forty
+/// unimplemented methods, which is how a drain ends up untested.
+#[async_trait]
+pub trait FavoriteSink: Send + Sync {
+    async fn push_favorite(&self, item_id: &str, is_favorite: bool) -> Result<(), RepoError>;
+}
+
+#[async_trait]
+impl<T: MediaRepository + ?Sized> FavoriteSink for T {
+    async fn push_favorite(&self, item_id: &str, is_favorite: bool) -> Result<(), RepoError> {
+        if is_favorite {
+            self.mark_favorite(item_id).await
+        } else {
+            self.unmark_favorite(item_id).await
+        }
+    }
+}
+
+/// A local favourite change still waiting to reach the server.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct PendingFavorite {
+    pub item_id: String,
+    pub is_favorite: bool,
+}
+
+/// Read every favourite change this user has pending.
+async fn read_pending(
+    db: &Arc<RusqliteService>,
+    user_id: &str,
+) -> Result<Vec<PendingFavorite>, String> {
+    db.query_many(
+        Query::with_params(
+            "SELECT item_id, is_favorite FROM user_data \
+             WHERE user_id = ? AND pending_sync = 1 AND is_favorite IS NOT NULL",
+            vec![QueryParam::String(user_id.to_string())],
+        ),
+        |row| {
+            Ok(PendingFavorite {
+                item_id: row.get::<_, String>(0)?,
+                is_favorite: row.get::<_, Option<i32>>(1)?.unwrap_or(0) != 0,
+            })
+        },
+    )
+    .await
+}
+
+/// Push pending favourite changes to the server and clear their flags.
+///
+/// Returns the ids that reached the server, for the `favorites-changed` event.
+/// A row whose push fails keeps `pending_sync = 1` and is retried on the next
+/// reconnect rather than being dropped.
+///
+/// TRACES: UR-069 | DR-120 | UT-103
+pub async fn drain_pending_favorites(
+    db: &Arc<RusqliteService>,
+    sink: &dyn FavoriteSink,
+    user_id: &str,
+) -> Result<Vec<String>, String> {
+    let pending = read_pending(db, user_id).await?;
+    if pending.is_empty() {
+        return Ok(Vec::new());
+    }
+
+    info!(
+        "[Favorites] Pushing {} favourite change(s) queued while offline",
+        pending.len()
+    );
+
+    let mut pushed = Vec::new();
+    for change in pending {
+        match sink
+            .push_favorite(&change.item_id, change.is_favorite)
+            .await
+        {
+            Ok(()) => {
+                let cleared = db
+                    .execute(Query::with_params(
+                        "UPDATE user_data SET pending_sync = 0, synced_at = ? \
+                         WHERE user_id = ? AND item_id = ?",
+                        vec![
+                            QueryParam::String(chrono::Utc::now().to_rfc3339()),
+                            QueryParam::String(user_id.to_string()),
+                            QueryParam::String(change.item_id.clone()),
+                        ],
+                    ))
+                    .await;
+
+                match cleared {
+                    Ok(_) => pushed.push(change.item_id),
+                    // The server took it; failing to clear the flag only means
+                    // we push it again next time, which is harmless.
+                    Err(e) => warn!(
+                        "[Favorites] Pushed {} but could not clear pending_sync: {}",
+                        change.item_id, e
+                    ),
+                }
+            }
+            Err(e) => {
+                // Still pending — retried on the next reconnect.
+                debug!(
+                    "[Favorites] Deferring {}, server rejected the push: {:?}",
+                    change.item_id, e
+                );
+            }
+        }
+    }
+
+    Ok(pushed)
+}
+
+/// Drain on every offline→online transition.
+///
+/// Hooks the `connectivity:reconnected` event the `ConnectivityMonitor`
+/// already emits, rather than polling — reachability is derived from real
+/// traffic (DR-055) and this just reacts to it.
+///
+/// TRACES: UR-069 | DR-120
+pub fn spawn_favorites_drain(app: tauri::AppHandle) {
+    let handle = app.clone();
+    app.listen("connectivity:reconnected", move |_event| {
+        let app = handle.clone();
+        tauri::async_runtime::spawn(async move {
+            if let Err(e) = run_drain(&app).await {
+                warn!("[Favorites] Drain skipped: {}", e);
+            }
+        });
+    });
+}
+
+async fn run_drain(app: &tauri::AppHandle) -> Result<(), String> {
+    let db_service: Arc<RusqliteService> = {
+        let db = app.state::<crate::commands::storage::DatabaseWrapper>();
+        let database = db.0.lock().map_err(|e| e.to_string())?;
+        Arc::new(database.service())
+    };
+
+    let (repo, user_id) = {
+        let manager = app.state::<crate::commands::repository::RepositoryManagerWrapper>();
+        let handles = manager.0.handles();
+        let Some(handle) = handles.first() else {
+            // Not signed in — nothing to push on behalf of.
+            return Ok(());
+        };
+        let repo = manager.0.get(handle).ok_or("Repository not found")?;
+        let user_id = repo.user_id().to_string();
+        (repo, user_id)
+    };
+
+    let pushed = drain_pending_favorites(&db_service, repo.as_ref(), &user_id).await?;
+
+    if !pushed.is_empty() {
+        let event = crate::commands::repository::FavoritesChangedEvent { item_ids: pushed };
+        if let Err(e) = app.emit(crate::commands::repository::FAVORITES_CHANGED_EVENT, &event) {
+            warn!("[Favorites] Failed to emit change event: {}", e);
+        }
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use rusqlite::Connection;
+    use std::sync::Mutex;
+
+    /// Records what the server was asked to do, and can be told to fail.
+    struct RecordingSink {
+        calls: Mutex<Vec<(String, bool)>>,
+        fail_for: Option<String>,
+    }
+
+    impl RecordingSink {
+        fn new() -> Self {
+            Self {
+                calls: Mutex::new(Vec::new()),
+                fail_for: None,
+            }
+        }
+
+        fn failing_for(item_id: &str) -> Self {
+            Self {
+                calls: Mutex::new(Vec::new()),
+                fail_for: Some(item_id.to_string()),
+            }
+        }
+
+        fn calls(&self) -> Vec<(String, bool)> {
+            let mut calls = self.calls.lock().unwrap().clone();
+            calls.sort();
+            calls
+        }
+    }
+
+    #[async_trait]
+    impl FavoriteSink for RecordingSink {
+        async fn push_favorite(&self, item_id: &str, is_favorite: bool) -> Result<(), RepoError> {
+            if self.fail_for.as_deref() == Some(item_id) {
+                return Err(RepoError::Offline);
+            }
+            self.calls
+                .lock()
+                .unwrap()
+                .push((item_id.to_string(), is_favorite));
+            Ok(())
+        }
+    }
+
+    fn test_db() -> Arc<RusqliteService> {
+        let conn = Connection::open_in_memory().unwrap();
+        conn.execute_batch(
+            r#"
+            CREATE TABLE user_data (
+                user_id TEXT NOT NULL,
+                item_id TEXT NOT NULL,
+                is_favorite INTEGER,
+                synced_at TEXT,
+                pending_sync INTEGER DEFAULT 0,
+                PRIMARY KEY (user_id, item_id)
+            );
+            "#,
+        )
+        .unwrap();
+        Arc::new(RusqliteService::new(Arc::new(Mutex::new(conn))))
+    }
+
+    async fn seed(db: &Arc<RusqliteService>, rows: &[(&str, &str, i32, i32)]) {
+        for (user, item, fav, pending) in rows {
+            db.execute(Query::with_params(
+                "INSERT INTO user_data (user_id, item_id, is_favorite, pending_sync) \
+                 VALUES (?, ?, ?, ?)",
+                vec![
+                    QueryParam::String(user.to_string()),
+                    QueryParam::String(item.to_string()),
+                    QueryParam::Int(*fav),
+                    QueryParam::Int(*pending),
+                ],
+            ))
+            .await
+            .unwrap();
+        }
+    }
+
+    async fn pending_flag(db: &Arc<RusqliteService>, item_id: &str) -> Option<i32> {
+        db.query_optional(
+            Query::with_params(
+                "SELECT pending_sync FROM user_data WHERE item_id = ?",
+                vec![QueryParam::String(item_id.to_string())],
+            ),
+            |row| row.get::<_, Option<i32>>(0),
+        )
+        .await
+        .unwrap()
+        .flatten()
+    }
+
+    /// UT-103 — the core of the bug: a favourite toggled while offline reaches
+    /// the server on reconnect, and stops being pending.
+    ///
+    /// TRACES: UR-069 | DR-120 | UT-103
+    #[tokio::test]
+    async fn test_drain_pushes_pending_favorites_and_clears_the_flag() {
+        let db = test_db();
+        seed(
+            &db,
+            &[
+                ("u1", "marked-offline", 1, 1),
+                ("u1", "unmarked-offline", 0, 1),
+                // Already synced — must not be pushed again.
+                ("u1", "already-synced", 1, 0),
+            ],
+        )
+        .await;
+
+        let sink = RecordingSink::new();
+        let pushed = drain_pending_favorites(&db, &sink, "u1").await.unwrap();
+
+        assert_eq!(
+            sink.calls(),
+            vec![
+                ("marked-offline".to_string(), true),
+                ("unmarked-offline".to_string(), false),
+            ],
+            "both pending changes push, with their direction preserved"
+        );
+        assert_eq!(pushed.len(), 2);
+        assert_eq!(pending_flag(&db, "marked-offline").await, Some(0));
+        assert_eq!(pending_flag(&db, "unmarked-offline").await, Some(0));
+    }
+
+    /// A push that fails keeps its row pending, so the change is retried rather
+    /// than dropped on the floor.
+    ///
+    /// TRACES: UR-069 | DR-120 | UT-103
+    #[tokio::test]
+    async fn test_drain_leaves_failed_pushes_pending() {
+        let db = test_db();
+        seed(&db, &[("u1", "ok", 1, 1), ("u1", "boom", 1, 1)]).await;
+
+        let sink = RecordingSink::failing_for("boom");
+        let pushed = drain_pending_favorites(&db, &sink, "u1").await.unwrap();
+
+        assert_eq!(pushed, vec!["ok".to_string()]);
+        assert_eq!(pending_flag(&db, "ok").await, Some(0));
+        assert_eq!(
+            pending_flag(&db, "boom").await,
+            Some(1),
+            "a failed push must stay queued for the next reconnect"
+        );
+    }
+
+    /// Another user's queued changes are not pushed with this user's token.
+    ///
+    /// TRACES: UR-069 | DR-120 | UT-103
+    #[tokio::test]
+    async fn test_drain_only_touches_the_given_user() {
+        let db = test_db();
+        seed(&db, &[("u1", "mine", 1, 1), ("u2", "theirs", 1, 1)]).await;
+
+        let sink = RecordingSink::new();
+        let pushed = drain_pending_favorites(&db, &sink, "u1").await.unwrap();
+
+        assert_eq!(pushed, vec!["mine".to_string()]);
+        assert_eq!(pending_flag(&db, "theirs").await, Some(1));
+    }
+
+    /// Nothing pending means no server calls at all — a reconnect must not
+    /// generate traffic just because it happened.
+    ///
+    /// TRACES: UR-069 | DR-120 | UT-103
+    #[tokio::test]
+    async fn test_drain_is_a_noop_when_nothing_is_pending() {
+        let db = test_db();
+        seed(&db, &[("u1", "synced", 1, 0)]).await;
+
+        let sink = RecordingSink::new();
+        let pushed = drain_pending_favorites(&db, &sink, "u1").await.unwrap();
+
+        assert!(pushed.is_empty());
+        assert!(sink.calls().is_empty());
+    }
+}

src-tauri/src/commands/mod.rs

@@ -7,6 +7,7 @@ pub mod connectivity;
 pub mod conversions;
 pub mod device;
 pub mod download;
+pub mod favorites;
 pub mod offline;
 pub mod playback_mode;
 pub mod playback_reporting;
@@ -16,6 +17,7 @@ pub mod repository;
 pub mod sessions;
 pub mod storage;
 pub mod sync;
+pub mod sync_drain;
 
 pub use auth::*;
 pub use catalog::*;
@@ -33,3 +35,4 @@ pub use repository::{RepositoryManager, RepositoryManagerWrapper, *};
 pub use sessions::*;
 pub use storage::*;
 pub use sync::*;
+pub use sync_drain::*;

src-tauri/src/commands/player/mod.rs

@@ -57,18 +57,6 @@ pub struct MediaSessionManagerWrapper(pub Mutex<MediaSessionManager>);
 /// @req: DR-048 - Video settings (auto-play toggle, countdown duration)
 pub struct VideoSettingsWrapper(pub Mutex<VideoSettings>);
 
-/// Base offset (seconds) for the active background-audio handoff.
-///
-/// The audio-only stream is requested with `StartTimeTicks` = the handoff
-/// position, so the server makes that point the stream's zero. ExoPlayer then
-/// reports position RELATIVE to that zero. To convert back to an absolute
-/// position on exit (so the video resumes where the audio actually reached), we
-/// add this stored base to the native player's reported position.
-///
-/// TRACES: UR-040 | DR-052
-#[derive(Default)]
-pub struct BackgroundAudioOffset(pub Mutex<f64>);
-
 /// Response for player state queries
 #[derive(specta::Type, Debug, Serialize)]
 #[serde(rename_all = "camelCase")]
@@ -205,6 +193,36 @@ pub struct PlayItemRequest {
     /// zero-duration session renders no scrubber, even with ACTION_SEEK_TO set.
     #[serde(default)]
     pub duration_seconds: Option<f64>,
+    /// Item type (e.g. "Episode", "Movie", "Audio"). Carried through the
+    /// background-audio handoff so an episode played as audio-only is still
+    /// recognised as an episode by autoplay (UR-040) and advances to the next one.
+    #[serde(default)]
+    pub item_type: Option<String>,
+    /// Series ID for TV episodes. Needed alongside `item_type` so the backend can
+    /// look up the next episode when a background-audio track ends.
+    #[serde(default)]
+    pub series_id: Option<String>,
+    /// Subtitle tracks to sideload, with URLs the frontend has already resolved.
+    ///
+    /// Only the native backends use these: on Android they become the
+    /// `MediaItem.SubtitleConfiguration`s ExoPlayer renders. The HTML5 path
+    /// builds its own `<track>` children instead and ignores this list.
+    ///
+    /// **Order is the contract.** `player_set_subtitle_track(n)` reaches
+    /// `JellyTauPlayer.setSubtitleTrack(n)`, which indexes into ExoPlayer's
+    /// *text track groups* — i.e. the position of the sideloaded configuration,
+    /// not the Jellyfin stream index (which is kept on each entry for the UI's
+    /// benefit). So `n` must be a position in this very array, and the array
+    /// must not be reordered or filtered between building it and sending it.
+    /// `nativeSubtitleArrayIndex()` on the frontend computes `n` from the same
+    /// list that is sent here, for exactly this reason.
+    ///
+    /// Defaulted so the background-audio handoff and the autoplay/next-episode
+    /// callers, which have no subtitles to offer, need not send the field.
+    ///
+    /// TRACES: UR-020 | IR-016, JA-008 | UT-145
+    #[serde(default)]
+    pub subtitles: Vec<crate::player::SubtitleTrack>,
 }
 
 /// Queue context for remote transfer - what type of queue is this?
@@ -324,6 +342,29 @@ pub enum AudioTrackSwitchResponse {
     },
 }
 
+/// Response for a mid-playback streaming-quality change.
+///
+/// Mirrors [`AudioTrackSwitchResponse`]: the backend decides whether the caller
+/// has to reload anything, so no strategy branch lives in the UI.
+///
+/// TRACES: UR-074 | DR-162
+#[derive(specta::Type, Debug, Serialize)]
+#[serde(tag = "strategy", rename_all = "camelCase")]
+pub enum StreamQualityResponse {
+    /// The native backend was reloaded here; nothing left for the frontend.
+    Native {
+        /// Position playback resumed at.
+        position: f64,
+    },
+    /// HTML5 must reload its element with this URL.
+    ReloadStream {
+        /// New stream URL, already transcoded to the requested ceiling.
+        new_url: String,
+        /// Position to resume from.
+        position: f64,
+    },
+}
+
 /// Helper function to create MediaItem from video request
 ///
 /// PlayItemRequest is now video-only, so we create a video MediaItem.
@@ -376,12 +417,124 @@ pub(super) async fn create_media_item(
         needs_transcoding: req.needs_transcoding,
         video_width: None,  // Not available from video-only request
         video_height: None, // Not available from video-only request
-        subtitles: vec![],
+        // Sideloaded subtitles, in the order the frontend sent them — that order
+        // is what `player_set_subtitle_track(n)` indexes into on Android.
+        // TRACES: UR-020 | IR-016 | UT-145
+        subtitles: req.subtitles,
         series_id: None, // Not available from video-only request
         server_id: None, // Not available from video-only request
     })
 }
 
+/// Pick the source for an audio-only handoff.
+///
+/// A downloaded file wins over the audio-only stream URL. No transcode or audio
+/// extraction is involved or wanted: the native backends already play a video
+/// container without decoding its video — the Linux MPV backend is configured
+/// with `video: no`, and ExoPlayer simply has no surface to render to when the
+/// item is `MediaType::Audio`. Producing a separate audio-only file would cost
+/// CPU and battery, need an encoder the project does not ship, and leave a
+/// second artifact to keep in step with the first.
+///
+/// TRACES: UR-071 | DR-128 | UT-119
+pub(super) fn background_audio_source(
+    local_path: Option<String>,
+    stream_url: String,
+    item_id: &str,
+) -> MediaSource {
+    match local_path {
+        Some(path) => MediaSource::Local {
+            file_path: PathBuf::from(path),
+            jellyfin_item_id: Some(item_id.to_string()),
+        },
+        None => MediaSource::Remote {
+            stream_url,
+            jellyfin_item_id: item_id.to_string(),
+        },
+    }
+}
+
+/// How a background-audio handoff must start playback, given where its audio
+/// actually begins.
+///
+/// TRACES: UR-040, UR-071 | DR-180 | UT-181
+pub(super) struct BackgroundAudioPlan {
+    /// The position the stream's own zero corresponds to, recorded as the
+    /// handoff base so later readings can be shifted back to the episode's
+    /// timeline.
+    pub base_seconds: f64,
+    /// Where to seek after loading, if the source does not already start there.
+    pub seek_to: Option<f64>,
+}
+
+/// Decide the base and the seek for a handoff at `position_seconds`.
+///
+/// The two sources start in different places. An audio-only **stream** is built
+/// with `StartTimeTicks`, so the server makes the handoff point that stream's
+/// zero: the base is the handoff position, and seeking would skip *past* the
+/// content by that much again. A downloaded **file** has no such parameter and
+/// begins at the episode's own zero, so it needs the opposite — no base, and a
+/// real seek. Treating a file like a stream is why backgrounding a downloaded
+/// episode restarted it from 0:00 while the lockscreen showed the right time.
+///
+/// TRACES: UR-040, UR-071 | DR-180 | UT-181
+pub(super) fn background_audio_plan(
+    is_local_file: bool,
+    position_seconds: f64,
+) -> BackgroundAudioPlan {
+    let position = position_seconds.max(0.0);
+
+    if is_local_file {
+        BackgroundAudioPlan {
+            base_seconds: 0.0,
+            seek_to: (position > 0.0).then_some(position),
+        }
+    } else {
+        BackgroundAudioPlan {
+            base_seconds: position,
+            seek_to: None,
+        }
+    }
+}
+
+/// Resolve the on-disk file backing a completed download, if there is one.
+///
+/// A `downloads` row is not proof of a file: it can outlive the bytes (manual
+/// deletion, a cleared cache directory, a restored database). Every caller wants
+/// "can I play this from disk right now", so existence is checked here rather
+/// than trusted from the row.
+///
+/// Split out from [`check_for_local_download`] so the resolution is testable
+/// without a `DatabaseWrapper`, and reusable by the video path.
+///
+/// TRACES: UR-071 | DR-123 | UT-116
+pub(super) async fn resolve_local_media_path<S: DatabaseService>(
+    db_service: &Arc<S>,
+    item_id: &str,
+) -> Result<Option<String>, String> {
+    let query = Query::with_params(
+        "SELECT file_path FROM downloads WHERE item_id = ? AND status = 'completed' LIMIT 1",
+        vec![QueryParam::String(item_id.to_string())],
+    );
+
+    let path: Option<String> = db_service
+        .query_optional(query, |row| row.get(0))
+        .await
+        .map_err(|e| e.to_string())?;
+
+    match path {
+        Some(ref file_path) if std::path::Path::new(file_path).exists() => Ok(path),
+        Some(file_path) => {
+            warn!(
+                "[Player] Download entry exists in DB but file not found: {}",
+                file_path
+            );
+            Ok(None)
+        }
+        None => Ok(None),
+    }
+}
+
 /// Check if an item has a completed download
 pub(super) async fn check_for_local_download(
     db: &DatabaseWrapper,
@@ -392,30 +545,33 @@ pub(super) async fn check_for_local_download(
         Arc::new(database.service())
     };
 
-    let query = Query::with_params(
-        "SELECT file_path FROM downloads WHERE item_id = ? AND status = 'completed' LIMIT 1",
-        vec![QueryParam::String(item_id.to_string())],
-    );
+    resolve_local_media_path(&db_service, item_id).await
+}
 
-    let path: Option<String> = db_service
-        .query_optional(query, |row| row.get(0))
-        .await
-        .map_err(|e| e.to_string())?;
+/// The on-disk path for a downloaded item, for playback surfaces that resolve
+/// their own source rather than going through the queue.
+///
+/// The video player is the reason this exists: audio has preferred local files
+/// since queue construction, but video asks the repository for a stream URL and
+/// never consults `downloads`, so a downloaded film was still streamed — costing
+/// bandwidth that had already been spent and failing outright when offline.
+///
+/// Returns `None` when nothing is downloaded *or* the file is missing, so the
+/// caller falls back to streaming.
+///
+/// TRACES: UR-071 | DR-123 | UT-116
+#[tauri::command]
+#[specta::specta]
+pub async fn player_local_media_path(
+    db: State<'_, DatabaseWrapper>,
+    item_id: String,
+) -> Result<Option<String>, String> {
+    let db_service = {
+        let database = db.0.lock().map_err(|e| e.to_string())?;
+        Arc::new(database.service())
+    };
 
-    // Verify the file actually exists on disk
-    if let Some(ref file_path) = path {
-        if std::path::Path::new(file_path).exists() {
-            Ok(path)
-        } else {
-            warn!(
-                "[Player] Download entry exists in DB but file not found: {}",
-                file_path
-            );
-            Ok(None)
-        }
-    } else {
-        Ok(None)
-    }
+    resolve_local_media_path(&db_service, &item_id).await
 }
 
 /// Re-point queued streaming items at completed local downloads.
@@ -577,7 +733,7 @@ pub async fn player_play_item(
 pub async fn player_enter_background_audio(
     player: State<'_, PlayerStateWrapper>,
     session: State<'_, MediaSessionManagerWrapper>,
-    bg_offset: State<'_, BackgroundAudioOffset>,
+    db: State<'_, DatabaseWrapper>,
     item: PlayItemRequest,
     position_seconds: f64,
 ) -> Result<PlayerStatus, String> {
@@ -586,6 +742,22 @@ pub async fn player_enter_background_audio(
         item.title, position_seconds
     );
 
+    // Prefer the downloaded file over the audio-only stream URL the frontend
+    // resolved. Handing the native backend a local video container yields
+    // audio-only playback for free — no transcode, no second artifact.
+    // TRACES: UR-071 | DR-128
+    let local_path = check_for_local_download(&db, &item.id).await?;
+    if local_path.is_some() {
+        info!(
+            "player_enter_background_audio: using downloaded file for {}",
+            item.id
+        );
+    }
+    // A downloaded file starts at the episode's zero; a stream starts at the
+    // handoff point. Only one of them has a base, and only the other needs a seek.
+    let plan = background_audio_plan(local_path.is_some(), position_seconds);
+    let source = background_audio_source(local_path, item.stream_url, &item.id);
+
     // Build an AUDIO media item pointing at the audio-only stream. We do not use
     // create_media_item() because that hardcodes MediaType::Video; background
     // audio must be Audio so no video decode is started.
@@ -601,22 +773,21 @@ pub async fn player_enter_background_audio(
         artists: None,
         primary_image_tag: item.primary_image_tag.clone(),
         image_id: item.primary_image_tag.clone(),
-        item_type: None,
+        // Carry episode identity so autoplay can advance to the next episode when
+        // this audio-only handoff ends while backgrounded (UR-040).
+        item_type: item.item_type.clone(),
         playlist_id: None,
         // Carry the real duration so the lockscreen MediaSession can draw a scrubber.
         duration: item.duration_seconds,
         artwork_url: None,
         media_type: MediaType::Audio,
-        source: MediaSource::Remote {
-            stream_url: item.stream_url,
-            jellyfin_item_id: item.id.clone(),
-        },
+        source,
         video_codec: None,
         needs_transcoding: false,
         video_width: None,
         video_height: None,
         subtitles: vec![],
-        series_id: None,
+        series_id: item.series_id.clone(),
         server_id: item.server_id.clone(),
     };
 
@@ -625,23 +796,31 @@ pub async fn player_enter_background_audio(
         session_mgr.start_audio_session(media_item.clone());
     }
 
-    // Remember where the video was: the audio stream's zero == this position
-    // (the URL was built with StartTimeTicks=position_seconds), so on exit we add
-    // this base to the native player's relative position to get the absolute one.
-    *bg_offset.0.lock().map_err(|e| e.to_string())? = position_seconds.max(0.0);
-
     // Same base offset drives the lockscreen scrubber: ExoPlayer reports position
     // relative to the stream's StartTimeTicks zero, but the metadata duration is
     // absolute, so shift the reported position back to absolute for the scrubber.
-    let _ = crate::player::set_lockscreen_position_offset(position_seconds.max(0.0));
+    let _ = crate::player::set_lockscreen_position_offset(plan.base_seconds);
 
     let controller = player.0.lock().await;
+    // Remember where the video was: for a stream the audio's zero == this
+    // position (the URL was built with StartTimeTicks=position_seconds), so on
+    // exit we add this base to the native player's relative position to get the
+    // absolute one. The controller owns it so a backend-driven advance to the
+    // next episode clears it along with the stream it described.
+    controller.enter_background_audio(plan.base_seconds);
     controller
         .play_item(media_item)
         .map_err(|e| e.to_string())?;
-    // NOTE: do NOT seek here. The audio-only URL already starts at the handoff
-    // position via StartTimeTicks; the stream's timeline begins at 0 == that
-    // point, so an extra seek(position_seconds) would jump PAST the content.
+    // Seek ONLY a local file. The audio-only URL already starts at the handoff
+    // position via StartTimeTicks — its timeline begins at 0 == that point — so
+    // seeking a stream would jump PAST the content by the handoff position again.
+    if let Some(seek_to) = plan.seek_to {
+        info!(
+            "player_enter_background_audio: seeking the downloaded file to {:.1}s",
+            seek_to
+        );
+        controller.seek(seek_to).map_err(|e| e.to_string())?;
+    }
 
     controller.emit_queue_changed();
     if let Some(emitter) = controller.event_emitter() {
@@ -666,29 +845,31 @@ pub async fn player_enter_background_audio(
 #[specta::specta]
 pub async fn player_exit_background_audio(
     player: State<'_, PlayerStateWrapper>,
-    bg_offset: State<'_, BackgroundAudioOffset>,
 ) -> Result<f64, String> {
-    // The base offset (handoff position) + native player's relative position =
-    // the absolute position to resume the video at. Read/reset the base first.
-    let base = {
-        let mut off = bg_offset.0.lock().map_err(|e| e.to_string())?;
-        let b = *off;
-        *off = 0.0;
-        b
-    };
-
-    // Back to foreground playback: the lockscreen scrubber is absolute again.
-    let _ = crate::player::set_lockscreen_position_offset(0.0);
-
     let controller = player.0.lock().await;
-    // Capture position into a `let` BEFORE stop() — never hold work across a lock
-    // re-entrant call (deadlock discipline, CLAUDE.md).
-    let relative = controller.position();
+
+    // Read the position BEFORE clearing either base. The position tick applies the
+    // base natively, so a tick landing between "base cleared" and "position read"
+    // would hand back a relative position — the whole bug, reintroduced at the one
+    // moment it matters most. Capturing into a `let` before stop() is also the
+    // lock discipline from CLAUDE.md: never hold work across a re-entrant call.
+    // (DR-159)
+    //
+    // `absolute_position` rather than `position`, because a tick that has not
+    // landed *yet* is the same hazard from the other side: returning to the
+    // foreground while the audio-only transcode is still opening read 0.0, and
+    // the video reloaded at StartTimeTicks=0 — the episode restarting from the
+    // beginning. Flooring at the handoff base cannot overshoot: the stream is
+    // physically incapable of being behind its own starting point. (DR-178)
+    let absolute = controller.absolute_position();
+
+    // Now safe to tear the handoff down, native side first.
+    let _ = crate::player::set_lockscreen_position_offset(0.0);
+    controller.exit_background_audio();
     controller.stop().map_err(|e| e.to_string())?;
-    let absolute = base + relative;
     info!(
-        "player_exit_background_audio: base={:.1}s + relative={:.1}s = {:.1}s",
-        base, relative, absolute
+        "player_exit_background_audio: resuming the video at {:.1}s",
+        absolute
     );
     Ok(absolute)
 }
@@ -906,6 +1087,16 @@ pub async fn player_stop(
                 .clone()
         };
         client.send_session_command(session_id, "Stop").await?;
+
+        // Stopping the remote session ends the cast, so the manager returns to
+        // Idle — same as a local stop. This is also what hands OS volume control
+        // back to this device: set_mode releases the Android remote volume
+        // provider on any exit from remote mode. Without it the mode stayed
+        // Remote and the system volume slider remained stuck on the remote
+        // session with no way back to the local speaker.
+        playback_mode
+            .0
+            .set_mode(crate::playback_mode::PlaybackMode::Idle);
     } else {
         // Local playback
         let controller = player.0.lock().await;
@@ -1100,9 +1291,12 @@ pub async fn player_seek(
         let position_ticks = (position * 10_000_000.0) as i64;
         client.session_seek(session_id, position_ticks).await?;
     } else {
-        // Local playback
+        // Local playback. seek_absolute, not seek: the position came from the UI,
+        // which shows the whole item, so during a background-audio handoff it has
+        // to be resolved against the episode's timeline rather than the handoff
+        // stream's. (DR-159)
         let controller = player.0.lock().await;
-        controller.seek(position).map_err(|e| e.to_string())?;
+        controller.seek_absolute(position).await?;
     }
 
     let controller = player.0.lock().await;
@@ -1198,7 +1392,6 @@ pub async fn player_seek_video(
                 .get_video_stream_url(
                     &jellyfin_item_id,
                     media_source_id.as_deref(),
-                    Some(position),
                     audio_stream_index,
                 )
                 .await
@@ -1209,6 +1402,13 @@ pub async fn player_seek_video(
                 position
             );
 
+            // `seek_offset` carries the position to RESUME AT, not a base to add
+            // to the element's clock. The reloaded stream starts at the item's
+            // zero — a position on an HLS playlist makes the server 400 every
+            // segment behind it (DR-181) — so the adapter reaches the position by
+            // seeking the element and leaves the transcode offset at zero. The
+            // field keeps its name only because renaming it means regenerating
+            // the specta bindings; `reloadSource` documents the contract.
             Ok(VideoSeekResponse::ReloadStream {
                 new_url,
                 seek_offset: position,
@@ -1222,7 +1422,6 @@ pub async fn player_seek_video(
                 .get_video_stream_url(
                     &jellyfin_item_id,
                     media_source_id.as_deref(),
-                    Some(position),
                     audio_stream_index,
                 )
                 .await
@@ -1260,6 +1459,11 @@ pub async fn player_seek_video(
                 } else {
                     return Err("No current item after URL update".to_string());
                 }
+
+                // The re-opened stream begins at zero — the position cannot ride
+                // along in the URL without 400ing every segment (DR-181) — so the
+                // seek that the reload was asked for happens here.
+                controller.seek(position).map_err(|e| e.to_string())?;
             }
 
             info!(
@@ -1310,12 +1514,13 @@ pub async fn player_switch_audio_track(
                 .to_string()
         };
 
-        // Get new stream URL with selected audio track
+        // Get new stream URL with selected audio track. It starts at zero — an
+        // HLS playlist cannot carry a position (DR-181) — and `position` below
+        // tells the frontend where to seek the reloaded element back to.
         let new_url = repository
             .get_video_stream_url(
                 &jellyfin_item_id,
                 media_source_id.as_deref(),
-                current_position,
                 Some(stream_index),
             )
             .await
@@ -1336,6 +1541,115 @@ pub async fn player_switch_audio_track(
     }
 }
 
+/// Change the bandwidth ceiling of the video that is playing *right now*.
+///
+/// A cap is a property of the stream the server is producing, so unlike a volume
+/// change it cannot be applied to a stream already in flight — the stream has to
+/// be re-opened at the new quality and resumed at the current position. That is
+/// the same reload the transcoded-seek and audio-track paths use, and the same
+/// two-sided split: HTML5 gets the URL back and reloads its own element, while a
+/// native backend is reloaded here.
+///
+/// The change applies to this playback *and* to everything started afterwards
+/// (it sets the process-wide ceiling), but it is deliberately **not** persisted:
+/// the in-player picker is a "this film, this connection" control, and the
+/// durable default belongs to Settings. `player_set_video_settings` is the one
+/// that writes to the database.
+///
+/// TRACES: UR-074 | DR-162
+#[tauri::command]
+#[specta::specta]
+pub async fn player_set_stream_quality(
+    player: State<'_, PlayerStateWrapper>,
+    repository_manager: State<'_, super::repository::RepositoryManagerWrapper>,
+    video_settings: State<'_, VideoSettingsWrapper>,
+    repository_handle: String,
+    quality: crate::settings::StreamingQuality,
+    use_html5: bool,
+    current_position: Option<f64>,
+    media_source_id: Option<String>,
+    audio_stream_index: Option<i32>,
+) -> Result<StreamQualityResponse, String> {
+    info!(
+        "[player_set_stream_quality] Switching to {} (use_html5: {}, position: {:?})",
+        quality.label(),
+        use_html5,
+        current_position
+    );
+
+    let repository = repository_manager
+        .0
+        .get(&repository_handle)
+        .ok_or("Repository not found - user may need to log in")?;
+
+    let jellyfin_item_id = {
+        let controller = player.0.lock().await;
+        let queue_arc = controller.queue();
+        let queue = queue_arc.lock().map_err(|e| e.to_string())?;
+
+        let current_item = queue.current().ok_or("No item currently playing")?;
+
+        if current_item.media_type != MediaType::Video {
+            return Err("Current item is not a video".to_string());
+        }
+
+        current_item
+            .jellyfin_id()
+            .ok_or("Current item has no Jellyfin ID")?
+            .to_string()
+    };
+
+    // Set the ceiling *before* building the URL — the builder reads it.
+    crate::repository::online::set_streaming_quality(quality);
+    {
+        let mut settings = video_settings.0.lock().map_err(|e| e.to_string())?;
+        settings.streaming_quality = quality;
+    }
+
+    let position = current_position.unwrap_or(0.0);
+    let new_url = repository
+        .get_video_stream_url(
+            &jellyfin_item_id,
+            media_source_id.as_deref(),
+            audio_stream_index,
+        )
+        .await
+        .map_err(|e| format!("Failed to get video stream URL: {:?}", e))?;
+
+    if use_html5 {
+        return Ok(StreamQualityResponse::ReloadStream { new_url, position });
+    }
+
+    // Native backend (Android/ExoPlayer): stop, repoint the queue entry at the
+    // new URL, and reload — mirroring `VideoSeekStrategy::BackendReloadStream`.
+    // The re-opened stream begins at zero (an HLS playlist cannot carry a start
+    // position without 400ing every segment — DR-181), so it is seeked back to
+    // where the picture was.
+    {
+        let controller = player.0.lock().await;
+        controller.stop().map_err(|e| e.to_string())?;
+
+        let queue_arc = controller.queue();
+        {
+            let mut queue = queue_arc.lock().map_err(|e| e.to_string())?;
+            if !queue.update_current_stream_url(new_url.clone()) {
+                return Err("Failed to update stream URL in queue".to_string());
+            }
+        }
+
+        let queue = queue_arc.lock().map_err(|e| e.to_string())?;
+        let updated_item = queue.current().ok_or("No current item after URL update")?;
+        controller
+            .load_and_play(updated_item)
+            .map_err(|e| e.to_string())?;
+        if position > 0.0 {
+            controller.seek(position).map_err(|e| e.to_string())?;
+        }
+    }
+
+    Ok(StreamQualityResponse::Native { position })
+}
+
 #[tauri::command]
 #[specta::specta]
 pub async fn player_set_audio_track(
@@ -1557,6 +1871,43 @@ pub async fn player_get_queue(
     Ok(get_queue_status(&controller))
 }
 
+/// What playback facilities this platform's backend actually provides.
+///
+/// The frontend is presentation-only and must not re-derive backend facts from
+/// `navigator.userAgent` — that sniffing was a second copy of the same platform
+/// decision Rust already makes with `cfg!`, and it drifted. These flags are the
+/// single source of truth; the frontend consumes them.
+///
+/// TRACES: UR-003, UR-005 | DR-004, DR-023, DR-024
+#[derive(specta::Type, Debug, Serialize)]
+#[serde(rename_all = "camelCase")]
+pub struct PlaybackCapabilities {
+    /// True when audio is rendered by a webview `<audio>` element rather than a
+    /// native backend. Native audio exists on Linux (mpv) and Android
+    /// (ExoPlayer); everything else (Windows, future desktops) uses the webview.
+    pub uses_webview_audio: bool,
+    /// True when video can be rendered by a native surface composited *behind*
+    /// a transparent webview. Android only: ExoPlayer draws into a SurfaceView
+    /// beneath the WebView. Linux cannot do this (WebKitGTK/Wayland
+    /// compositing), so it stays on the HTML5 element.
+    pub supports_native_video: bool,
+}
+
+/// Report this platform's playback capabilities to the frontend.
+///
+/// TRACES: UR-003, UR-005 | DR-004, DR-023, DR-024
+#[tauri::command]
+#[specta::specta]
+pub async fn player_get_capabilities() -> Result<PlaybackCapabilities, String> {
+    // Mirrors the cfg gates the backends themselves are built under.
+    let native_audio = cfg!(any(target_os = "android", target_os = "linux"));
+
+    Ok(PlaybackCapabilities {
+        uses_webview_audio: !native_audio,
+        supports_native_video: cfg!(target_os = "android"),
+    })
+}
+
 pub(super) fn get_player_status(controller: &PlayerController) -> PlayerStatus {
     // Determine backend at compile time based on platform
     let (backend, use_html5_element) = if cfg!(target_os = "android") {
@@ -1569,7 +1920,11 @@ pub(super) fn get_player_status(controller: &PlayerController) -> PlayerStatus {
 
     PlayerStatus {
         state: controller.state(),
-        position: controller.position(),
+        // The position on the item's timeline, whichever of the three paths is
+        // rendering it — the native backend answers for only one of them, and
+        // reads 0 for webview video and for a handoff that has not ticked yet.
+        // TRACES: UR-005 | DR-178
+        position: controller.absolute_position(),
         duration: controller.duration(),
         volume: controller.volume(),
         muted: controller.muted(),
@@ -2386,6 +2741,306 @@ pub async fn player_disable_jellyfin(player: State<'_, PlayerStateWrapper>) -> R
 
 #[cfg(test)]
 mod tests {
+    /// The subtitle list the frontend resolved must survive the IPC hop and end
+    /// up on the `MediaItem` the native backend loads.
+    ///
+    /// The bug: `VideoPlayer.svelte` built a fully-resolved subtitle array and
+    /// then dropped it on the floor — `PlayItemRequest` had no field to put it
+    /// in — so `create_media_item` always produced `subtitles: vec![]`,
+    /// `android/mod.rs` serialized `[]` across JNI, and ExoPlayer was handed a
+    /// `MediaItem` with zero `SubtitleConfiguration`s. Every later
+    /// `setSubtitleTrack(n)` then found no text track groups and logged
+    /// "Invalid subtitle track index".
+    ///
+    /// The payload below is exactly what the frontend sends: camelCase for the
+    /// top-level command params (Tauri v2 converts them), and the subtitle
+    /// entries in the casing of `SubtitleTrack` itself — note `mime_type`.
+    ///
+    /// TRACES: UR-020 | IR-016 | UT-145
+    #[tokio::test]
+    async fn test_play_item_request_carries_subtitles_into_media_item() {
+        use super::{create_media_item, PlayItemRequest};
+
+        let payload = serde_json::json!({
+            "id": "ep-1",
+            "title": "Pilot",
+            "streamUrl": "https://jelly.example/Videos/ep-1/master.m3u8",
+            "videoCodec": "h264",
+            "needsTranscoding": false,
+            "subtitles": [
+                {
+                    "index": 2,
+                    "url": "https://jelly.example/Videos/ep-1/2/Subtitles/subtitles.vtt",
+                    "language": "eng",
+                    "label": "English (SRT)",
+                    "mime_type": "text/vtt"
+                },
+                {
+                    "index": 3,
+                    "url": "https://jelly.example/Videos/ep-1/3/Subtitles/subtitles.vtt",
+                    "language": null,
+                    "label": null,
+                    "mime_type": "text/vtt"
+                }
+            ]
+        });
+
+        let req: PlayItemRequest =
+            serde_json::from_value(payload).expect("frontend payload must deserialize");
+        assert_eq!(
+            req.subtitles.len(),
+            2,
+            "PlayItemRequest must carry the subtitle tracks, not silently ignore them"
+        );
+
+        let media = create_media_item(req, None).await.unwrap();
+        assert_eq!(
+            media.subtitles.len(),
+            2,
+            "create_media_item must thread the tracks onto the MediaItem the backend loads"
+        );
+        assert_eq!(media.subtitles[0].index, 2);
+        assert_eq!(media.subtitles[0].language.as_deref(), Some("eng"));
+        assert_eq!(media.subtitles[0].label.as_deref(), Some("English (SRT)"));
+        assert_eq!(media.subtitles[0].mime_type, "text/vtt");
+        // Order is the contract: `player_set_subtitle_track(n)` is a position in
+        // this list (see the note on `PlayItemRequest::subtitles`).
+        assert_eq!(media.subtitles[1].index, 3);
+        assert!(media.subtitles[1].language.is_none());
+    }
+
+    /// A request without subtitles must still deserialize — the field is
+    /// defaulted so the background-audio handoff and the autoplay/next-episode
+    /// callers keep compiling and sending what they always sent.
+    ///
+    /// TRACES: UR-020 | IR-016 | UT-145
+    #[tokio::test]
+    async fn test_play_item_request_without_subtitles_defaults_to_empty() {
+        use super::{create_media_item, PlayItemRequest};
+
+        let req: PlayItemRequest = serde_json::from_value(serde_json::json!({
+            "id": "movie-1",
+            "title": "Movie",
+            "streamUrl": "https://jelly.example/Videos/movie-1/stream.mp4",
+            "videoCodec": "h264",
+            "needsTranscoding": false
+        }))
+        .expect("a subtitle-less payload must still deserialize");
+
+        assert!(req.subtitles.is_empty());
+        assert!(create_media_item(req, None)
+            .await
+            .unwrap()
+            .subtitles
+            .is_empty());
+    }
+
+    /// The JSON handed to Kotlin over JNI must use the keys
+    /// `JellyTauPlayer.load()` actually reads.
+    ///
+    /// `MediaItem` is `rename_all = "camelCase"`, and the instinct (and the
+    /// house IPC rule) is to camelCase nested structs too — but
+    /// `JellyTauPlayer.kt` reads `subtitle.optString("mime_type", …)`. Renaming
+    /// the field would not fail to compile or fail the IPC; it would silently
+    /// fall back to the default MIME type for every track, so this is asserted
+    /// on the exact bytes `android/mod.rs` sends.
+    ///
+    /// TRACES: UR-020 | IR-016, JA-008 | UT-146
+    #[test]
+    fn test_subtitle_json_for_jni_uses_the_keys_kotlin_reads() {
+        use crate::player::media::SubtitleTrack;
+
+        let subtitles = vec![SubtitleTrack {
+            index: 2,
+            url: "https://jelly.example/subs.vtt".to_string(),
+            language: Some("eng".to_string()),
+            label: Some("English".to_string()),
+            mime_type: "text/vtt".to_string(),
+        }];
+
+        // Exactly what player/android/mod.rs passes to loadWithMetadata.
+        let json = serde_json::to_string(&subtitles).unwrap();
+        let parsed: Vec<serde_json::Value> = serde_json::from_str(&json).unwrap();
+        let obj = parsed[0].as_object().unwrap();
+
+        for key in ["url", "language", "label", "mime_type"] {
+            assert!(
+                obj.contains_key(key),
+                "JellyTauPlayer.load() reads `{key}`; serialized keys were {:?}",
+                obj.keys().collect::<Vec<_>>()
+            );
+        }
+        assert!(
+            !obj.contains_key("mimeType"),
+            "camelCasing mime_type silently drops every track's MIME type on Android"
+        );
+    }
+
+    /// The audio-only handoff must play a downloaded file when there is one,
+    /// rather than fetching an audio-only stream for media already on disk.
+    ///
+    /// TRACES: UR-071 | DR-128 | UT-119
+    #[test]
+    fn test_background_audio_source_prefers_local_file() {
+        use super::background_audio_source;
+        use crate::player::MediaSource;
+        use std::path::PathBuf;
+
+        let local = background_audio_source(
+            Some("/downloads/ep1.mkv".to_string()),
+            "https://server/audio-only".to_string(),
+            "ep-1",
+        );
+        match local {
+            MediaSource::Local {
+                file_path,
+                jellyfin_item_id,
+            } => {
+                assert_eq!(file_path, PathBuf::from("/downloads/ep1.mkv"));
+                // The Jellyfin id must survive so progress still syncs back.
+                assert_eq!(jellyfin_item_id.as_deref(), Some("ep-1"));
+            }
+            other => panic!("expected a local source, got {:?}", other),
+        }
+
+        let remote = background_audio_source(None, "https://server/audio-only".to_string(), "ep-1");
+        match remote {
+            MediaSource::Remote {
+                stream_url,
+                jellyfin_item_id,
+            } => {
+                assert_eq!(stream_url, "https://server/audio-only");
+                assert_eq!(jellyfin_item_id, "ep-1");
+            }
+            other => panic!("expected a remote source, got {:?}", other),
+        }
+    }
+
+    /// The two sources start in different places, so the handoff cannot treat
+    /// them alike.
+    ///
+    /// An audio-only *stream* is built with `StartTimeTicks`, so the server makes
+    /// the handoff point that stream's zero: the base is the handoff position and
+    /// seeking would jump past the content. A *downloaded file* has no such
+    /// parameter — it starts at the episode's own zero — so basing it at the
+    /// handoff position claims 18 minutes of audio that is about to play from the
+    /// beginning. That is the downloaded-episode version of "it restarts when the
+    /// screen sleeps", and it needs the opposite treatment: no base, and a seek.
+    ///
+    /// TRACES: UR-040, UR-071 | DR-180 | UT-181
+    #[test]
+    fn test_background_audio_plan_seeks_a_file_and_bases_a_stream() {
+        use super::background_audio_plan;
+
+        let local = background_audio_plan(true, 1104.0);
+        assert_eq!(local.base_seconds, 0.0);
+        assert_eq!(local.seek_to, Some(1104.0));
+
+        let streamed = background_audio_plan(false, 1104.0);
+        assert_eq!(streamed.base_seconds, 1104.0);
+        assert_eq!(
+            streamed.seek_to, None,
+            "the URL already starts at the handoff point; seeking again skips past it"
+        );
+    }
+
+    /// Handing off at the very start has nothing to seek to and nothing to base:
+    /// both sources are already where they need to be.
+    ///
+    /// TRACES: UR-040, UR-071 | DR-180 | UT-181
+    #[test]
+    fn test_background_audio_plan_at_the_start_neither_seeks_nor_bases() {
+        use super::background_audio_plan;
+
+        for local in [true, false] {
+            let plan = background_audio_plan(local, 0.0);
+            assert_eq!(plan.base_seconds, 0.0);
+            assert_eq!(plan.seek_to, None);
+        }
+    }
+
+    /// A downloaded item must resolve to its file, and a `downloads` row whose
+    /// file has gone must resolve to `None` so the caller falls back to
+    /// streaming instead of handing the player a path that cannot be opened.
+    ///
+    /// TRACES: UR-071 | DR-123 | UT-116
+    #[tokio::test]
+    async fn test_resolve_local_media_path() {
+        use super::resolve_local_media_path;
+        use crate::storage::db_service::{DatabaseService, Query, RusqliteService};
+        use rusqlite::Connection;
+        use std::sync::{Arc, Mutex};
+
+        let conn = Connection::open_in_memory().unwrap();
+        conn.execute(
+            "CREATE TABLE downloads (id INTEGER PRIMARY KEY, item_id TEXT, status TEXT, file_path TEXT)",
+            [],
+        )
+        .unwrap();
+        let db_service = Arc::new(RusqliteService::new(Arc::new(Mutex::new(conn))));
+
+        // A real file on disk, so the existence check passes.
+        let present = std::env::temp_dir().join("jellytau-resolve-local-test.mp4");
+        std::fs::write(&present, b"x").unwrap();
+        let present_str = present.to_string_lossy().to_string();
+
+        for (item, status, path) in [
+            ("downloaded", "completed", present_str.as_str()),
+            ("still-going", "downloading", present_str.as_str()),
+            (
+                "file-gone",
+                "completed",
+                "/nonexistent/jellytau/missing.mp4",
+            ),
+        ] {
+            db_service
+                .execute(Query::with_params(
+                    "INSERT INTO downloads (item_id, status, file_path) VALUES (?, ?, ?)",
+                    vec![
+                        crate::storage::db_service::QueryParam::String(item.to_string()),
+                        crate::storage::db_service::QueryParam::String(status.to_string()),
+                        crate::storage::db_service::QueryParam::String(path.to_string()),
+                    ],
+                ))
+                .await
+                .unwrap();
+        }
+
+        assert_eq!(
+            resolve_local_media_path(&db_service, "downloaded")
+                .await
+                .unwrap()
+                .as_deref(),
+            Some(present_str.as_str()),
+            "a completed download with its file present must resolve"
+        );
+
+        assert_eq!(
+            resolve_local_media_path(&db_service, "still-going")
+                .await
+                .unwrap(),
+            None,
+            "an in-progress download is not playable from disk"
+        );
+
+        assert_eq!(
+            resolve_local_media_path(&db_service, "file-gone")
+                .await
+                .unwrap(),
+            None,
+            "a row whose file has gone must fall back to streaming, not hand over a dead path"
+        );
+
+        assert_eq!(
+            resolve_local_media_path(&db_service, "never-heard-of-it")
+                .await
+                .unwrap(),
+            None
+        );
+
+        let _ = std::fs::remove_file(&present);
+    }
+
     /// Queue items enqueued as Remote must flip to Local once a completed
     /// download exists on disk — this is what makes preloaded tracks (and
     /// offline playback after a connection drop) actually use the cache.

src-tauri/src/commands/player/settings.rs

@@ -1,12 +1,29 @@
 //! Audio and video playback settings commands.
 //!
-//! TRACES: UR-022, UR-031, UR-032, UR-033 | DR-025, DR-034, DR-035, DR-036
+//! TRACES: UR-022, UR-027, UR-031, UR-032, UR-033, UR-074 | DR-025, DR-030, DR-034, DR-035, DR-036, DR-162, IR-020
 
-use tauri::State;
+use std::sync::Arc;
+
+use log::{info, warn};
+use tauri::{Manager, State};
 
 use super::{PlayerStateWrapper, VideoSettingsWrapper};
+use crate::commands::storage::DatabaseWrapper;
 use crate::player::AutoplaySettings;
-use crate::settings::{AudioSettings, VideoSettings};
+use crate::settings::{AudioSettings, EqPreset, StreamingQuality, VideoSettings};
+use crate::storage::db_service::{DatabaseService, Query, QueryParam};
+use crate::utils::lock::MutexSafe;
+
+/// `app_settings` key holding the persisted streaming bandwidth ceiling.
+///
+/// The cap is persisted (unlike the rest of `VideoSettings`, which is
+/// process-lifetime state) because forgetting it is the one failure that costs
+/// the user something real: a limit set for a metered connection that silently
+/// reverts to uncapped on the next launch spends their data allowance without
+/// ever showing them a changed setting.
+///
+/// TRACES: UR-074 | DR-162
+const STREAMING_QUALITY_KEY: &str = "streaming_quality";
 
 #[tauri::command]
 #[specta::specta]
@@ -14,13 +31,32 @@ pub async fn player_set_audio_settings(
     player: State<'_, PlayerStateWrapper>,
     settings: AudioSettings,
 ) -> Result<AudioSettings, String> {
+    // Validate/normalise domain values before applying: clamp crossfade to its
+    // range and normalise the equalizer band vector (length + gain clamps).
+    let validated = settings
+        .with_crossfade_clamped()
+        .with_equalizer_normalised();
     let mut controller = player.0.lock().await;
     controller
-        .set_audio_settings(&settings)
+        .set_audio_settings(&validated)
         .map_err(|e| e.to_string())?;
     Ok(controller.audio_settings())
 }
 
+/// The built-in equalizer presets and their per-band gain curves (dB), for the
+/// settings UI. The curve numbers are domain data defined by the band layout,
+/// so the frontend reads them here rather than encoding them.
+///
+/// TRACES: UR-027 | DR-030
+#[tauri::command]
+#[specta::specta]
+pub async fn player_get_eq_presets() -> Result<Vec<(EqPreset, Vec<f32>)>, String> {
+    Ok(EqPreset::ALL
+        .iter()
+        .map(|p| (*p, p.gains().to_vec()))
+        .collect())
+}
+
 #[tauri::command]
 #[specta::specta]
 pub async fn player_get_audio_settings(
@@ -35,6 +71,7 @@ pub async fn player_get_audio_settings(
 pub async fn player_set_video_settings(
     video_settings: State<'_, VideoSettingsWrapper>,
     player: State<'_, PlayerStateWrapper>,
+    db: State<'_, DatabaseWrapper>,
     settings: VideoSettings,
 ) -> Result<VideoSettings, String> {
     let validated = settings.with_countdown_clamped();
@@ -43,6 +80,12 @@ pub async fn player_set_video_settings(
         *current = validated.clone();
     } // Drop MutexGuard before await
 
+    // The bandwidth ceiling is read by the repository's URL builders and by the
+    // PlaybackInfo negotiation, neither of which can see this wrapper.
+    // TRACES: UR-074 | DR-162
+    crate::repository::online::set_streaming_quality(validated.streaming_quality);
+    persist_streaming_quality(&db, validated.streaming_quality).await;
+
     // Sync to PlayerController's autoplay settings so on_playback_ended() uses current values
     let controller = player.0.lock().await;
     controller.set_autoplay_settings(AutoplaySettings {
@@ -54,6 +97,110 @@ pub async fn player_set_video_settings(
     Ok(validated)
 }
 
+/// The bandwidth ceilings the quality picker may offer, each with the label and
+/// one-line detail to show for it, highest first.
+///
+/// The ladder and its numbers are Jellyfin encoding domain vocabulary, so the
+/// frontend reads them here rather than encoding them — the same arrangement as
+/// [`player_get_eq_presets`].
+///
+/// TRACES: UR-074 | DR-162
+#[tauri::command]
+#[specta::specta]
+pub async fn player_get_streaming_qualities(
+) -> Result<Vec<(StreamingQuality, String, String)>, String> {
+    Ok(StreamingQuality::ALL
+        .iter()
+        .map(|q| (*q, q.label().to_string(), q.detail().to_string()))
+        .collect())
+}
+
+/// Write the ceiling to `app_settings`. Failure is logged, not returned: the
+/// setting has already been applied in memory, and refusing the whole call
+/// because the write failed would leave the UI showing a cap that *is* active.
+///
+/// TRACES: UR-074 | DR-162
+async fn persist_streaming_quality(db: &State<'_, DatabaseWrapper>, quality: StreamingQuality) {
+    let db_service = {
+        let database = db.0.lock_safe();
+        Arc::new(database.service())
+    };
+
+    let encoded = match serde_json::to_string(&quality) {
+        Ok(value) => value,
+        Err(e) => {
+            warn!("[VideoSettings] Failed to encode streaming quality: {}", e);
+            return;
+        }
+    };
+
+    let query = Query::with_params(
+        "INSERT OR REPLACE INTO app_settings (key, value, updated_at)
+         VALUES (?, ?, CURRENT_TIMESTAMP)",
+        vec![
+            QueryParam::String(STREAMING_QUALITY_KEY.to_string()),
+            QueryParam::String(encoded),
+        ],
+    );
+
+    if let Err(e) = db_service.execute(query).await {
+        warn!("[VideoSettings] Failed to persist streaming quality: {}", e);
+    }
+}
+
+/// Restore the persisted bandwidth ceiling at startup, into both the repository
+/// (which enforces it) and `VideoSettings` (which the settings UI reads).
+///
+/// Called from the Tauri `setup` hook. A missing or unreadable row leaves the
+/// default — uncapped — in place, so a database problem degrades to the old
+/// behaviour rather than to an arbitrary limit.
+///
+/// TRACES: UR-074 | DR-162
+pub async fn restore_streaming_quality(app: &tauri::AppHandle) {
+    let db_service = {
+        let Some(db) = app.try_state::<DatabaseWrapper>() else {
+            warn!("[VideoSettings] No database available; streaming quality stays uncapped");
+            return;
+        };
+        let database = db.0.lock_safe();
+        Arc::new(database.service())
+    };
+
+    let query = Query::with_params(
+        "SELECT value FROM app_settings WHERE key = ?",
+        vec![QueryParam::String(STREAMING_QUALITY_KEY.to_string())],
+    );
+
+    let stored: Option<String> = match db_service.query_optional(query, |row| row.get(0)).await {
+        Ok(value) => value,
+        Err(e) => {
+            warn!("[VideoSettings] Failed to read streaming quality: {}", e);
+            return;
+        }
+    };
+
+    let Some(stored) = stored else { return };
+    let quality: StreamingQuality = match serde_json::from_str(&stored) {
+        Ok(quality) => quality,
+        Err(e) => {
+            warn!(
+                "[VideoSettings] Ignoring unrecognised persisted streaming quality {:?}: {}",
+                stored, e
+            );
+            return;
+        }
+    };
+
+    crate::repository::online::set_streaming_quality(quality);
+    if let Some(video_settings) = app.try_state::<VideoSettingsWrapper>() {
+        video_settings.0.lock_safe().streaming_quality = quality;
+    }
+    info!(
+        "[VideoSettings] Restored streaming quality cap: {}",
+        quality.label()
+    );
+}
+
 #[tauri::command]
 #[specta::specta]
 pub async fn player_get_video_settings(

src-tauri/src/commands/player/timers.rs

@@ -141,6 +141,8 @@ pub async fn player_play_next_episode(
 /// - Frontend when HTML5 video ends (Linux/desktop) - passes itemId + repositoryHandle for the video
 /// - Frontend when audio track ends via backend event - no itemId/repositoryHandle needed
 /// - Android JNI callback also triggers this logic directly
+///
+/// TRACES: UR-023, UR-026, UR-040 | DR-047, DR-052, DR-129
 #[tauri::command]
 #[specta::specta]
 pub async fn player_on_playback_ended(
@@ -242,12 +244,34 @@ pub async fn player_on_playback_ended(
                 });
             }
 
-            // Start countdown if auto_advance enabled
+            // Advance if auto_advance is enabled. This is the path that actually
+            // runs on Android: the JNI callback's own decision is swallowed by the
+            // NewTrackLoaded end reason set at load, so it returns Stop, emits
+            // PlaybackEnded, and the frontend echoes it back into this command —
+            // which is where the real decision lands.
             if auto_advance {
                 controller_arc
                     .lock()
                     .await
-                    .start_autoplay_countdown(next_episode, countdown_seconds);
+                    .auto_advance_to_next_episode(next_episode, countdown_seconds)
+                    .await;
+            }
+        }
+        AutoplayDecision::ResumeStream { position } => {
+            // The stream was cut short by the network, not by the media ending.
+            // Re-open it where it died — no queue clearing, no PlaybackEnded, and
+            // above all no leaving the player parked in ExoPlayer's STATE_ENDED,
+            // where the next play intent restarts the item from 0:00.
+            log::info!(
+                "[Autoplay] Decision: Resume truncated stream at {:.1}s",
+                position
+            );
+            let controller = controller_arc.lock().await;
+            if let Err(e) = controller.resume_stream_at(position).await {
+                log::error!("[Autoplay] Failed to resume truncated stream: {}", e);
+                if let Some(emitter) = controller.event_emitter() {
+                    emitter.emit(PlayerStatusEvent::PlaybackEnded);
+                }
             }
         }
     }
@@ -255,6 +279,53 @@ pub async fn player_on_playback_ended(
     Ok(())
 }
 
+/// Try to recover playback after a **recoverable** player error, reporting
+/// whether it was handled.
+///
+/// The frontend's error handler stops the player, which is right for a real
+/// failure and wrong for a network blip — it turned every hiccup into "playback
+/// died". This is the echo path for backends that cannot decide in-process:
+/// MpvBackend is constructed before `PlayerController` exists ([`lib.rs`]), so
+/// its event thread has no controller to ask. It emits the error, the frontend
+/// echoes it here, and the decision stays in Rust — the same shape as
+/// `PlaybackEnded` → `player_on_playback_ended`.
+///
+/// Returns `true` when the stream was re-opened and the caller must NOT stop the
+/// player; `false` when the error is real and should be surfaced as before.
+/// Android decides inside its JNI callback and only emits errors it has already
+/// declined to recover, so this reports `false` for those without a second
+/// opinion — the shared attempt budget is spent by then either way.
+///
+/// TRACES: UR-004, UR-040 | DR-130 | UT-117
+#[tauri::command]
+#[specta::specta]
+pub async fn player_recover_stream(player: State<'_, PlayerStateWrapper>) -> Result<bool, String> {
+    let (position, delay_secs) = {
+        let controller = player.0.lock().await;
+        match controller.recoverable_error_resume() {
+            Some(resume) => resume,
+            None => return Ok(false),
+        }
+    };
+
+    log::warn!(
+        "[Recovery] Stream failed — re-opening at {:.1}s in {}s",
+        position,
+        delay_secs
+    );
+    // Give a brief outage time to clear; retrying instantly just burns the budget.
+    tokio::time::sleep(std::time::Duration::from_secs(delay_secs)).await;
+
+    let controller = player.0.lock().await;
+    match controller.resume_stream_at(position).await {
+        Ok(()) => Ok(true),
+        Err(e) => {
+            log::error!("[Recovery] Failed to re-open stream: {}", e);
+            Ok(false)
+        }
+    }
+}
+
 // ===== HTML5 video state-report commands =====
 //
 // On platforms where video renders in the webview (Linux WebKitGTK HTML5

src-tauri/src/commands/repository.rs

@@ -12,9 +12,11 @@ use serde::{Deserialize, Serialize};
 use tauri::{AppHandle, Emitter, State};
 use uuid::Uuid;
 
+use crate::domain::rank_search_results;
 use crate::jellyfin::HttpClient;
 use crate::repository::{
-    types::*, HybridRepository, MediaRepository, OfflineRepository, OnlineRepository,
+    series_progress, types::*, HybridRepository, MediaRepository, OfflineRepository,
+    OnlineRepository,
 };
 
 /// Repository handle manager
@@ -39,6 +41,19 @@ impl RepositoryManager {
         repos.get(handle).cloned()
     }
 
+    /// Handles of every live repository.
+    ///
+    /// The background catalog indexer (DR-109) runs outside any command, so it
+    /// has no handle passed in and needs to discover one. In practice there is a
+    /// single signed-in repository; returning all of them avoids inventing an
+    /// "active" concept the rest of the code does not have.
+    ///
+    /// TRACES: UR-065 | DR-109
+    pub fn handles(&self) -> Vec<String> {
+        let repos = self.repositories.lock_safe();
+        repos.keys().cloned().collect()
+    }
+
     pub fn destroy(&self, handle: &str) {
         let mut repos = self.repositories.lock_safe();
         repos.remove(handle);
@@ -319,6 +334,71 @@ pub async fn repository_get_next_up_episodes(
         .map_err(|e| format!("{:?}", e))
 }
 
+/// Every episode of a series, across all seasons, in series order.
+///
+/// Jellyfin hangs episodes off season folders — except for "flat" series whose
+/// children are episodes directly. Both shapes are provider vocabulary, so the
+/// fan-out and its fallback live in Rust rather than being reimplemented in the
+/// frontend (which is what it used to do).
+///
+/// TRACES: UR-062 | DR-101
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_get_series_episodes(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    series_id: String,
+) -> Result<Vec<MediaItem>, String> {
+    let repo = manager.0.get(&handle).ok_or("Repository not found")?;
+    series_progress::fetch_series_episodes(repo.as_ref(), &series_id)
+        .await
+        .map_err(|e| format!("{:?}", e))
+}
+
+/// The episode a viewer should land on when they open a series.
+///
+/// "Current" is domain policy, not layout: an episode in progress, else the
+/// server's Next Up for the series, else the first unwatched episode, else the
+/// first. The third rung is what makes this work offline, where Next Up is
+/// always empty. Returns `None` only when the series has no episodes at all.
+///
+/// TRACES: UR-062 | DR-101
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_get_series_current_episode(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    series_id: String,
+) -> Result<Option<MediaItem>, String> {
+    let repo = manager.0.get(&handle).ok_or("Repository not found")?;
+    series_progress::resolve_current_episode(repo.as_ref(), &series_id)
+        .await
+        .map_err(|e| format!("{:?}", e))
+}
+
+/// Erase the viewer's watch history for an item.
+///
+/// Clears the played flag and the resume position; on a series or season the
+/// server applies it to everything inside. A series cleared this way is "never
+/// watched" again, so `repository_get_series_current_episode` returns its
+/// premiere. Requires the server — offline this fails rather than diverging
+/// local state the next sync would overwrite.
+///
+/// TRACES: UR-064 | DR-106
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_clear_watch_history(
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    item_id: String,
+) -> Result<(), String> {
+    let repo = manager.0.get(&handle).ok_or("Repository not found")?;
+    repo.as_ref()
+        .clear_watch_history(&item_id)
+        .await
+        .map_err(|e| format!("{:?}", e))
+}
+
 /// Get recently played audio
 #[tauri::command]
 #[specta::specta]
@@ -394,7 +474,12 @@ pub struct SearchUpdateEvent {
     pub result: SearchResult,
 }
 
-/// Search for items
+/// Search for items.
+///
+/// Resolves `SearchOptions::scope` into concrete Jellyfin item types before
+/// dispatching, so scope taxonomy stays in Rust.
+///
+/// TRACES: UR-049, UR-050 | DR-063
 #[tauri::command]
 #[specta::specta]
 pub async fn repository_search(
@@ -407,9 +492,19 @@ pub async fn repository_search(
 ) -> Result<SearchResult, String> {
     let repo = manager.0.get(&handle).ok_or("Repository not found")?;
 
+    // Expand the opaque scope into item types HERE — once, before the cache and
+    // server paths diverge — so both phases filter identically. Doing it later
+    // (or in only one path) makes offline results disagree with online ones.
+    // The frontend sends `scope` and never names a Jellyfin item type for
+    // search; see docs/specs/scoped-search-boundary.md.
+    let options = options.map(|mut o| {
+        o.resolve_scope();
+        o
+    });
+
     // Phase 1: instant local results from the cache (downloaded content) so the
     // UI can render immediately while the server is still being queried.
-    let cache_result = repo
+    let mut cache_result = repo
         .search_cache_only(&query, options.clone())
         .await
         .unwrap_or_else(|e| {
@@ -420,6 +515,12 @@ pub async fn repository_search(
             }
         });
 
+    // Neither backend orders by *where* the query matched, so a mid-word hit
+    // ("Sparks" for "parks") can outrank a prefix hit ("Parks and Recreation").
+    // Both phases are ranked with the same rules so the list does not reshuffle
+    // when the server results land.
+    rank_search_results(&mut cache_result.items, &query);
+
     // Phase 2: query the live server in the background, merge with the cache,
     // and push the union to the frontend via a `search-event`. Tagged with
     // `request_id` so the frontend can discard results from superseded queries.
@@ -428,7 +529,11 @@ pub async fn repository_search(
     tauri::async_runtime::spawn(async move {
         match repo_bg.search_server_only(&query, options).await {
             Ok(server_result) => {
-                let merged = HybridRepository::merge_search_results(cache_for_merge, server_result);
+                let mut merged =
+                    HybridRepository::merge_search_results(cache_for_merge, server_result);
+                // Rank the union, not each half: a server-only prefix match must
+                // be able to outrank a cached mid-word one.
+                rank_search_results(&mut merged.items, &query);
                 let event = SearchUpdateEvent {
                     request_id,
                     result: merged,
@@ -466,7 +571,13 @@ pub async fn repository_get_playback_info(
         .map_err(|e| format!("{:?}", e))
 }
 
-/// Get video stream URL with optional seeking support
+/// Get a video stream URL.
+///
+/// There is no start-position parameter on purpose: the URL is an HLS playlist
+/// covering the whole item, and a position on it makes the server reject every
+/// segment with `400` (DR-181). Callers resume by seeking after load.
+///
+/// TRACES: UR-004 | DR-181 | UT-182
 #[tauri::command]
 #[specta::specta]
 pub async fn repository_get_video_stream_url(
@@ -474,17 +585,11 @@ pub async fn repository_get_video_stream_url(
     handle: String,
     item_id: String,
     media_source_id: Option<String>,
-    start_time_seconds: Option<f64>,
     audio_stream_index: Option<i32>,
 ) -> Result<String, String> {
     let repo = manager.0.get(&handle).ok_or("Repository not found")?;
     repo.as_ref()
-        .get_video_stream_url(
-            &item_id,
-            media_source_id.as_deref(),
-            start_time_seconds,
-            audio_stream_index,
-        )
+        .get_video_stream_url(&item_id, media_source_id.as_deref(), audio_stream_index)
         .await
         .map_err(|e| format!("{:?}", e))
 }
@@ -607,9 +712,19 @@ pub async fn repository_report_playback_progress(
 }
 
 /// Report playback stopped
+///
+/// A stop-report that cannot reach the server is queued rather than dropped:
+/// this is the position the resume point is built from, and losing it is
+/// exactly the "it forgot where I was" the sync queue exists to prevent. The
+/// drain (DR-131) pushes it on the next reconnect. Queueing is best-effort —
+/// failing the command because the *queue* write failed would tell the caller
+/// the report was lost when the local position was already saved.
+///
+/// TRACES: UR-025 | DR-154 | UT-151
 #[tauri::command]
 #[specta::specta]
 pub async fn repository_report_playback_stopped(
+    db: State<'_, crate::commands::storage::DatabaseWrapper>,
     manager: State<'_, RepositoryManagerWrapper>,
     handle: String,
     item_id: String,
@@ -618,10 +733,39 @@ pub async fn repository_report_playback_stopped(
     // Milliseconds across the boundary; the Jellyfin API wants ticks.
     let position_ticks = position_ms * 10_000;
     let repo = manager.0.get(&handle).ok_or("Repository not found")?;
-    repo.as_ref()
+
+    let result = repo
+        .as_ref()
         .report_playback_stopped(&item_id, position_ticks)
+        .await;
+
+    if let Err(e) = &result {
+        let db_service = {
+            let database = db.0.lock().map_err(|err| err.to_string())?;
+            Arc::new(database.service())
+        };
+        let user_id = repo.user_id().to_string();
+        if let Err(queue_err) = crate::commands::sync_drain::enqueue_playback_stopped(
+            &db_service,
+            &user_id,
+            &item_id,
+            position_ticks,
+        )
         .await
-        .map_err(|e| format!("{:?}", e))
+        {
+            warn!(
+                "[Repository] Stop-report for {} failed ({:?}) and could not be queued: {}",
+                item_id, e, queue_err
+            );
+        } else {
+            debug!(
+                "[Repository] Stop-report for {} failed ({:?}); queued for the next reconnect",
+                item_id, e
+            );
+        }
+    }
+
+    result.map_err(|e| format!("{:?}", e))
 }
 
 /// Get image URL for an item
@@ -660,7 +804,7 @@ pub fn repository_get_subtitle_url(
 #[tauri::command]
 #[specta::specta]
 #[allow(dead_code)]
-pub fn repository_get_video_download_url(
+pub async fn repository_get_video_download_url(
     manager: State<'_, RepositoryManagerWrapper>,
     handle: String,
     item_id: String,
@@ -668,9 +812,16 @@ pub fn repository_get_video_download_url(
     media_source_id: Option<String>,
 ) -> Result<String, String> {
     let repo = manager.0.get(&handle).ok_or("Repository not found")?;
-    Ok(repo
-        .as_ref()
-        .get_video_download_url(&item_id, &quality, media_source_id.as_deref()))
+    // Async because the audio-codec policy has to know what the source's audio
+    // is before it can decide whether the file may be copied verbatim (DR-171).
+    // The frontend calls this exactly as before — the decision stays in Rust.
+    Ok(crate::repository::resolve_video_download_url(
+        repo.as_ref(),
+        &item_id,
+        &quality,
+        media_source_id.as_deref(),
+    )
+    .await)
 }
 
 /// Mark an item as favorite
@@ -688,6 +839,125 @@ pub async fn repository_mark_favorite(
         .map_err(|e| format!("{:?}", e))
 }
 
+/// Tauri event announcing that favourite state changed behind the UI's back —
+/// either because the server disagreed with the cache on a background refresh,
+/// or because pending offline toggles were pushed on reconnect.
+///
+/// TRACES: UR-069 | DR-120
+pub const FAVORITES_CHANGED_EVENT: &str = "favorites-changed";
+
+/// Payload for [`FAVORITES_CHANGED_EVENT`] — the ids whose favourite state
+/// actually flipped, so the frontend refreshes those rather than everything.
+///
+/// TRACES: UR-069 | DR-120 | UT-107
+#[derive(specta::Type, Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct FavoritesChangedEvent {
+    pub item_ids: Vec<String>,
+}
+
+/// Ids whose favourite state differs between what we showed and what the server
+/// has — favourited elsewhere since the cache was written, or un-favourited
+/// elsewhere.
+///
+/// Pulled out of the command so the "emit nothing when nothing changed" rule is
+/// testable: an unchanged set must leave a quiet page quiet rather than
+/// triggering a refetch on every visit.
+///
+/// TRACES: UR-069 | DR-120 | UT-107
+fn changed_favorite_ids(
+    cached: &std::collections::HashSet<String>,
+    server: &std::collections::HashSet<String>,
+) -> Vec<String> {
+    let mut changed: Vec<String> = server.symmetric_difference(cached).cloned().collect();
+    // Deterministic order so the event payload does not depend on hash seeding.
+    changed.sort();
+    changed
+}
+
+/// Everything the viewer has favourited, across libraries, narrowed by scope.
+///
+/// Two-phase like `repository_search`: the local answer returns immediately and
+/// a background server pass emits `favorites-changed` when the server's set
+/// differs. Without the second phase a favourite marked in another client shows
+/// up only on the *second* visit to the page, since the cache-first read hands
+/// back local rows and the refresh is invisible to the frontend.
+///
+/// TRACES: UR-067 | DR-115, DR-120, JA-033 | UT-107
+#[tauri::command]
+#[specta::specta]
+pub async fn repository_get_favorites(
+    app: AppHandle,
+    manager: State<'_, RepositoryManagerWrapper>,
+    handle: String,
+    scope: SearchScope,
+    options: Option<GetItemsOptions>,
+) -> Result<SearchResult, String> {
+    let repo = manager.0.get(&handle).ok_or("Repository not found")?;
+
+    let cache_result = repo
+        .get_favorites_cache_only(scope, options.clone())
+        .await
+        .unwrap_or_else(|e| {
+            debug!("[Favorites] Cache miss/timeout: {:?}", e);
+            SearchResult {
+                items: Vec::new(),
+                total_record_count: 0,
+            }
+        });
+
+    // With "Show all server media" off the local answer is authoritative
+    // (DR-080) — don't go behind the user's back to the server.
+    if !crate::repository::offline::include_catalog_browse() {
+        return Ok(cache_result);
+    }
+
+    // Nothing cached yet — a fresh install, or a viewer whose favourites were
+    // all marked on another client. Returning the empty result here paints
+    // "Nothing favourited yet — tap the heart on anything you like", which is a
+    // *wrong* answer, corrected a server round trip later when the background
+    // refresh fires `favorites-changed`. Ask the repository for a real answer
+    // instead: its `get_favorites` is exactly this read — cache first, server on
+    // a miss, saving through — and it applies the same DR-080 gate.
+    //
+    // TRACES: UR-067 | DR-115
+    if !cache_result.has_content() {
+        debug!("[Favorites] Nothing cached; answering from the server");
+        return repo
+            .get_favorites(scope, options)
+            .await
+            .map_err(|e| format!("{:?}", e));
+    }
+
+    let repo_bg = repo.clone();
+    let cached_ids: std::collections::HashSet<String> =
+        cache_result.items.iter().map(|i| i.id.clone()).collect();
+    tauri::async_runtime::spawn(async move {
+        match repo_bg.get_favorites_server_only(scope, options).await {
+            Ok(server_result) => {
+                let server_ids: std::collections::HashSet<String> =
+                    server_result.items.iter().map(|i| i.id.clone()).collect();
+                let changed = changed_favorite_ids(&cached_ids, &server_ids);
+
+                if !changed.is_empty() {
+                    let event = FavoritesChangedEvent { item_ids: changed };
+                    if let Err(e) = app.emit(FAVORITES_CHANGED_EVENT, &event) {
+                        error!("[Favorites] Failed to emit change event: {}", e);
+                    }
+                }
+            }
+            Err(e) => {
+                warn!(
+                    "[Favorites] Server refresh failed, keeping cached favourites: {:?}",
+                    e
+                );
+            }
+        }
+    });
+
+    Ok(cache_result)
+}
+
 /// Unmark an item as favorite
 #[tauri::command]
 #[specta::specta]
@@ -761,6 +1031,44 @@ mod tests {
         assert!(manager.get("any-handle").is_none());
     }
 
+    fn ids(values: &[&str]) -> std::collections::HashSet<String> {
+        values.iter().map(|v| v.to_string()).collect()
+    }
+
+    /// UT-107 — the background refresh reports only what actually changed.
+    ///
+    /// TRACES: UR-069 | DR-120 | UT-107
+    #[test]
+    fn test_changed_favorite_ids_reports_both_directions() {
+        // Favourited in another client since we cached.
+        assert_eq!(
+            changed_favorite_ids(&ids(&["a"]), &ids(&["a", "b"])),
+            vec!["b".to_string()]
+        );
+
+        // Un-favourited in another client.
+        assert_eq!(
+            changed_favorite_ids(&ids(&["a", "b"]), &ids(&["a"])),
+            vec!["b".to_string()]
+        );
+
+        // Both at once, in a stable order.
+        assert_eq!(
+            changed_favorite_ids(&ids(&["a", "b"]), &ids(&["b", "c"])),
+            vec!["a".to_string(), "c".to_string()]
+        );
+    }
+
+    /// An unchanged set emits nothing — otherwise every visit to the page would
+    /// fire an event and trigger a pointless refetch.
+    ///
+    /// TRACES: UR-069 | DR-120 | UT-107
+    #[test]
+    fn test_changed_favorite_ids_is_empty_when_nothing_moved() {
+        assert!(changed_favorite_ids(&ids(&["a", "b"]), &ids(&["b", "a"])).is_empty());
+        assert!(changed_favorite_ids(&ids(&[]), &ids(&[])).is_empty());
+    }
+
     #[test]
     fn test_repository_manager_wrapper_structure() {
         let manager = RepositoryManager::new();

src-tauri/src/commands/storage/mod.rs

@@ -80,6 +80,30 @@ pub fn storage_init(db: State<DatabaseWrapper>) -> Result<String, String> {
     Ok(database.path().to_string_lossy().to_string())
 }
 
+/// A playable URL for a downloaded file on disk.
+///
+/// Local media is served over a loopback HTTP server rather than handed to the
+/// webview as a `file://`/asset URL, because the asset protocol cannot stream a
+/// large file — it answers a range-less request with the whole thing, which
+/// Chromium abandons. See `media_server` for why real HTTP is used.
+///
+/// The returned URL carries the server's per-session token, so it is only valid
+/// for this run of the app and must not be persisted.
+///
+/// TRACES: UR-071 | DR-137
+#[tauri::command]
+#[specta::specta]
+pub fn media_local_url(
+    server: State<crate::media_server::MediaServerWrapper>,
+    path: String,
+) -> Result<String, String> {
+    server
+        .0
+        .as_ref()
+        .map(|s| s.url_for(&path))
+        .ok_or_else(|| "Local media server is not running".to_string())
+}
+
 /// Get storage directory path (parent directory of the database file)
 #[tauri::command]
 #[specta::specta]
@@ -843,6 +867,86 @@ pub async fn storage_mark_played(
     }
 }
 
+/// Set the watched flag locally for an item **and everything inside it**.
+///
+/// This backs the watched toggle, and is deliberately separate from
+/// [`storage_mark_played`] — which reports a single track/episode finishing and
+/// increments `play_count` — because the toggle has two directions and applies
+/// to containers.
+///
+/// The recursion is what makes the toggle honest offline. Jellyfin applies
+/// `POST`/`DELETE /PlayedItems/{id}` recursively over a season or series, so
+/// online the server fixes up the children on the next read; with no server to
+/// ask, marking a season watched would otherwise tick the season and leave every
+/// episode inside it unwatched. Targets are drawn from `items` by the same link
+/// columns the rest of the offline layer uses, so an id that is not cached
+/// selects nothing and the statement is a no-op rather than a foreign-key error.
+///
+/// Un-marking clears the resume position too, matching the server, so an item
+/// un-marked offline does not come back offering to resume from a position it is
+/// no longer meant to have.
+///
+/// `pending_sync = 1` hands the rows to the sync drain.
+///
+/// TRACES: UR-073 | DR-158
+#[tauri::command]
+#[specta::specta]
+pub async fn storage_set_watched(
+    db: State<'_, DatabaseWrapper>,
+    user_id: String,
+    item_id: String,
+    watched: bool,
+) -> Result<(), String> {
+    let db_service = {
+        let database = db.0.lock().map_err(|e| e.to_string())?;
+        Arc::new(database.service())
+    };
+
+    // The item itself plus its descendants: a season's episodes reach it by
+    // season_id, a series' by series_id, its seasons by parent_id, an album's
+    // tracks by album_id.
+    let targets = "SELECT id FROM items
+                   WHERE id = ? OR parent_id = ? OR album_id = ?
+                      OR season_id = ? OR series_id = ?";
+
+    let sql = if watched {
+        format!(
+            "INSERT INTO user_data (user_id, item_id, is_played, play_count, last_played_at, pending_sync)
+             SELECT ?, id, 1, 1, CURRENT_TIMESTAMP, 1 FROM ({targets})
+             ON CONFLICT(user_id, item_id) DO UPDATE SET
+                is_played = 1,
+                play_count = MAX(user_data.play_count, 1),
+                last_played_at = CURRENT_TIMESTAMP,
+                pending_sync = 1"
+        )
+    } else {
+        format!(
+            "INSERT INTO user_data (user_id, item_id, is_played, play_count, playback_position_ticks, pending_sync)
+             SELECT ?, id, 0, 0, 0, 1 FROM ({targets})
+             ON CONFLICT(user_id, item_id) DO UPDATE SET
+                is_played = 0,
+                play_count = 0,
+                playback_position_ticks = 0,
+                pending_sync = 1"
+        )
+    };
+
+    let query = Query::with_params(
+        sql,
+        vec![
+            QueryParam::String(user_id),
+            QueryParam::String(item_id.clone()),
+            QueryParam::String(item_id.clone()),
+            QueryParam::String(item_id.clone()),
+            QueryParam::String(item_id.clone()),
+            QueryParam::String(item_id.clone()),
+        ],
+    );
+
+    db_service.execute(query).await.map_err(|e| e.to_string())?;
+    Ok(())
+}
+
 /// Get playback progress for an item
 #[tauri::command]
 #[specta::specta]

src-tauri/src/commands/storage/people.rs

@@ -47,10 +47,25 @@ pub async fn storage_save_person(
     };
 
     let query = Query::with_params(
-        "INSERT OR REPLACE INTO people (
+        // A real UPSERT, not INSERT OR REPLACE — `people` is now backed by the
+        // `people_fts` index (migration 022), and REPLACE would orphan an index
+        // entry on every re-cache: it fires no AFTER DELETE trigger without
+        // `recursive_triggers`, and reassigns the rowid that `content_rowid`
+        // refers to. Same defect as DR-110 fixed for `items`.
+        //
+        // TRACES: UR-065 | DR-110, DR-111
+        "INSERT INTO people (
             id, server_id, name, overview, primary_image_tag,
             premiere_date, end_date, synced_at
-        ) VALUES (?, ?, ?, ?, ?, ?, ?, CURRENT_TIMESTAMP)",
+        ) VALUES (?, ?, ?, ?, ?, ?, ?, CURRENT_TIMESTAMP)
+        ON CONFLICT(id) DO UPDATE SET
+            server_id = excluded.server_id,
+            name = excluded.name,
+            overview = excluded.overview,
+            primary_image_tag = excluded.primary_image_tag,
+            premiere_date = excluded.premiere_date,
+            end_date = excluded.end_date,
+            synced_at = CURRENT_TIMESTAMP",
         vec![
             QueryParam::String(person.id),
             QueryParam::String(person.server_id),

src-tauri/src/commands/sync.rs

@@ -2,7 +2,9 @@
 //!
 //! The sync queue stores mutations (favorites, playback progress, etc.)
 //! that need to be synced to the Jellyfin server when connectivity is restored.
-//! TRACES: UR-002, UR-017, UR-025 | DR-014
+//! Draining it lives in `sync_drain` (DR-131); this module is the storage and
+//! read side the UI lists from (DR-132).
+//! TRACES: UR-002, UR-017, UR-025 | DR-014, DR-131, DR-132
 
 use serde::{Deserialize, Serialize};
 use std::sync::Arc;
@@ -24,6 +26,12 @@ pub struct SyncQueueItem {
     pub retry_count: i32,
     pub created_at: Option<String>,
     pub error_message: Option<String>,
+    /// Cached title of the item the operation is about, when the catalog knows
+    /// it. Resolved here rather than by a per-row frontend fetch — the queue
+    /// list is otherwise a wall of opaque ids.
+    ///
+    /// TRACES: UR-025 | DR-132
+    pub item_name: Option<String>,
 }
 
 /// Queue a mutation for sync to server
@@ -74,20 +82,20 @@ pub async fn sync_get_pending(
         Arc::new(database.service())
     };
 
-    let sql = if let Some(l) = limit {
-        format!(
-            "SELECT id, user_id, operation, item_id, payload, status, retry_count, created_at, error_message
-             FROM sync_queue
-             WHERE user_id = ? AND status IN ('pending', 'failed')
-             ORDER BY created_at ASC
-             LIMIT {}",
-            l
-        )
-    } else {
-        "SELECT id, user_id, operation, item_id, payload, status, retry_count, created_at, error_message
-         FROM sync_queue
-         WHERE user_id = ? AND status IN ('pending', 'failed')
-         ORDER BY created_at ASC".to_string()
+    // The `items` join names the queued item where the catalog has it; a row for
+    // an item that was never cached still lists, with a null name.
+    // `abandoned` rows (DR-131 gave up on them) are excluded here for the same
+    // reason they are excluded from the count — they are no longer waiting.
+    const SELECT: &str = "SELECT q.id, q.user_id, q.operation, q.item_id, q.payload, q.status,
+                                 COALESCE(q.retry_count, 0), q.created_at, q.error_message, i.name
+                          FROM sync_queue q
+                          LEFT JOIN items i ON i.id = q.item_id
+                          WHERE q.user_id = ? AND q.status IN ('pending', 'failed')
+                          ORDER BY q.created_at ASC, q.id ASC";
+
+    let sql = match limit {
+        Some(l) => format!("{} LIMIT {}", SELECT, l),
+        None => SELECT.to_string(),
     };
 
     let query = Query::with_params(sql, vec![QueryParam::String(user_id)]);
@@ -104,6 +112,7 @@ pub async fn sync_get_pending(
                 retry_count: row.get(6)?,
                 created_at: row.get(7)?,
                 error_message: row.get(8)?,
+                item_name: row.get(9)?,
             })
         })
         .await
@@ -256,6 +265,7 @@ mod tests {
             retry_count: 0,
             created_at: Some("2024-02-14T08:00:00Z".to_string()),
             error_message: None,
+            item_name: None,
         };
 
         // Should serialize successfully
@@ -280,6 +290,7 @@ mod tests {
             retry_count: 3,
             created_at: Some("2024-02-14T07:00:00Z".to_string()),
             error_message: Some("Connection timeout".to_string()),
+            item_name: None,
         };
 
         let json = serde_json::to_string(&item).unwrap();
@@ -300,6 +311,7 @@ mod tests {
             retry_count: 0,
             created_at: None,
             error_message: None,
+            item_name: None,
         };
 
         let json = serde_json::to_string(&item).unwrap();
@@ -323,6 +335,7 @@ mod tests {
                 retry_count: 0,
                 created_at: None,
                 error_message: None,
+                item_name: None,
             };
 
             let json = serde_json::to_string(&item).unwrap();
@@ -363,6 +376,7 @@ mod tests {
             retry_count: 0,
             created_at: None,
             error_message: None,
+            item_name: None,
         };
 
         // Simulate retries

src-tauri/src/domain/mod.rs

@@ -5,6 +5,8 @@
 
 pub mod from_jellyfin;
 pub mod media;
+pub mod search_rank;
 
 pub use from_jellyfin::{kind_from_jellyfin, stream_kind_from_jellyfin, ticks_to_ms};
 pub use media::{MediaKind, StreamKind};
+pub use search_rank::rank_search_results;

src-tauri/src/domain/search_rank.rs

@@ -0,0 +1,313 @@
+//! Relevance ranking for search results.
+//!
+//! Both search paths (the SQLite FTS cache and the Jellyfin server) return items
+//! in an order that ignores *where* in the name the query matched: a server
+//! substring hit like "Sparks of Love" can outrank "Parks and Recreation" for
+//! the query "parks". Neither backend is going to change, so the app imposes its
+//! own ordering on the union.
+//!
+//! Ranking is domain logic, not presentation: it encodes what a "better match"
+//! means and which media kinds outrank which. The frontend only renders the
+//! order it is given.
+//!
+//! Two rules, in priority order:
+//!
+//! 1. **Match position** — a prefix match beats a word-start match, which beats
+//!    a mid-word substring match. This is what makes "parks" find
+//!    "Parks and Recreation" before "Sparks of Love".
+//! 2. **Kind** — containers before their contents at equal match quality, so a
+//!    series outranks its own episodes.
+//!
+//! Ties fall back to the input order, so a backend's own relevance signal (FTS
+//! `rank`) still breaks ties it was never overruled on.
+
+use crate::domain::MediaKind;
+use crate::repository::types::MediaItem;
+
+/// How well a query matched an item's name — better matches sort first.
+///
+/// Ordered by discriminant: `Prefix` is the strongest. Derived `Ord` gives the
+/// comparison for free, so adding a tier in the right position is all it takes.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+pub enum MatchQuality {
+    /// The name starts with the query — "parks" in "Parks and Recreation".
+    Prefix,
+    /// Some later *word* starts with the query — "recreation" in "Parks and
+    /// Recreation". Still a deliberate hit: users type whole words.
+    WordStart,
+    /// The query appears mid-word — "parks" in "Sparks of Love". Weakest hit
+    /// that still counts as a match.
+    Substring,
+    /// No match on the name at all. The backend returned it for some other
+    /// reason (overview, artist, album), so it is kept but sorted last.
+    None,
+}
+
+/// Rank of a media kind when match quality ties — lower sorts first.
+///
+/// Containers outrank the items they contain: searching a show's name should
+/// surface the show, not an arbitrary episode of it. Within a tier the order is
+/// arbitrary but stable, and equal ranks fall through to input order.
+fn kind_rank(kind: MediaKind) -> u8 {
+    match kind {
+        // Top-level containers a user is most likely to be looking for.
+        MediaKind::Series | MediaKind::Movie | MediaKind::Album | MediaKind::Artist => 0,
+        // Sub-containers and standalone collections.
+        MediaKind::Season | MediaKind::Playlist | MediaKind::Channel | MediaKind::Folder => 1,
+        // Leaves — an episode/track is a match *inside* something bigger.
+        MediaKind::Episode | MediaKind::Track | MediaKind::LiveChannel | MediaKind::ChannelItem => {
+            2
+        }
+        // Peripheral matches.
+        MediaKind::Person | MediaKind::Other => 3,
+    }
+}
+
+/// Classify how `query` matches `name`, case-insensitively.
+///
+/// Both sides are trimmed and lowercased; an empty query matches everything
+/// equally (`Prefix`), which leaves the input order untouched.
+pub fn match_quality(name: &str, query: &str) -> MatchQuality {
+    let query = query.trim().to_lowercase();
+    if query.is_empty() {
+        return MatchQuality::Prefix;
+    }
+    let name = name.trim().to_lowercase();
+
+    let Some(index) = name.find(&query) else {
+        return MatchQuality::None;
+    };
+
+    if index == 0 {
+        return MatchQuality::Prefix;
+    }
+
+    // A word start is any match preceded by a non-alphanumeric character, so
+    // "the-office" and "The Office" behave the same. Indexing back one char is
+    // safe on the byte index `find` returned only via `char_indices`, since a
+    // multi-byte char would panic on a raw slice.
+    let preceded_by_boundary = name[..index]
+        .chars()
+        .next_back()
+        .is_some_and(|c| !c.is_alphanumeric());
+
+    if preceded_by_boundary {
+        MatchQuality::WordStart
+    } else {
+        MatchQuality::Substring
+    }
+}
+
+/// Sort search results by relevance to `query`, in place.
+///
+/// Stable, so items the rules rank equally keep the order the backend supplied
+/// (FTS `rank` for cache hits, Jellyfin's own ordering for server hits).
+///
+/// TRACES: UR-060 | DR-090
+pub fn rank_search_results(items: &mut [MediaItem], query: &str) {
+    // An empty query carries no relevance signal, so there is nothing to rank
+    // by — reordering on kind alone would shuffle the backend's own ordering
+    // for no reason.
+    if query.trim().is_empty() {
+        return;
+    }
+    items.sort_by_key(|item| (match_quality(&item.name, query), kind_rank(item.kind)));
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn item(name: &str, kind: MediaKind) -> MediaItem {
+        let mut item = MediaItem::default();
+        item.id = format!("id-{}-{:?}", name, kind);
+        item.name = name.to_string();
+        item.kind = kind;
+        item
+    }
+
+    fn names(items: &[MediaItem]) -> Vec<&str> {
+        items.iter().map(|i| i.name.as_str()).collect()
+    }
+
+    /// UT-085: a prefix match outranks a mid-word substring match.
+    #[test]
+    fn prefix_match_beats_midword_substring() {
+        assert_eq!(
+            match_quality("Parks and Recreation", "parks"),
+            MatchQuality::Prefix
+        );
+        assert_eq!(
+            match_quality("Sparks of Love", "parks"),
+            MatchQuality::Substring
+        );
+        assert!(MatchQuality::Prefix < MatchQuality::Substring);
+    }
+
+    /// UT-085: the reported bug — "parks" must find the show, not "Sparks".
+    #[test]
+    fn ranks_prefix_match_before_substring_match() {
+        let mut items = vec![
+            item("Sparks of Love", MediaKind::Series),
+            item("Parks and Recreation", MediaKind::Series),
+        ];
+
+        rank_search_results(&mut items, "parks");
+
+        assert_eq!(
+            names(&items),
+            vec!["Parks and Recreation", "Sparks of Love"]
+        );
+    }
+
+    /// A match at a later word start beats a mid-word one but loses to a prefix.
+    #[test]
+    fn word_start_ranks_between_prefix_and_substring() {
+        assert_eq!(
+            match_quality("The Office", "office"),
+            MatchQuality::WordStart
+        );
+        assert_eq!(match_quality("Bofficer", "office"), MatchQuality::Substring);
+
+        let mut items = vec![
+            item("Bofficer", MediaKind::Series),
+            item("The Office", MediaKind::Series),
+            item("Office Space", MediaKind::Movie),
+        ];
+
+        rank_search_results(&mut items, "office");
+
+        assert_eq!(
+            names(&items),
+            vec!["Office Space", "The Office", "Bofficer"]
+        );
+    }
+
+    /// UT-086: at equal match quality a series outranks an episode.
+    #[test]
+    fn series_ranks_before_episode_at_equal_match_quality() {
+        let mut items = vec![
+            item("Parks and Recreation S01E01", MediaKind::Episode),
+            item("Parks and Recreation", MediaKind::Series),
+        ];
+
+        rank_search_results(&mut items, "parks");
+
+        assert_eq!(
+            names(&items),
+            vec!["Parks and Recreation", "Parks and Recreation S01E01"]
+        );
+    }
+
+    /// Albums outrank their tracks for the same reason series outrank episodes.
+    #[test]
+    fn album_ranks_before_track_at_equal_match_quality() {
+        let mut items = vec![
+            item("Rumours", MediaKind::Track),
+            item("Rumours", MediaKind::Album),
+        ];
+
+        rank_search_results(&mut items, "rumours");
+
+        assert_eq!(items[0].kind, MediaKind::Album);
+    }
+
+    /// Match quality dominates kind: a better-matching episode beats a
+    /// worse-matching series, so kind never drags an irrelevant show to the top.
+    #[test]
+    fn match_quality_outranks_kind() {
+        let mut items = vec![
+            item("Sparks of Love", MediaKind::Series),
+            item("Parks Cleanup", MediaKind::Episode),
+        ];
+
+        rank_search_results(&mut items, "parks");
+
+        assert_eq!(names(&items), vec!["Parks Cleanup", "Sparks of Love"]);
+    }
+
+    /// Items the backend returned for a non-name reason (overview, artist) are
+    /// kept, but sort below everything that actually matched the name.
+    #[test]
+    fn non_matching_names_sort_last_without_being_dropped() {
+        let mut items = vec![
+            item("Unrelated Documentary", MediaKind::Movie),
+            item("Parks and Recreation", MediaKind::Series),
+        ];
+
+        rank_search_results(&mut items, "parks");
+
+        assert_eq!(
+            names(&items),
+            vec!["Parks and Recreation", "Unrelated Documentary"]
+        );
+    }
+
+    /// Ranking is stable: equally-ranked items keep the backend's order, so the
+    /// FTS/server relevance signal still breaks ties.
+    #[test]
+    fn equal_rank_preserves_input_order() {
+        let mut items = vec![
+            item("Parks A", MediaKind::Series),
+            item("Parks B", MediaKind::Series),
+            item("Parks C", MediaKind::Series),
+        ];
+
+        rank_search_results(&mut items, "parks");
+
+        assert_eq!(names(&items), vec!["Parks A", "Parks B", "Parks C"]);
+    }
+
+    /// Case and surrounding whitespace never change the tier.
+    #[test]
+    fn matching_is_case_and_whitespace_insensitive() {
+        assert_eq!(
+            match_quality("PARKS AND RECREATION", "  parks "),
+            MatchQuality::Prefix
+        );
+        assert_eq!(
+            match_quality("Parks and Recreation", "PARKS"),
+            MatchQuality::Prefix
+        );
+    }
+
+    /// An empty query leaves the order alone rather than reshuffling on kind.
+    #[test]
+    fn empty_query_preserves_input_order() {
+        let mut items = vec![
+            item("Zebra", MediaKind::Episode),
+            item("Apple", MediaKind::Series),
+        ];
+
+        rank_search_results(&mut items, "");
+
+        assert_eq!(names(&items), vec!["Zebra", "Apple"]);
+    }
+
+    /// A multi-byte name must not panic when the match is mid-string — the
+    /// boundary check walks chars rather than slicing raw bytes.
+    #[test]
+    fn handles_multibyte_names_without_panicking() {
+        assert_eq!(
+            match_quality("Pokémon Journeys", "journeys"),
+            MatchQuality::WordStart
+        );
+        assert_eq!(
+            match_quality("Café Parks", "parks"),
+            MatchQuality::WordStart
+        );
+    }
+
+    /// Punctuation counts as a word boundary, so "office" hits "The-Office".
+    #[test]
+    fn punctuation_counts_as_a_word_boundary() {
+        assert_eq!(
+            match_quality("The-Office", "office"),
+            MatchQuality::WordStart
+        );
+        assert_eq!(
+            match_quality("Show: Parks", "parks"),
+            MatchQuality::WordStart
+        );
+    }
+}

src-tauri/src/download/cache.rs

@@ -26,6 +26,12 @@ pub struct CacheConfig {
     pub storage_limit: u64,
     /// Only cache on WiFi
     pub wifi_only: bool,
+    /// How long a temporary (`download_source = 'auto'`) download lives before
+    /// it is reclaimed, in hours. 0 disables expiry, leaving space pressure as
+    /// the only reclaim trigger.
+    ///
+    /// TRACES: UR-071 | DR-127
+    pub temporary_ttl_hours: u64,
 }
 
 impl Default for CacheConfig {
@@ -37,6 +43,10 @@ impl Default for CacheConfig {
             album_affinity_threshold: 3,
             storage_limit: 10 * 1024 * 1024 * 1024, // 10GB
             wifi_only: false,                       // Allow preloading on any connection by default
+            // A week: long enough that re-watching over a weekend still hits
+            // disk, short enough that a one-off play does not hold space
+            // indefinitely.
+            temporary_ttl_hours: 24 * 7,
         }
     }
 }
@@ -183,7 +193,92 @@ impl SmartCache {
         current_size + new_size <= storage_limit
     }
 
-    /// Evict least recently used items to make space (async version)
+    /// Reclaim temporary downloads whose life limit has passed.
+    ///
+    /// The time-based half of the temporary tier (DR-127); [`evict_lru_async`]
+    /// is the space-pressure half. A row is reclaimed by whichever fires first.
+    ///
+    /// Scoped to `download_source = 'auto'` for the same reason eviction is: a
+    /// `'user'` row is someone's own download and has no expiry. `COALESCE`
+    /// guards rows predating migration 012, whose source is NULL and whose
+    /// provenance must therefore be treated as the user's.
+    ///
+    /// Expiry is normally *derived* — `completed_at` plus the configured TTL —
+    /// rather than stamped at completion. That means a TTL change applies to
+    /// entries already on disk instead of only to future ones, and entries
+    /// predating the column expire without a backfill. `expires_at` is honoured
+    /// as a per-row override when something sets it.
+    ///
+    /// `now` is passed in rather than read from the clock so the policy is
+    /// testable without sleeping. Both sides go through SQLite's `datetime()`
+    /// because `completed_at` is written as `CURRENT_TIMESTAMP`
+    /// (`YYYY-MM-DD HH:MM:SS`) while callers pass RFC-3339 (`…T…+00:00`) — a raw
+    /// string comparison between the two formats is wrong, since `' ' < 'T'`.
+    ///
+    /// Returns the number of entries reclaimed.
+    ///
+    /// TRACES: UR-071 | DR-127 | UT-120
+    pub async fn reclaim_expired_async<S: DatabaseService>(
+        &self,
+        db_service: &Arc<S>,
+        user_id: &str,
+        now: &str,
+    ) -> Result<usize, String> {
+        let ttl_hours = {
+            let config = self.config.lock().map_err(|e| e.to_string())?;
+            config.temporary_ttl_hours
+        };
+        // 0 disables time-based reclaim; space pressure remains the only trigger.
+        if ttl_hours == 0 {
+            return Ok(0);
+        }
+
+        let expired: Vec<(i64, String)> = db_service
+            .query_many(
+                Query::with_params(
+                    "SELECT id, file_path FROM downloads
+                      WHERE user_id = ?
+                        AND COALESCE(download_source, 'user') = 'auto'
+                        AND status = 'completed'
+                        AND datetime(
+                              COALESCE(expires_at, datetime(completed_at, '+' || ? || ' hours'))
+                            ) < datetime(?)",
+                    vec![
+                        QueryParam::String(user_id.to_string()),
+                        QueryParam::String(ttl_hours.to_string()),
+                        QueryParam::String(now.to_string()),
+                    ],
+                ),
+                |row| Ok((row.get(0)?, row.get(1)?)),
+            )
+            .await
+            .map_err(|e| e.to_string())?;
+
+        let mut reclaimed = 0usize;
+        for (id, file_path) in expired {
+            // Best-effort on the file: a missing one still needs its row gone,
+            // or the sweep retries it forever.
+            let _ = std::fs::remove_file(&file_path);
+            db_service
+                .execute(Query::with_params(
+                    "DELETE FROM downloads WHERE id = ?",
+                    vec![QueryParam::Int64(id)],
+                ))
+                .await
+                .map_err(|e| e.to_string())?;
+            reclaimed += 1;
+        }
+
+        if reclaimed > 0 {
+            info!("[SmartCache] Reclaimed {} expired entries", reclaimed);
+        }
+        Ok(reclaimed)
+    }
+
+    /// Evict least recently used items to make space (async version).
+    ///
+    /// The space-pressure half of the temporary tier; [`reclaim_expired_async`]
+    /// is the time-based half.
     pub async fn evict_lru_async<S: DatabaseService>(
         &self,
         db_service: &Arc<S>,
@@ -207,10 +302,26 @@ impl SmartCache {
         let to_free = (current_size + space_needed) - limit;
         let mut freed: u64 = 0;
 
-        // Get downloads ordered by last access (oldest first)
+        // Only the *temporary* tier is evictable. `download_source = 'auto'` is
+        // precache — the cache put it there, the cache may reclaim it. A 'user'
+        // row is a download someone explicitly asked for; deleting it to make
+        // room for a predictive fetch is data loss, and because the old query
+        // ordered purely by `completed_at ASC` it took the oldest — typically
+        // exactly the film saved for a flight.
+        //
+        // COALESCE, not `= 'auto'` alone: migration 012 added the column with a
+        // 'user' default, but rows predating it can be NULL, and an unknown
+        // provenance must be treated as the user's, never as disposable.
+        //
+        // Freeing less than requested is the correct outcome when only user
+        // downloads remain — the caller surfaces "unable to free enough space"
+        // rather than silently deleting them.
+        //
+        // TRACES: UR-071 | DR-126 | UT-108
         let query = Query::with_params(
             "SELECT id, file_size, file_path FROM downloads
              WHERE user_id = ? AND status = 'completed'
+               AND COALESCE(download_source, 'user') = 'auto'
              ORDER BY completed_at ASC",
             vec![QueryParam::String(user_id.to_string())],
         );
@@ -304,6 +415,205 @@ mod tests {
         assert!(cache.should_precache_queue());
     }
 
+    /// Expiry reclaims only temporary entries that are actually past their life
+    /// limit — never a user's download (which has no expiry), and never a
+    /// temporary entry still within its life.
+    ///
+    /// TRACES: UR-071 | DR-127 | UT-120
+    #[tokio::test]
+    async fn test_reclaim_expired_only_takes_expired_temporary_entries() {
+        use crate::storage::db_service::{DatabaseService, RusqliteService};
+        use rusqlite::Connection;
+        use std::sync::{Arc, Mutex};
+
+        let conn = Connection::open_in_memory().unwrap();
+        conn.execute(
+            "CREATE TABLE downloads (
+                id INTEGER PRIMARY KEY,
+                user_id TEXT,
+                status TEXT,
+                file_size INTEGER,
+                file_path TEXT,
+                completed_at TEXT,
+                download_source TEXT DEFAULT 'user',
+                expires_at TEXT
+            )",
+            [],
+        )
+        .unwrap();
+
+        // `completed_at` is in SQLite's CURRENT_TIMESTAMP format (space, not
+        // 'T'), deliberately: the query has to compare it against an RFC-3339
+        // "now" and must not do so as raw strings.
+        for (path, source, completed, expires) in [
+            // Completed long ago, no override => derived expiry has passed.
+            ("/tmp/jt-expired.mp4", "auto", "2026-01-01 00:00:00", None),
+            // Completed yesterday => still inside the 7-day default TTL.
+            ("/tmp/jt-fresh.mp4", "auto", "2026-05-31 00:00:00", None),
+            // Old, but an explicit override keeps it alive.
+            (
+                "/tmp/jt-override.mp4",
+                "auto",
+                "2026-01-01 00:00:00",
+                Some("2026-12-01T00:00:00+00:00"),
+            ),
+            // A user download must never carry an expiry, but assert the sweep
+            // ignores it even if one were somehow set.
+            (
+                "/tmp/jt-user.mp4",
+                "user",
+                "2026-01-01 00:00:00",
+                Some("2026-01-01T00:00:00+00:00"),
+            ),
+            (
+                "/tmp/jt-user-noexp.mp4",
+                "user",
+                "2026-01-01 00:00:00",
+                None,
+            ),
+        ] {
+            conn.execute(
+                "INSERT INTO downloads (user_id, status, file_size, file_path, download_source, completed_at, expires_at)
+                 VALUES ('user1', 'completed', 10, ?1, ?2, ?3, ?4)",
+                rusqlite::params![path, source, completed, expires],
+            )
+            .unwrap();
+        }
+
+        let conn_arc = Arc::new(Mutex::new(conn));
+        let db_service = Arc::new(RusqliteService::new(conn_arc.clone()));
+        let cache = SmartCache::new(CacheConfig::default());
+
+        let reclaimed = cache
+            .reclaim_expired_async(&db_service, "user1", "2026-06-01T00:00:00+00:00")
+            .await
+            .unwrap();
+        assert_eq!(
+            reclaimed, 1,
+            "only the expired temporary entry is reclaimed"
+        );
+
+        let surviving: Vec<String> = {
+            let guard = conn_arc.lock_safe();
+            let mut stmt = guard
+                .prepare("SELECT file_path FROM downloads ORDER BY id")
+                .unwrap();
+            let rows = stmt
+                .query_map([], |row| row.get::<_, String>(0))
+                .unwrap()
+                .map(|r| r.unwrap())
+                .collect();
+            rows
+        };
+        assert_eq!(
+            surviving,
+            vec![
+                "/tmp/jt-fresh.mp4".to_string(),
+                "/tmp/jt-override.mp4".to_string(),
+                "/tmp/jt-user.mp4".to_string(),
+                "/tmp/jt-user-noexp.mp4".to_string(),
+            ],
+            "entries within their life, those with a later override, and every user download must survive"
+        );
+
+        // TTL of 0 disables time-based reclaim entirely.
+        let cache_no_ttl = SmartCache::new(CacheConfig {
+            temporary_ttl_hours: 0,
+            ..Default::default()
+        });
+        assert_eq!(
+            cache_no_ttl
+                .reclaim_expired_async(&db_service, "user1", "2027-01-01T00:00:00+00:00")
+                .await
+                .unwrap(),
+            0,
+            "a zero TTL leaves space pressure as the only reclaim trigger"
+        );
+    }
+
+    /// Eviction must only reclaim *temporary* (`download_source = 'auto'`)
+    /// downloads — the precache tier. A download the user explicitly asked for
+    /// is their file: it may be deleted by them, never by the cache making room
+    /// for a predictive fetch.
+    ///
+    /// Before the fix, `evict_lru_async` selected every completed row ordered by
+    /// `completed_at ASC` with no source filter, so hitting the storage limit
+    /// deleted the *oldest* download — typically the film someone downloaded for
+    /// a flight — in favour of a newer auto-precached track.
+    ///
+    /// TRACES: UR-071 | DR-126 | UT-108
+    #[tokio::test]
+    async fn test_evict_lru_never_deletes_user_downloads() {
+        use crate::storage::db_service::RusqliteService;
+        use rusqlite::Connection;
+        use std::sync::{Arc, Mutex};
+
+        let conn = Connection::open_in_memory().unwrap();
+        conn.execute(
+            "CREATE TABLE downloads (
+                id INTEGER PRIMARY KEY,
+                user_id TEXT,
+                status TEXT,
+                file_size INTEGER,
+                file_path TEXT,
+                completed_at TEXT,
+                download_source TEXT DEFAULT 'user'
+            )",
+            [],
+        )
+        .unwrap();
+
+        // The user's own download is the OLDEST, so a purely time-ordered
+        // eviction would take it first.
+        conn.execute(
+            "INSERT INTO downloads (user_id, status, file_size, file_path, completed_at, download_source)
+             VALUES ('user1', 'completed', 600, '/tmp/jellytau-test-user.mp4', '2026-01-01', 'user')",
+            [],
+        )
+        .unwrap();
+        // A newer, auto-precached item.
+        conn.execute(
+            "INSERT INTO downloads (user_id, status, file_size, file_path, completed_at, download_source)
+             VALUES ('user1', 'completed', 600, '/tmp/jellytau-test-auto.mp4', '2026-06-01', 'auto')",
+            [],
+        )
+        .unwrap();
+
+        let conn_arc = Arc::new(Mutex::new(conn));
+        let db_service = Arc::new(RusqliteService::new(conn_arc.clone()));
+
+        let cache = SmartCache::new(CacheConfig {
+            storage_limit: 1000,
+            ..Default::default()
+        });
+
+        // 1200 bytes held against a 1000 limit: eviction must free something.
+        let freed = cache
+            .evict_lru_async(&db_service, "user1", 0)
+            .await
+            .unwrap();
+        assert!(freed > 0, "eviction should have reclaimed the auto entry");
+
+        let surviving: Vec<String> = {
+            let guard = conn_arc.lock_safe();
+            let mut stmt = guard
+                .prepare("SELECT download_source FROM downloads ORDER BY id")
+                .unwrap();
+            let rows = stmt
+                .query_map([], |row| row.get::<_, String>(0))
+                .unwrap()
+                .map(|r| r.unwrap())
+                .collect();
+            rows
+        };
+
+        assert_eq!(
+            surviving,
+            vec!["user".to_string()],
+            "the user's own download must survive; only the 'auto' entry is evictable"
+        );
+    }
+
     #[tokio::test]
     async fn test_storage_limit_check() {
         use crate::storage::db_service::RusqliteService;

src-tauri/src/download/mod.rs

@@ -9,6 +9,7 @@
 pub mod cache;
 pub mod events;
 pub mod network;
+pub mod stop;
 pub mod worker;
 
 use crate::utils::lock::MutexSafe;

src-tauri/src/download/stop.rs

@@ -0,0 +1,150 @@
+//! Stop signalling for in-flight downloads.
+//!
+//! TRACES: UR-055 | DR-168
+//!
+//! Pausing and cancelling used to be database-only: `pause_download` wrote
+//! `status = 'paused'` and nothing else. No cancellation existed anywhere in the
+//! download stack — no token, no flag, no abort — so the streaming task kept
+//! running, kept writing bytes, and on finishing overwrote the row with
+//! `completed` or `failed`. The row flicked to "paused" and then undid itself,
+//! which is precisely the reported "pause does not work".
+//!
+//! This is the missing half: a flag per in-flight download that the worker reads
+//! between chunks. Setting it makes the worker return [`Stopped`] promptly and
+//! leave the `.part` file **intact**, which is what lets a resume pick up from
+//! where it stopped via the existing HTTP Range request.
+//!
+//! Kept as a module-level registry rather than on `DownloadManager` because the
+//! two sides never meet: the command handler holds the manager's lock, while the
+//! worker runs detached inside `tauri::async_runtime::spawn` with no access to
+//! Tauri state. A registry both can reach is the smallest thing that works.
+//!
+//! [`Stopped`]: crate::download::worker::DownloadError::Stopped
+
+use std::collections::HashMap;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::{Arc, Mutex, OnceLock};
+
+use crate::utils::lock::MutexSafe;
+
+/// download id → its stop flag, for downloads currently in flight.
+fn registry() -> &'static Mutex<HashMap<i64, Arc<AtomicBool>>> {
+    static REGISTRY: OnceLock<Mutex<HashMap<i64, Arc<AtomicBool>>>> = OnceLock::new();
+    REGISTRY.get_or_init(|| Mutex::new(HashMap::new()))
+}
+
+/// Register `download_id` as in-flight and hand back its stop flag.
+///
+/// Called by the worker as it starts. A previous flag for the same id is
+/// replaced, so a download that is paused and later resumed does not inherit the
+/// set flag from its last run and stop immediately.
+pub fn register(download_id: i64) -> Arc<AtomicBool> {
+    let flag = Arc::new(AtomicBool::new(false));
+    registry().lock_safe().insert(download_id, flag.clone());
+    flag
+}
+
+/// Ask an in-flight download to stop.
+///
+/// Returns whether one was actually in flight — the caller uses this to tell a
+/// running download (which will stop shortly) from a merely queued one (which
+/// the database update alone has already handled).
+pub fn signal(download_id: i64) -> bool {
+    match registry().lock_safe().get(&download_id) {
+        Some(flag) => {
+            flag.store(true, Ordering::SeqCst);
+            true
+        }
+        None => false,
+    }
+}
+
+/// Forget a download's flag. Called when its task finishes, however it ended.
+pub fn clear(download_id: i64) {
+    registry().lock_safe().remove(&download_id);
+}
+
+/// Whether a stop has been requested for `download_id`.
+///
+/// The worker reads its own `Arc<AtomicBool>` directly rather than looking the id
+/// up, so this exists for the tests that assert the registry's behaviour.
+#[cfg(test)]
+pub fn is_stopping(download_id: i64) -> bool {
+    registry()
+        .lock_safe()
+        .get(&download_id)
+        .map(|f| f.load(Ordering::SeqCst))
+        .unwrap_or(false)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Ids are per-test so the shared registry cannot leak between them.
+    fn unique_id(seed: i64) -> i64 {
+        900_000 + seed
+    }
+
+    #[test]
+    fn test_a_registered_download_starts_unflagged() {
+        let id = unique_id(1);
+        let flag = register(id);
+        assert!(!flag.load(Ordering::SeqCst));
+        assert!(!is_stopping(id));
+        clear(id);
+    }
+
+    #[test]
+    fn test_signal_sets_the_flag_the_worker_reads() {
+        let id = unique_id(2);
+        let flag = register(id);
+
+        assert!(signal(id), "a registered download reports as in flight");
+        assert!(
+            flag.load(Ordering::SeqCst),
+            "the worker's own handle sees it"
+        );
+        assert!(is_stopping(id));
+
+        clear(id);
+    }
+
+    /// The pump only needs to abort a task that exists; a queued row is handled
+    /// by its database status alone.
+    #[test]
+    fn test_signalling_an_unregistered_download_reports_not_in_flight() {
+        assert!(!signal(unique_id(3)));
+    }
+
+    #[test]
+    fn test_clear_forgets_the_download() {
+        let id = unique_id(4);
+        register(id);
+        signal(id);
+        clear(id);
+
+        assert!(!is_stopping(id));
+        assert!(!signal(id), "a cleared download is no longer in flight");
+    }
+
+    /// The bug this guards: pause sets the flag, and resume re-runs the same
+    /// download id. If registering reused the old flag, the resumed run would see
+    /// a set flag and stop instantly — a download that could never be resumed.
+    #[test]
+    fn test_reregistering_clears_a_previous_stop() {
+        let id = unique_id(5);
+        register(id);
+        signal(id);
+        assert!(is_stopping(id));
+
+        let fresh = register(id);
+        assert!(!fresh.load(Ordering::SeqCst));
+        assert!(
+            !is_stopping(id),
+            "a resumed download must not inherit the pause"
+        );
+
+        clear(id);
+    }
+}

src-tauri/src/download/worker.rs

@@ -1,6 +1,7 @@
 //! Download worker for HTTP streaming with progress tracking and retry logic
 
 use log::warn;
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
 
 use futures_util::StreamExt;
@@ -31,10 +32,18 @@ impl DownloadWorker {
         }
     }
 
-    /// Download a file with retry logic and progress tracking
+    /// Download a file with retry logic and progress tracking.
+    ///
+    /// `stop` is the pause/cancel flag (see [`crate::download::stop`]). It is
+    /// checked between chunks and again between retries, so a paused download
+    /// stops promptly rather than after its next backoff — up to 45 seconds
+    /// away, which reads as the pause having done nothing.
+    ///
+    /// TRACES: UR-055 | DR-168
     pub async fn download<F>(
         &self,
         task: &DownloadTask,
+        stop: &AtomicBool,
         on_progress: F,
     ) -> Result<DownloadResult, DownloadError>
     where
@@ -43,7 +52,10 @@ impl DownloadWorker {
         let mut retries = 0;
 
         loop {
-            match self.try_download(task, &on_progress).await {
+            if stop.load(Ordering::SeqCst) {
+                return Err(DownloadError::Stopped);
+            }
+            match self.try_download(task, stop, &on_progress).await {
                 Ok(result) => return Ok(result),
                 Err(e) if retries < self.max_retries && e.is_retryable() => {
                     retries += 1;
@@ -63,6 +75,7 @@ impl DownloadWorker {
     async fn try_download<F>(
         &self,
         task: &DownloadTask,
+        stop: &AtomicBool,
         on_progress: &F,
     ) -> Result<DownloadResult, DownloadError>
     where
@@ -76,7 +89,7 @@ impl DownloadWorker {
         }
 
         // Check for partial download
-        let temp_path = task.target_path.with_extension("part");
+        let temp_path = partial_path(&task.target_path);
         let existing_bytes = if temp_path.exists() {
             fs::metadata(&temp_path).await.map(|m| m.len()).unwrap_or(0)
         } else {
@@ -100,22 +113,32 @@ impl DownloadWorker {
             return Err(DownloadError::Http(response.status().as_u16()));
         }
 
-        // Get content length
+        // Did the server actually honour the Range? A transcode does not, and
+        // answers 200 with the whole stream — appending that would duplicate what
+        // we already hold. (DR-170)
+        let resume_from = resume_offset(existing_bytes, response.status().as_u16());
+        if existing_bytes > 0 && resume_from == 0 {
+            warn!(
+                "Server ignored the Range request (HTTP {}) — restarting {} from the beginning \
+                 instead of appending to {} existing bytes",
+                response.status().as_u16(),
+                task.target_path.display(),
+                existing_bytes
+            );
+        }
+
+        // Get content length. Absent on a chunked transcode, which is why progress
+        // for a non-`original` preset has no percentage to show.
         let _total_bytes = response
             .headers()
             .get(reqwest::header::CONTENT_LENGTH)
             .and_then(|v| v.to_str().ok())
             .and_then(|v| v.parse::<u64>().ok())
-            .map(|len| {
-                if existing_bytes > 0 {
-                    len + existing_bytes
-                } else {
-                    len
-                }
-            });
+            .map(|len| len + resume_from);
 
-        // Open file for appending
-        let mut file = if existing_bytes > 0 {
+        // Append only when resuming a range the server agreed to; otherwise
+        // create/truncate so the restarted stream replaces the stale bytes.
+        let mut file = if resume_from > 0 {
             fs::OpenOptions::new().append(true).open(&temp_path).await
         } else {
             fs::File::create(&temp_path).await
@@ -123,11 +146,22 @@ impl DownloadWorker {
         .map_err(|e| DownloadError::FileSystem(e.to_string()))?;
 
         // Stream download with progress tracking
-        let mut downloaded = existing_bytes;
+        let mut downloaded = resume_from;
         let mut stream = response.bytes_stream();
         let mut last_progress_emit = std::time::Instant::now();
 
         while let Some(chunk) = stream.next().await {
+            // Checked before writing, so a paused download stops on a byte
+            // boundary the `.part` file already accounts for — the Range request
+            // on resume then asks for exactly what is missing. Flushing what we
+            // have and leaving the file in place is the whole mechanism behind
+            // "resume", so this must never delete it. (DR-168)
+            if stop.load(Ordering::SeqCst) {
+                let _ = file.flush().await;
+                let _ = file.sync_all().await;
+                return Err(DownloadError::Stopped);
+            }
+
             let chunk = chunk.map_err(|e| DownloadError::Network(e.to_string()))?;
 
             file.write_all(&chunk)
@@ -138,7 +172,7 @@ impl DownloadWorker {
 
             // Emit progress every 500ms or every MB
             if last_progress_emit.elapsed() > Duration::from_millis(500)
-                || downloaded % (1024 * 1024) == 0
+                || downloaded.is_multiple_of(1024 * 1024)
             {
                 last_progress_emit = std::time::Instant::now();
                 on_progress(downloaded, _total_bytes);
@@ -167,6 +201,56 @@ impl DownloadWorker {
     }
 }
 
+/// Where to resume writing a partial download, given how the server answered.
+///
+/// A byte offset of 0 means "start the file again"; anything else means "append
+/// from here".
+///
+/// This is what makes non-`original` downloads survive. Those presets ask
+/// Jellyfin to **transcode**, and a live transcode is chunked with no
+/// `Content-Length` and cannot be byte-seeked: the server ignores `Range` and
+/// answers `200` with the whole stream from the beginning, not `206` with the
+/// requested tail. The worker sent the header and appended the body regardless,
+/// so every retry — and every resume — concatenated a fresh copy of the whole
+/// transcode onto the bytes already on disk. The file grew past its real size
+/// and would not play. Only a `206` actually promises the tail; a `200` means we
+/// must discard what we have and take the stream from the top.
+///
+/// TRACES: UR-071 | DR-170
+pub fn resume_offset(existing_bytes: u64, status: u16) -> u64 {
+    if existing_bytes == 0 {
+        return 0;
+    }
+    // 206 Partial Content is the only answer that honours the Range request.
+    if status == 206 {
+        existing_bytes
+    } else {
+        0
+    }
+}
+
+/// The partial-download sidecar for `target`.
+///
+/// **Appends** `.part` rather than replacing the extension. The worker used
+/// `Path::with_extension("part")`, which replaces: `movie.mp4` became
+/// `movie.part`. Every cleanup path meanwhile deleted `"{file_path}.part"` —
+/// `movie.mp4.part` — so nothing ever matched and the partial file of every
+/// cancelled or failed download was left on disk forever, invisible to the
+/// disk-usage totals because no `downloads` row pointed at it. That is the
+/// reported "failure is not cleaned".
+///
+/// Appending also removes a collision the old form had: `movie.mp4` and
+/// `movie.mkv` both mapped to `movie.part` and would have fought over one file.
+///
+/// One function so the writer and the cleaners cannot disagree again.
+///
+/// TRACES: UR-055 | DR-169
+pub fn partial_path(target: &std::path::Path) -> std::path::PathBuf {
+    let mut s = target.as_os_str().to_os_string();
+    s.push(".part");
+    std::path::PathBuf::from(s)
+}
+
 /// Result of a successful download
 #[derive(Debug)]
 pub struct DownloadResult {
@@ -179,6 +263,10 @@ pub enum DownloadError {
     Network(String),
     Http(u16),
     FileSystem(String),
+    /// The download was asked to stop (paused or cancelled). Not a failure: the
+    /// row's status already says what happened, and the partial file is kept so a
+    /// resume can continue from it.
+    Stopped,
 }
 
 impl DownloadError {
@@ -188,8 +276,16 @@ impl DownloadError {
             DownloadError::Network(_) => true,
             DownloadError::Http(status) => *status >= 500, // Retry server errors
             DownloadError::FileSystem(_) => false,
+            // Retrying would restart the very download the user just paused.
+            DownloadError::Stopped => false,
         }
     }
+
+    /// Whether this outcome means "the user stopped it", rather than a failure to
+    /// record and report.
+    pub fn is_stopped(&self) -> bool {
+        matches!(self, DownloadError::Stopped)
+    }
 }
 
 impl std::fmt::Display for DownloadError {
@@ -198,6 +294,7 @@ impl std::fmt::Display for DownloadError {
             DownloadError::Network(msg) => write!(f, "Network error: {}", msg),
             DownloadError::Http(status) => write!(f, "HTTP error {}", status),
             DownloadError::FileSystem(msg) => write!(f, "File system error: {}", msg),
+            DownloadError::Stopped => write!(f, "Download stopped by request"),
         }
     }
 }
@@ -208,6 +305,64 @@ impl std::error::Error for DownloadError {}
 mod tests {
     use super::*;
 
+    /// The bitrate-download corruption: a transcode ignores `Range` and answers
+    /// `200` with the whole stream. Appending that to the bytes already on disk
+    /// duplicated them, so every retry grew the file past its real size and left
+    /// it unplayable. Only `206` promises the requested tail.
+    ///
+    /// TRACES: UR-071 | DR-170 | UT-164
+    #[test]
+    fn test_resume_offset_only_appends_when_the_server_honoured_the_range() {
+        // Nothing on disk: start at the beginning either way.
+        assert_eq!(resume_offset(0, 200), 0);
+        assert_eq!(resume_offset(0, 206), 0);
+
+        // The server agreed to the range — append to what we have.
+        assert_eq!(resume_offset(5_000, 206), 5_000);
+
+        // The server ignored it and is sending the whole file (a transcode).
+        // Restart, or the bytes are duplicated.
+        assert_eq!(
+            resume_offset(5_000, 200),
+            0,
+            "a 200 carries the whole stream; appending it corrupts the file"
+        );
+    }
+
+    /// The regression: `with_extension` replaced the extension, so the worker
+    /// wrote `movie.part` while every cleanup path deleted `movie.mp4.part`.
+    /// Nothing matched, and partial files accumulated forever.
+    ///
+    /// TRACES: UR-055 | DR-169 | UT-163
+    #[test]
+    fn test_partial_path_appends_rather_than_replacing_the_extension() {
+        use std::path::Path;
+
+        assert_eq!(
+            partial_path(Path::new("/media/movie.mp4")),
+            Path::new("/media/movie.mp4.part"),
+            "the cleanup paths delete \"{{file_path}}.part\"; this must produce it"
+        );
+
+        // Two sources for one title must not fight over a single partial file.
+        assert_ne!(
+            partial_path(Path::new("/media/movie.mp4")),
+            partial_path(Path::new("/media/movie.mkv")),
+        );
+
+        // Extension-less targets still get a sidecar rather than being clobbered.
+        assert_eq!(
+            partial_path(Path::new("/media/track")),
+            Path::new("/media/track.part"),
+        );
+
+        // A dotted name keeps every part of its own name.
+        assert_eq!(
+            partial_path(Path::new("/media/S01.E02.episode.mkv")),
+            Path::new("/media/S01.E02.episode.mkv.part"),
+        );
+    }
+
     #[test]
     fn test_exponential_backoff() {
         assert_eq!(
@@ -231,5 +386,9 @@ mod tests {
         assert!(DownloadError::Http(503).is_retryable());
         assert!(!DownloadError::Http(404).is_retryable());
         assert!(!DownloadError::FileSystem("disk full".to_string()).is_retryable());
+        // Retrying a paused download would restart what the user just stopped.
+        assert!(!DownloadError::Stopped.is_retryable());
+        assert!(DownloadError::Stopped.is_stopped());
+        assert!(!DownloadError::Network("timeout".to_string()).is_stopped());
     }
 }

src-tauri/src/lib.rs

@@ -5,6 +5,7 @@ mod credentials;
 mod domain;
 mod download;
 mod jellyfin;
+mod media_server;
 mod playback_mode;
 mod playback_reporting;
 mod player;
@@ -85,6 +86,7 @@ use commands::{
     lms_unsync_player,
     mark_download_completed,
     mark_download_failed,
+    media_local_url,
     offline_get_items,
     offline_is_available,
     offline_search,
@@ -121,12 +123,17 @@ use commands::{
     player_get_audio_settings,
     player_get_autoplay_settings,
     player_get_cache_config,
+    player_get_capabilities,
+    player_get_eq_presets,
     player_get_queue,
     // Session management commands
     player_get_session,
     player_get_sleep_timer,
     player_get_status,
+    player_get_streaming_qualities,
     player_get_video_settings,
+    // Preload commands
+    player_local_media_path,
     player_move_in_queue,
     player_next,
     player_on_playback_ended,
@@ -137,9 +144,9 @@ use commands::{
     player_play_next_episode,
     player_play_queue,
     player_play_tracks,
-    // Preload commands
     player_preload_upcoming,
     player_previous,
+    player_recover_stream,
     player_remove_from_queue,
     player_report_media_loaded,
     player_report_position,
@@ -153,6 +160,7 @@ use commands::{
     player_set_cache_config,
     // Sleep timer and autoplay commands
     player_set_sleep_timer,
+    player_set_stream_quality,
     player_set_subtitle_track,
     player_set_video_settings,
     player_set_volume,
@@ -177,6 +185,7 @@ use commands::{
     remote_session_set_volume,
     remote_session_toggle_mute,
     // Repository commands
+    repository_clear_watch_history,
     repository_create,
     repository_destroy,
     repository_get_audio_only_stream_url_for_video,
@@ -185,6 +194,7 @@ use commands::{
     repository_get_download_disk_usage,
     repository_get_downloaded_items,
     repository_get_downloaded_libraries,
+    repository_get_favorites,
     repository_get_genres,
     repository_get_image_url,
     repository_get_item,
@@ -200,6 +210,8 @@ use commands::{
     repository_get_rediscover_albums,
     repository_get_resume_items,
     repository_get_resume_movies,
+    repository_get_series_current_episode,
+    repository_get_series_episodes,
     repository_get_similar_items,
     repository_get_subtitle_url,
     repository_get_video_download_url,
@@ -255,6 +267,7 @@ use commands::{
     storage_save_user,
     storage_search_items,
     storage_set_active_user,
+    storage_set_watched,
     storage_toggle_favorite,
     storage_update_playback_context,
     storage_update_playback_progress,
@@ -266,6 +279,7 @@ use commands::{
     sync_mark_completed,
     sync_mark_failed,
     sync_mark_processing,
+    sync_process_pending,
     // Sync queue commands
     sync_queue_mutation,
     thumbnail_clear_cache,
@@ -413,6 +427,28 @@ impl MediaSessionHandler {
 
     /// Drive the local player for a transport command.
     fn handle_local_command(&self, command: &str) {
+        // A lockscreen scrub is an ABSOLUTE position — the scrubber shows the
+        // whole episode — and resolving it during a background-audio handoff means
+        // re-opening the stream, which is async. So it runs on the runtime and,
+        // critically, is handled *before* the blocking lock below: taking that
+        // guard and then spawning a task that waits for the same mutex would
+        // deadlock the media session. (DR-159)
+        if let Some(raw) = command.strip_prefix("seek:") {
+            match raw.parse::<f64>() {
+                Ok(position) => {
+                    let player = self.player.clone();
+                    tokio::spawn(async move {
+                        let controller = player.lock().await;
+                        if let Err(e) = controller.seek_absolute(position).await {
+                            error!("[MediaSession] Seek to {:.1}s failed: {}", position, e);
+                        }
+                    });
+                }
+                Err(_) => warn!("[MediaSession] Bad seek command: {}", command),
+            }
+            return;
+        }
+
         // Use blocking_lock since this is called from a non-async JNI callback
         let controller = self.player.blocking_lock();
 
@@ -422,13 +458,6 @@ impl MediaSessionHandler {
             "next" => controller.next(),
             "previous" => controller.previous(),
             "stop" => controller.stop(),
-            cmd if cmd.starts_with("seek:") => match cmd[5..].parse::<f64>() {
-                Ok(pos) => controller.seek(pos),
-                Err(_) => {
-                    warn!("[MediaSession] Bad seek command: {}", command);
-                    Ok(())
-                }
-            },
             _ => {
                 warn!("[MediaSession] Unknown command: {}", command);
                 Ok(())
@@ -515,6 +544,12 @@ fn emit_backend_init_failed(app_handle: &tauri::AppHandle, backend: &'static str
 }
 
 /// Create the appropriate player backend for the current platform.
+// playback_reporter/position_throttler are consumed only by the native audio
+// backends (mpv/exo); on platforms using the webview audio backend they're unused.
+#[cfg_attr(
+    not(any(target_os = "linux", target_os = "android")),
+    allow(unused_variables)
+)]
 fn create_player_backend(
     app_handle: tauri::AppHandle,
     playback_reporter: Arc<tokio::sync::Mutex<Option<playback_reporting::PlaybackReporter>>>,
@@ -615,11 +650,19 @@ fn create_player_backend(
         }
     }
 
-    // Fallback for other platforms
+    // Platforms with no native audio backend (e.g. Windows): render audio-only
+    // playback through a webview <audio> element (all video already renders in
+    // the webview). Falls back to NullBackend only if the backend can't init.
     #[cfg(not(any(target_os = "linux", target_os = "android")))]
     {
-        warn!("WARNING: No audio backend available for this platform");
-        Box::new(NullBackend::new())
+        info!("No native audio backend for this platform - using webview <audio> backend");
+        match player::WebviewAudioBackend::new(_event_emitter) {
+            Ok(backend) => Box::new(backend),
+            Err(e) => {
+                emit_backend_init_failed(&app_handle, "webview-audio", e.to_string());
+                Box::new(NullBackend::new())
+            }
+        }
     }
 }
 
@@ -658,6 +701,7 @@ fn specta_builder() -> Builder<tauri::Wry> {
             player_cycle_repeat,
             player_get_status,
             player_get_queue,
+            player_get_capabilities,
             player_add_to_queue,
             player_add_track_by_id,
             player_add_tracks_by_ids,
@@ -666,8 +710,11 @@ fn specta_builder() -> Builder<tauri::Wry> {
             player_skip_to,
             player_set_audio_settings,
             player_get_audio_settings,
+            player_get_eq_presets,
             player_set_video_settings,
             player_get_video_settings,
+            player_get_streaming_qualities,
+            player_set_stream_quality,
             // Sleep timer and autoplay commands
             player_set_sleep_timer,
             player_cancel_sleep_timer,
@@ -677,10 +724,12 @@ fn specta_builder() -> Builder<tauri::Wry> {
             player_cancel_autoplay_countdown,
             player_play_next_episode,
             player_on_playback_ended,
+            player_recover_stream,
             player_report_state,
             player_report_position,
             player_report_media_loaded,
             // Preload commands
+            player_local_media_path,
             player_preload_upcoming,
             player_set_cache_config,
             player_get_cache_config,
@@ -760,6 +809,7 @@ fn specta_builder() -> Builder<tauri::Wry> {
             storage_update_playback_progress,
             storage_update_playback_context,
             storage_mark_played,
+            storage_set_watched,
             storage_get_playback_progress,
             storage_mark_synced,
             storage_toggle_favorite,
@@ -782,6 +832,7 @@ fn specta_builder() -> Builder<tauri::Wry> {
             get_download_storage_stats,
             mark_download_completed,
             mark_download_failed,
+            media_local_url,
             start_download,
             enqueue_download,
             enqueue_video_downloads,
@@ -822,6 +873,7 @@ fn specta_builder() -> Builder<tauri::Wry> {
             sync_mark_completed,
             sync_mark_failed,
             sync_get_pending_count,
+            sync_process_pending,
             sync_cleanup_completed,
             sync_clear_user,
             // Thumbnail cache and image commands
@@ -853,6 +905,9 @@ fn specta_builder() -> Builder<tauri::Wry> {
             repository_get_latest_items,
             repository_get_resume_items,
             repository_get_next_up_episodes,
+            repository_get_series_episodes,
+            repository_get_series_current_episode,
+            repository_clear_watch_history,
             repository_get_recently_played_audio,
             repository_get_resume_movies,
             repository_get_rediscover_albums,
@@ -871,6 +926,7 @@ fn specta_builder() -> Builder<tauri::Wry> {
             repository_get_image_url,
             repository_mark_favorite,
             repository_unmark_favorite,
+            repository_get_favorites,
             repository_get_person,
             repository_get_items_by_person,
             repository_get_similar_items,
@@ -973,6 +1029,16 @@ fn set_env_if_unset(key: &str, value: &str) {
     }
 }
 
+/// Downloaded media and cached thumbnails are handed to the webview as
+/// `http://asset.localhost/…` URLs by `convertFileSrc`. Tauri only answers that
+/// origin when the `protocol-asset` cargo feature is compiled in *and*
+/// `app.security.assetProtocol.enable` is set in `tauri.conf.json`, which also
+/// scopes it to `$APPDATA/**` — the storage root holding the database,
+/// `downloads/` and the thumbnail cache. Both are required together: with either
+/// missing the URL resolves to nothing and the webview reports
+/// `NETWORK_NO_SOURCE`, which is how offline video came to fail silently.
+///
+/// TRACES: UR-071 | DR-134
 #[cfg_attr(mobile, tauri::mobile_entry_point)]
 pub fn run() {
     // Initialize logger
@@ -1180,8 +1246,19 @@ pub fn run() {
             let video_settings = VideoSettingsWrapper(Mutex::new(VideoSettings::default()));
             app.manage(video_settings);
 
-            // Background-audio handoff base offset (UR-040).
-            app.manage(commands::player::BackgroundAudioOffset::default());
+            // Restore the persisted streaming bandwidth ceiling. Deferred to the
+            // async runtime because the read is async, and ordered after the
+            // wrapper above because it writes into it. Until it lands, streams
+            // are uncapped — the pre-existing behaviour — and no playback can
+            // have started this early anyway (login happens after setup).
+            //
+            // TRACES: UR-074 | DR-162
+            {
+                let handle = app.handle().clone();
+                tauri::async_runtime::spawn(async move {
+                    crate::commands::restore_streaming_quality(&handle).await;
+                });
+            }
 
             // Initialize thumbnail cache
             info!("[INIT] Initializing thumbnail cache...");
@@ -1203,6 +1280,22 @@ pub fn run() {
             let smart_cache_wrapper = SmartCacheWrapper(Mutex::new(smart_cache));
             app.manage(smart_cache_wrapper);
 
+            // Serve downloaded media over loopback HTTP. The webview cannot
+            // stream a large file through the asset protocol (see media_server),
+            // so local playback resolves its URL from here instead.
+            // TRACES: UR-071 | DR-137
+            info!("[INIT] Starting local media server...");
+            let media_server = match media_server::start(app_data_dir.clone()) {
+                Ok(s) => Some(s),
+                Err(e) => {
+                    // Not fatal: streaming still works, and the command reports
+                    // a clear error if local playback is attempted.
+                    error!("[INIT ERROR] Local media server failed to start: {}", e);
+                    None
+                }
+            };
+            app.manage(media_server::MediaServerWrapper(media_server));
+
             // Initialize download manager
             info!("[INIT] Initializing download manager...");
             let download_dir = app_data_dir.join("downloads");
@@ -1269,6 +1362,28 @@ pub fn run() {
             let playback_reporter_wrapper = PlaybackReporterWrapper(playback_reporter.clone());
             app.manage(playback_reporter_wrapper);
 
+            // Keep the local search index fresh. Ownership of *when* to re-index
+            // sits here rather than in the frontend: it is sync policy over
+            // domain data, and a startup-only trigger left a long session
+            // searching a stale catalog.
+            // TRACES: UR-065 | DR-109, IR-030
+            info!("[INIT] Starting background catalog indexer...");
+            commands::catalog::spawn_catalog_indexer(app.handle().clone());
+
+            // Push favourite toggles made while the server was unreachable, on
+            // every reconnect. In Rust rather than the frontend so it runs
+            // whether or not the screen that made the change is still mounted.
+            // TRACES: UR-069 | DR-120
+            info!("[INIT] Starting favourites drain...");
+            commands::favorites::spawn_favorites_drain(app.handle().clone());
+
+            // Push playback reports queued while the server was unreachable.
+            // Without this the `sync_queue` rows the reporter writes offline
+            // are never sent and the offline banner's count only grows.
+            // TRACES: UR-025, UR-002 | DR-131
+            info!("[INIT] Starting sync-queue drain...");
+            commands::sync_drain::spawn_sync_queue_drain(app.handle().clone());
+
             info!("[INIT] Application setup completed successfully");
             Ok(())
         })

src-tauri/src/media_server.rs

@@ -0,0 +1,588 @@
+//! A loopback HTTP server for locally downloaded media.
+//!
+//! Tauri's built-in `asset` protocol cannot serve a downloaded film to the
+//! webview. Its response to a request *without* a `Range` header reads the whole
+//! file into a `Vec<u8>`, and it only advertises `Accept-Ranges: bytes` from
+//! inside the range branch — so the first request never learns ranges are
+//! available and a multi-gigabyte body is attempted instead. Chromium abandoned
+//! it with `PIPELINE_ERROR_READ` after ~31s, which reached the user as
+//! "downloaded video does not play offline".
+//!
+//! Serving over real HTTP on 127.0.0.1 rather than a custom URI scheme is
+//! deliberate: it makes range support a property of the transport instead of
+//! depending on whether a platform's webview forwards `Range` to a custom
+//! scheme, which differs between Android and the desktop webviews.
+//!
+//! Two things confine it, because **loopback is shared between apps on
+//! Android** — any other installed app can connect to this port:
+//!
+//! - it binds `127.0.0.1` only, so nothing off-device can reach it; and
+//! - every URL carries a random per-session token, so another app cannot guess a
+//!   working URL, and paths are confined to the app data directory even if one
+//!   did.
+//!
+//! Phase 1 serves local files only. The same origin is the intended home for
+//! remote passthrough (and download-while-watching) later; see the stage-2 spec.
+//!
+//! TRACES: UR-071 | DR-137 | UT-127
+
+use std::fs::File;
+use std::io::{Read, Seek, SeekFrom};
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use log::{debug, error, info, warn};
+use rand::Rng;
+use tiny_http::{Header, Response, Server, StatusCode};
+
+/// Bytes per response. Large enough that a film needs relatively few round
+/// trips, small enough that one response is never a memory problem on a phone.
+/// Tauri's asset protocol uses 1 MiB; 4 MiB quarters the request count for the
+/// multi-gigabyte files this exists to serve.
+const CHUNK_LEN: u64 = 4 * 1024 * 1024;
+
+/// Managed state. `None` when the server could not bind — local playback then
+/// fails with a clear error instead of the app refusing to start.
+pub struct MediaServerWrapper(pub Option<MediaServer>);
+
+/// A running server. Dropping this does not stop the thread; the server lives
+/// for the life of the process by design, since playback can start at any time.
+pub struct MediaServer {
+    port: u16,
+    token: String,
+}
+
+impl MediaServer {
+    /// The base a media URL is built on, e.g. `http://127.0.0.1:53412/<token>`.
+    pub fn base_url(&self) -> String {
+        format!("http://127.0.0.1:{}/{}", self.port, self.token)
+    }
+
+    /// A playable URL for an absolute on-disk path.
+    pub fn url_for(&self, path: &str) -> String {
+        format!("{}/{}", self.base_url(), urlencoding::encode(path))
+    }
+}
+
+/// Bind to an ephemeral loopback port and start serving `root` in a background
+/// thread.
+///
+/// TRACES: UR-071 | DR-137
+pub fn start(root: PathBuf) -> Result<MediaServer, String> {
+    // Port 0 → the OS picks a free one. Binding 127.0.0.1 (not 0.0.0.0) keeps
+    // this off the network.
+    let server =
+        Server::http("127.0.0.1:0").map_err(|e| format!("Failed to bind media server: {e}"))?;
+
+    let port = server
+        .server_addr()
+        .to_ip()
+        .ok_or_else(|| "Media server bound to a non-IP address".to_string())?
+        .port();
+
+    let token: String = {
+        let mut rng = rand::thread_rng();
+        (0..32)
+            .map(|_| char::from_digit(rng.gen_range(0..16), 16).unwrap())
+            .collect()
+    };
+
+    info!(
+        "[MediaServer] Serving {} on 127.0.0.1:{}",
+        root.display(),
+        port
+    );
+
+    let server = Arc::new(server);
+    let shared = Arc::new((root, token.clone()));
+
+    std::thread::Builder::new()
+        .name("media-server".into())
+        .spawn(move || loop {
+            let request = match server.recv() {
+                Ok(r) => r,
+                Err(e) => {
+                    error!("[MediaServer] accept failed: {e}");
+                    continue;
+                }
+            };
+            let shared = Arc::clone(&shared);
+            // A thread per request: media clients open several connections at
+            // once, and a blocking read of one must not stall the others.
+            if let Err(e) = std::thread::Builder::new()
+                .name("media-server-req".into())
+                .spawn(move || {
+                    let (root, token) = &*shared;
+                    handle(request, root, token);
+                })
+            {
+                error!("[MediaServer] could not spawn handler: {e}");
+            }
+        })
+        .map_err(|e| format!("Failed to start media server thread: {e}"))?;
+
+    Ok(MediaServer { port, token })
+}
+
+fn header(name: &str, value: &str) -> Header {
+    Header::from_bytes(name.as_bytes(), value.as_bytes())
+        .expect("static header name/value are valid")
+}
+
+fn empty(status: u16) -> Response<std::io::Empty> {
+    Response::empty(StatusCode(status)).with_header(header("Accept-Ranges", "bytes"))
+}
+
+fn handle(request: tiny_http::Request, root: &Path, token: &str) {
+    let url = request.url().to_string();
+    let method = request.method().as_str().to_string();
+
+    let outcome = match route(&url, root, token) {
+        Ok(path) => path,
+        Err(status) => {
+            let _ = request.respond(empty(status));
+            return;
+        }
+    };
+
+    if method != "GET" && method != "HEAD" {
+        let _ = request.respond(empty(405));
+        return;
+    }
+
+    let mut file = match File::open(&outcome) {
+        Ok(f) => f,
+        Err(e) => {
+            warn!("[MediaServer] {}: {}", outcome.display(), e);
+            let _ = request.respond(empty(404));
+            return;
+        }
+    };
+    let len = match file.metadata() {
+        Ok(m) => m.len(),
+        Err(e) => {
+            warn!(
+                "[MediaServer] metadata failed for {}: {}",
+                outcome.display(),
+                e
+            );
+            let _ = request.respond(empty(404));
+            return;
+        }
+    };
+
+    let range = request
+        .headers()
+        .iter()
+        .find(|h| h.field.equiv("Range"))
+        .map(|h| h.value.as_str().to_string());
+
+    debug!(
+        "[MediaServer] {} {} ({} bytes) range={:?}",
+        method,
+        outcome.display(),
+        len,
+        range
+    );
+
+    let Some(span) = span_for(range.as_deref(), len) else {
+        let _ = request
+            .respond(empty(416).with_header(header("Content-Range", &format!("bytes */{len}"))));
+        return;
+    };
+
+    // Sniff before seeking to the span, for extension-less files.
+    let mut head = [0u8; 16];
+    let head_len = file.read(&mut head).unwrap_or(0);
+    let mime = content_type(&outcome, &head[..head_len]);
+
+    if method == "HEAD" {
+        let _ = request.respond(
+            empty(200)
+                .with_header(header("Content-Type", mime))
+                .with_header(header("Content-Length", &len.to_string())),
+        );
+        return;
+    }
+
+    if let Err(e) = file.seek(SeekFrom::Start(span.start)) {
+        warn!("[MediaServer] seek failed for {}: {}", outcome.display(), e);
+        let _ = request.respond(empty(500));
+        return;
+    }
+
+    // Streamed straight from the file handle: at no point is more than the
+    // span in memory, and the span is capped at CHUNK_LEN.
+    let nbytes = span.len();
+    let body = file.take(nbytes);
+    // tiny_http switches to chunked transfer above a 32 KiB default, which drops
+    // Content-Length — and a 206 without one is unusable to Chromium's media
+    // loader, which needs the range's size. Raising the threshold past our own
+    // cap keeps every response length-delimited.
+    let response = Response::new(
+        StatusCode(206),
+        vec![
+            header("Accept-Ranges", "bytes"),
+            header("Content-Type", mime),
+            header(
+                "Content-Range",
+                &format!("bytes {}-{}/{}", span.start, span.end, len),
+            ),
+        ],
+        body,
+        Some(nbytes as usize),
+        None,
+    )
+    .with_chunked_threshold(usize::MAX);
+
+    if let Err(e) = request.respond(response) {
+        // A client that seeks away closes the connection mid-body; that is
+        // normal and must not be logged as a failure.
+        debug!("[MediaServer] response ended early: {e}");
+    }
+}
+
+/// Check the token and resolve the path, or return the status to answer with.
+fn route(url: &str, root: &Path, token: &str) -> Result<PathBuf, u16> {
+    let trimmed = url.trim_start_matches('/');
+    let (got_token, rest) = trimmed.split_once('/').ok_or(404u16)?;
+
+    // Constant-time-ish: length check first, then a byte compare. The token is
+    // the only thing standing between another app on the device and this server.
+    if got_token.len() != token.len() || got_token != token {
+        warn!("[MediaServer] Rejected a request with a bad token");
+        return Err(403);
+    }
+
+    // Strip any query string before decoding.
+    let raw = rest.split('?').next().unwrap_or("");
+    match resolve_path(raw, root) {
+        Resolved::Allow(p) => Ok(p),
+        Resolved::Forbidden => {
+            warn!("[MediaServer] Refused a path outside the app data directory");
+            Err(403)
+        }
+    }
+}
+
+/// What a request path resolved to, before any file is touched.
+#[derive(Debug, PartialEq, Eq)]
+pub enum Resolved {
+    Allow(PathBuf),
+    /// Escaped the allowed root.
+    Forbidden,
+}
+
+/// Resolve a percent-encoded request path to a file inside `root`.
+///
+/// `..` segments are folded away lexically rather than through `canonicalize`,
+/// so a missing file still resolves (and then 404s) instead of being reported as
+/// a scope violation.
+///
+/// TRACES: UR-071 | DR-137 | UT-127
+pub fn resolve_path(raw: &str, root: &Path) -> Resolved {
+    let decoded = match urlencoding::decode(raw) {
+        Ok(d) => d.into_owned(),
+        Err(_) => raw.to_string(),
+    };
+
+    let mut normalised = PathBuf::new();
+    for part in Path::new(&decoded).components() {
+        match part {
+            std::path::Component::ParentDir => {
+                normalised.pop();
+            }
+            std::path::Component::CurDir => {}
+            other => normalised.push(other),
+        }
+    }
+
+    if normalised.starts_with(root) {
+        Resolved::Allow(normalised)
+    } else {
+        Resolved::Forbidden
+    }
+}
+
+/// The byte range a response should carry. `end` is inclusive.
+#[derive(Debug, PartialEq, Eq)]
+pub struct Span {
+    pub start: u64,
+    pub end: u64,
+}
+
+impl Span {
+    pub fn len(&self) -> u64 {
+        self.end + 1 - self.start
+    }
+}
+
+/// Decide which span to send for a `Range` header (or its absence).
+///
+/// `None` means unsatisfiable — answer 416. A missing or unparseable header
+/// yields the first chunk, so a client that did not ask for a range still gets a
+/// bounded response it can continue from, which is exactly the case the asset
+/// protocol answered with the whole file.
+///
+/// TRACES: UR-071 | DR-137 | UT-127
+pub fn span_for(range: Option<&str>, len: u64) -> Option<Span> {
+    if len == 0 {
+        return Some(Span { start: 0, end: 0 });
+    }
+    let last = len - 1;
+    let first_chunk = Span {
+        start: 0,
+        end: (CHUNK_LEN - 1).min(last),
+    };
+
+    let Some(raw) = range else {
+        return Some(first_chunk);
+    };
+    let Some(spec) = raw.trim().strip_prefix("bytes=") else {
+        return Some(first_chunk);
+    };
+    // Only the first range of a multi-range request is honoured; media clients
+    // ask for one, and a single 206 is a valid answer either way.
+    let spec = spec.split(',').next().unwrap_or("").trim();
+    let Some((from, to)) = spec.split_once('-') else {
+        return Some(first_chunk);
+    };
+
+    let (start, end) = if from.is_empty() {
+        // Suffix form: `-500` is the final 500 bytes.
+        let suffix: u64 = match to.parse() {
+            Ok(n) => n,
+            Err(_) => return Some(first_chunk),
+        };
+        if suffix == 0 {
+            return None;
+        }
+        (len.saturating_sub(suffix), last)
+    } else {
+        let start: u64 = match from.parse() {
+            Ok(n) => n,
+            Err(_) => return Some(first_chunk),
+        };
+        let end = if to.is_empty() {
+            last
+        } else {
+            match to.parse::<u64>() {
+                Ok(n) => n.min(last),
+                Err(_) => return Some(first_chunk),
+            }
+        };
+        (start, end)
+    };
+
+    if start > last || end < start {
+        return None;
+    }
+
+    Some(Span {
+        start,
+        end: end.min(start + CHUNK_LEN - 1),
+    })
+}
+
+/// Guess a content type.
+///
+/// **Magic bytes win over the extension.** Downloading at `original` quality
+/// asks Jellyfin for a direct static copy, which returns the *source file's*
+/// bytes under a `.mp4` name whatever the real container is — a downloaded film
+/// named `.mp4` turned out to be an AVI holding XVID. Trusting the extension
+/// there labels it `video/mp4` and the player is handed a container that is not
+/// what the header claims. The extension is only a fallback for a file whose
+/// bytes are unrecognised, and for the extension-less files the offline queue
+/// writes under an item id.
+///
+/// TRACES: UR-071 | DR-137 | UT-127
+fn content_type(path: &Path, head: &[u8]) -> &'static str {
+    // `ftyp` at offset 4 marks an ISO base media file (mp4 and friends).
+    if head.len() > 11 && &head[4..8] == b"ftyp" {
+        return "video/mp4";
+    }
+    if head.len() > 11 && head.starts_with(b"RIFF") && &head[8..11] == b"AVI" {
+        return "video/x-msvideo";
+    }
+    if head.starts_with(b"\x1aE\xdf\xa3") {
+        return "video/x-matroska";
+    }
+    if head.starts_with(b"ID3") || head.starts_with(b"\xff\xfb") {
+        return "audio/mpeg";
+    }
+    if head.starts_with(b"OggS") {
+        return "audio/ogg";
+    }
+    if head.starts_with(b"fLaC") {
+        return "audio/flac";
+    }
+
+    match path
+        .extension()
+        .and_then(|e| e.to_str())
+        .map(|e| e.to_ascii_lowercase())
+        .as_deref()
+    {
+        Some("mp4" | "m4v" | "mov") => return "video/mp4",
+        Some("mkv") => return "video/x-matroska",
+        Some("webm") => return "video/webm",
+        Some("mp3") => return "audio/mpeg",
+        Some("m4a" | "aac") => return "audio/mp4",
+        Some("flac") => return "audio/flac",
+        Some("ogg" | "opus") => return "audio/ogg",
+        Some("wav") => return "audio/wav",
+        Some("avi") => return "video/x-msvideo",
+        _ => {}
+    }
+
+    "application/octet-stream"
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The whole point: a request with no `Range` must still come back bounded.
+    /// That is the case Tauri's asset protocol answers with the entire file —
+    /// the read Chromium abandoned after 31s.
+    ///
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn a_rangeless_request_is_answered_with_one_chunk_not_the_file() {
+        let huge = 8 * 1024 * 1024 * 1024; // 8 GiB
+        let span = span_for(None, huge).unwrap();
+        assert_eq!(span.start, 0);
+        assert_eq!(span.len(), CHUNK_LEN);
+    }
+
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn no_response_ever_exceeds_one_chunk() {
+        let len = 8 * 1024 * 1024 * 1024;
+        for h in [
+            "bytes=0-",
+            "bytes=0-99999999999",
+            "bytes=1024-",
+            "bytes=-99999999",
+        ] {
+            let span = span_for(Some(h), len).unwrap();
+            assert!(span.len() <= CHUNK_LEN, "{h} produced {} bytes", span.len());
+        }
+    }
+
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn ranges_are_honoured() {
+        let len = 1000u64;
+        assert_eq!(
+            span_for(Some("bytes=100-199"), len).unwrap(),
+            Span {
+                start: 100,
+                end: 199
+            }
+        );
+        // Open-ended runs to the end of a small file.
+        assert_eq!(
+            span_for(Some("bytes=900-"), len).unwrap(),
+            Span {
+                start: 900,
+                end: 999
+            }
+        );
+        // Suffix form.
+        assert_eq!(
+            span_for(Some("bytes=-100"), len).unwrap(),
+            Span {
+                start: 900,
+                end: 999
+            }
+        );
+        // Past the end is unsatisfiable, not a clamp — a clamp would make a
+        // seek past the end silently replay earlier bytes.
+        assert!(span_for(Some("bytes=1000-"), len).is_none());
+    }
+
+    /// A malformed header must not fail the request: playing from the start is
+    /// strictly better than refusing to open the file.
+    ///
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn a_malformed_range_falls_back_to_the_first_chunk() {
+        assert_eq!(span_for(Some("pages=1-2"), 5000).unwrap().start, 0);
+        assert_eq!(span_for(Some("bytes=abc-def"), 5000).unwrap().start, 0);
+    }
+
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn reads_are_confined_to_the_app_data_directory() {
+        let root = Path::new("/data/user/0/app");
+
+        assert_eq!(
+            resolve_path("/data/user/0/app/videos/f.mp4", root),
+            Resolved::Allow(PathBuf::from("/data/user/0/app/videos/f.mp4"))
+        );
+        // Percent-encoded, as the URL builder produces.
+        assert_eq!(
+            resolve_path("%2Fdata%2Fuser%2F0%2Fapp%2Fa%20b.mp4", root),
+            Resolved::Allow(PathBuf::from("/data/user/0/app/a b.mp4"))
+        );
+        // Traversal out of the root, and an unrelated absolute path, are refused.
+        assert_eq!(
+            resolve_path("/data/user/0/app/../../../etc/passwd", root),
+            Resolved::Forbidden
+        );
+        assert_eq!(resolve_path("/etc/passwd", root), Resolved::Forbidden);
+    }
+
+    /// Loopback is shared between apps on Android, so the token is the only
+    /// thing stopping another installed app from reading downloaded media.
+    ///
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn a_request_without_the_right_token_is_refused() {
+        let root = Path::new("/data/user/0/app");
+        let good = "0123456789abcdef0123456789abcdef";
+
+        assert_eq!(
+            route(
+                &format!("/{good}/%2Fdata%2Fuser%2F0%2Fapp%2Ff.mp4"),
+                root,
+                good
+            ),
+            Ok(PathBuf::from("/data/user/0/app/f.mp4"))
+        );
+        assert_eq!(
+            route("/wrong-token/%2Fdata%2Fuser%2F0%2Fapp%2Ff.mp4", root, good),
+            Err(403)
+        );
+        // No token segment at all.
+        assert_eq!(route("/f.mp4", root, good), Err(404));
+        // Right token, but a path outside the root is still refused.
+        assert_eq!(
+            route(&format!("/{good}/%2Fetc%2Fpasswd"), root, good),
+            Err(403)
+        );
+    }
+
+    /// TRACES: UR-071 | DR-137 | UT-127
+    #[test]
+    fn content_type_uses_the_extension_then_the_magic_bytes() {
+        assert_eq!(content_type(Path::new("/a/f.mp4"), &[]), "video/mp4");
+        assert_eq!(content_type(Path::new("/a/f.mp3"), &[]), "audio/mpeg");
+        // A `.mp4` that is really an AVI: downloading at `original` quality
+        // copies the source bytes under an mp4 name, so the extension lies and
+        // the magic bytes must win.
+        let avi_head = b"RIFF\xcc\xf3\xbc\x2bAVI LIST";
+        assert_eq!(
+            content_type(Path::new("/a/film.mp4"), avi_head),
+            "video/x-msvideo"
+        );
+        // Extension-less, as the offline queue writes them: sniff instead.
+        let mp4_head = b"\x00\x00\x00\x20ftypisom\x00\x00\x02\x00";
+        assert_eq!(content_type(Path::new("/a/abc123"), mp4_head), "video/mp4");
+        assert_eq!(
+            content_type(Path::new("/a/abc123"), b"ID3\x03junk"),
+            "audio/mpeg"
+        );
+    }
+}

src-tauri/src/playback_mode/mod.rs

@@ -27,6 +27,10 @@ const TICKS_PER_SECOND: f64 = 10_000_000.0;
 /// send a resume position, so a fresh track casts from 0 rather than ~0.
 const RESUME_THRESHOLD_SECONDS: f64 = 0.5;
 
+/// Volume level (0-100) the remote volume slider starts at. The real level is
+/// corrected by the session poller once the remote session reports its volume.
+const DEFAULT_REMOTE_VOLUME: i32 = 50;
+
 /// Convert a live playback position (seconds) into the `StartPositionTicks` to
 /// hand to a remote session, or `None` if we're effectively at the start.
 ///
@@ -42,6 +46,50 @@ fn start_position_ticks_from_seconds(position_seconds: f64) -> Option<i64> {
     }
 }
 
+/// Platform hook for attaching/detaching the OS remote-volume control.
+///
+/// On Android, entering remote mode hands the `MediaSession` a
+/// `VolumeProviderCompat` so hardware volume buttons and the system slider drive
+/// the *remote* session; leaving remote mode must hand it back to the local
+/// media stream. Behind a trait so the routing rule (see
+/// [`PlaybackModeManager::set_mode`]) is unit-testable off-device — the real
+/// implementation is JNI and only exists on Android.
+pub trait RemoteVolumeControl: Send + Sync {
+    /// Attach remote-volume control (and, on Android, start the playback service).
+    fn enable(&self, initial_volume: i32);
+    /// Return volume control to the local device speaker.
+    fn disable(&self);
+}
+
+/// Production hook: forwards to the Android JNI bridge; no-op elsewhere.
+struct PlatformRemoteVolumeControl;
+
+impl RemoteVolumeControl for PlatformRemoteVolumeControl {
+    #[allow(unused_variables)]
+    fn enable(&self, initial_volume: i32) {
+        #[cfg(target_os = "android")]
+        {
+            if let Err(e) = crate::player::enable_remote_volume(initial_volume) {
+                log::warn!(
+                    "[PlaybackMode] Failed to enable remote volume/service: {}",
+                    e
+                );
+                // Non-fatal - continue; the next poll tick will retry metadata.
+            }
+        }
+    }
+
+    fn disable(&self) {
+        #[cfg(target_os = "android")]
+        {
+            if let Err(e) = crate::player::disable_remote_volume() {
+                log::warn!("[PlaybackMode] Failed to disable remote volume: {}", e);
+                // Non-fatal - the mode change itself has already happened.
+            }
+        }
+    }
+}
+
 /// Manages playback mode transfers between local and remote sessions
 pub struct PlaybackModeManager {
     jellyfin_client: Arc<Mutex<Option<JellyfinClient>>>,
@@ -51,6 +99,8 @@ pub struct PlaybackModeManager {
     /// Optional emitter used to notify the frontend when the mode changes, so its
     /// mirror store stays in sync with this authoritative one. `None` in tests.
     event_emitter: Arc<Mutex<Option<Arc<dyn PlayerEventEmitter>>>>,
+    /// Platform hook for OS-level remote volume routing (swapped in tests).
+    remote_volume: Arc<dyn RemoteVolumeControl>,
 }
 
 impl PlaybackModeManager {
@@ -65,6 +115,24 @@ impl PlaybackModeManager {
             current_mode: Arc::new(RwLock::new(PlaybackMode::Idle)),
             is_transferring: Arc::new(AtomicBool::new(false)),
             event_emitter: Arc::new(Mutex::new(None)),
+            remote_volume: Arc::new(PlatformRemoteVolumeControl),
+        }
+    }
+
+    /// Construct with a custom remote-volume hook (tests).
+    #[cfg(test)]
+    fn with_remote_volume(
+        jellyfin_client: Arc<Mutex<Option<JellyfinClient>>>,
+        player_controller: Arc<TokioMutex<PlayerController>>,
+        remote_volume: Arc<dyn RemoteVolumeControl>,
+    ) -> Self {
+        Self {
+            jellyfin_client,
+            player_controller,
+            current_mode: Arc::new(RwLock::new(PlaybackMode::Idle)),
+            is_transferring: Arc::new(AtomicBool::new(false)),
+            event_emitter: Arc::new(Mutex::new(None)),
+            remote_volume,
         }
     }
 
@@ -86,19 +154,39 @@ impl PlaybackModeManager {
     /// the frontend's mirror store reconciles to this authoritative value. The
     /// write lock is released before emitting to avoid holding it across the
     /// emitter call.
+    ///
+    /// Also owns **OS volume routing**, which is derived from the transition
+    /// rather than from each call site: entering remote mode attaches the remote
+    /// volume control, and *any* exit from remote mode hands it back to the local
+    /// speaker. Doing this per-call-site is what caused the bug where stopping a
+    /// remote session (`player_stop` → Idle) left Android stuck on the remote
+    /// volume slider — only the transfer-to-local path tore it down.
+    ///
+    /// TRACES: UR-010 | DR-059, IR-021
     pub fn set_mode(&self, mode: PlaybackMode) {
         log::info!("[PlaybackMode] Setting mode to: {:?}", mode);
-        let changed = {
+        let (changed, was_remote) = {
             let mut current = self.current_mode.write_safe();
             let changed = *current != mode;
+            let was_remote = matches!(*current, PlaybackMode::Remote { .. });
             *current = mode.clone();
-            changed
+            (changed, was_remote)
         };
 
         if !changed {
             return;
         }
 
+        // Volume routing follows the transition. Note remote->remote (switching
+        // target session) re-arms rather than releasing control.
+        let is_remote = matches!(mode, PlaybackMode::Remote { .. });
+        if is_remote {
+            self.remote_volume.enable(DEFAULT_REMOTE_VOLUME);
+        } else if was_remote {
+            log::info!("[PlaybackMode] Leaving remote mode - restoring local volume control");
+            self.remote_volume.disable();
+        }
+
         let (mode_str, session_id) = match &mode {
             PlaybackMode::Local => ("local".to_string(), None),
             PlaybackMode::Idle => ("idle".to_string(), None),
@@ -122,18 +210,13 @@ impl PlaybackModeManager {
     /// Both symptoms share this one cause, so this must not be skipped on any
     /// remote-entry path (notably the empty-queue early return in
     /// `transfer_to_remote_inner`). No-op / non-Android builds do nothing.
-    #[allow(unused_variables)]
+    ///
+    /// [`set_mode`](Self::set_mode) already arms this on entry into remote mode;
+    /// calling it again is harmless (the service start is idempotent) and keeps
+    /// the guarantee when the mode was already remote, which `set_mode` skips as
+    /// a no-op transition.
     fn enable_remote_control(&self) {
-        #[cfg(target_os = "android")]
-        {
-            if let Err(e) = crate::player::enable_remote_volume(50) {
-                log::warn!(
-                    "[PlaybackMode] Failed to enable remote volume/service: {}",
-                    e
-                );
-                // Non-fatal - continue; the next poll tick will retry metadata.
-            }
-        }
+        self.remote_volume.enable(DEFAULT_REMOTE_VOLUME);
     }
 
     /// Check if currently transferring
@@ -766,18 +849,10 @@ impl PlaybackModeManager {
         // This will be improved in Phase 3 when repository is migrated to Rust.
         log::debug!("[PlaybackMode] Cannot load media item in Rust yet - frontend handled it");
 
-        // Update mode to local
+        // Update mode to local. This also returns volume control to the local
+        // device speaker — set_mode owns that for every exit from remote mode.
         self.set_mode(PlaybackMode::Local);
 
-        // Disable remote volume control on Android (return to system volume)
-        #[cfg(target_os = "android")]
-        {
-            if let Err(e) = crate::player::disable_remote_volume() {
-                log::warn!("[PlaybackMode] Failed to disable remote volume: {}", e);
-                // Non-fatal - continue with transfer
-            }
-        }
-
         log::info!("[PlaybackMode] Successfully transferred to local");
         Ok(())
     }
@@ -893,6 +968,118 @@ mod tests {
         }
     }
 
+    /// Records enable/disable calls so tests can assert volume routing.
+    struct RecordingVolumeControl {
+        calls: Mutex<Vec<&'static str>>,
+    }
+
+    impl RemoteVolumeControl for RecordingVolumeControl {
+        fn enable(&self, _initial_volume: i32) {
+            self.calls.lock().unwrap().push("enable");
+        }
+        fn disable(&self) {
+            self.calls.lock().unwrap().push("disable");
+        }
+    }
+
+    fn manager_with_volume_control() -> (PlaybackModeManager, Arc<RecordingVolumeControl>) {
+        let volume = Arc::new(RecordingVolumeControl {
+            calls: Mutex::new(Vec::new()),
+        });
+        let manager = PlaybackModeManager::with_remote_volume(
+            Arc::new(Mutex::new(None)),
+            Arc::new(TokioMutex::new(crate::player::PlayerController::default())),
+            volume.clone(),
+        );
+        (manager, volume)
+    }
+
+    /// Leaving remote mode must hand volume control back to the local device.
+    ///
+    /// Stopping a remote session (`player_stop`) drives the manager
+    /// Remote -> Idle without going through `transfer_to_local`. Before this was
+    /// centralised in `set_mode`, only the transfer path tore the Android
+    /// `VolumeProviderCompat` down, so a plain stop left the system stuck on the
+    /// remote volume slider with no way back to the phone speaker.
+    ///
+    /// @req-test: UR-010 - Control playback of Jellyfin remote sessions
+    #[test]
+    fn test_leaving_remote_mode_restores_local_volume() {
+        let (manager, volume) = manager_with_volume_control();
+
+        manager.set_mode(PlaybackMode::Remote {
+            session_id: "sess-1".to_string(),
+        });
+        // The stop path: remote -> idle, no transfer involved.
+        manager.set_mode(PlaybackMode::Idle);
+
+        assert_eq!(
+            *volume.calls.lock().unwrap(),
+            vec!["enable", "disable"],
+            "remote->idle must return volume control to the local speaker"
+        );
+    }
+
+    /// The same must hold for remote -> local (transfer back to this device).
+    ///
+    /// @req-test: UR-010 - Control playback of Jellyfin remote sessions
+    #[test]
+    fn test_remote_to_local_restores_local_volume() {
+        let (manager, volume) = manager_with_volume_control();
+
+        manager.set_mode(PlaybackMode::Remote {
+            session_id: "sess-1".to_string(),
+        });
+        manager.set_mode(PlaybackMode::Local);
+
+        assert_eq!(
+            *volume.calls.lock().unwrap(),
+            vec!["enable", "disable"],
+            "remote->local must return volume control to the local speaker"
+        );
+    }
+
+    /// Volume routing must not be touched by transitions that never involve
+    /// remote mode — an idle->local start would otherwise issue a pointless
+    /// `setPlaybackToLocal` on every playback start.
+    ///
+    /// @req-test: UR-010 - Control playback of Jellyfin remote sessions
+    #[test]
+    fn test_non_remote_transitions_leave_volume_routing_alone() {
+        let (manager, volume) = manager_with_volume_control();
+
+        manager.set_mode(PlaybackMode::Local);
+        manager.set_mode(PlaybackMode::Idle);
+        manager.set_mode(PlaybackMode::Local);
+
+        assert!(
+            volume.calls.lock().unwrap().is_empty(),
+            "local/idle transitions must not touch remote volume routing"
+        );
+    }
+
+    /// Switching directly between two remote sessions stays remote: control must
+    /// remain attached (re-armed for the new session), never handed back local.
+    ///
+    /// @req-test: UR-010 - Control playback of Jellyfin remote sessions
+    #[test]
+    fn test_remote_to_remote_keeps_remote_volume() {
+        let (manager, volume) = manager_with_volume_control();
+
+        manager.set_mode(PlaybackMode::Remote {
+            session_id: "sess-1".to_string(),
+        });
+        manager.set_mode(PlaybackMode::Remote {
+            session_id: "sess-2".to_string(),
+        });
+
+        assert_eq!(
+            *volume.calls.lock().unwrap(),
+            vec!["enable", "enable"],
+            "remote->remote re-arms control without releasing it to local"
+        );
+    }
+
     /// Setting the same mode twice must not re-emit — the frontend reconciler
     /// (and the event channel) shouldn't be spammed on no-op transitions.
     #[test]

src-tauri/src/player/android/mod.rs

@@ -18,6 +18,7 @@ use super::events::{PlayerStatusEvent, SharedEventEmitter};
 use super::media::{MediaItem, MediaType};
 use super::state::PlayerState;
 use crate::playback_reporting::{EventThrottler, PlaybackOperation, PlaybackReporter};
+use crate::settings::{audio_settings_jni_payload, AudioSettings};
 use crate::utils::conversions::seconds_to_ticks;
 
 /// Global reference to the JavaVM for JNI callbacks
@@ -57,11 +58,20 @@ static POSITION_THROTTLER: OnceLock<Arc<EventThrottler>> = OnceLock::new();
 struct DetectedCodecs {
     video_codecs: Vec<String>,
     audio_codecs: Vec<String>,
+    /// Channels the *current audio output route* accepts, as reported by
+    /// media3's `AudioCapabilities`. Distinct from the codec lists: a device
+    /// decodes 5.1 happily and still has only two channels to play it out of.
+    /// `None` when the platform had no answer.
+    max_audio_channels: Option<u32>,
 }
 
 impl DetectedCodecs {
     /// Create from comma-separated codec strings (from JNI)
-    fn from_jni_strings(video_codecs: &str, audio_codecs: &str) -> Self {
+    fn from_jni_strings(
+        video_codecs: &str,
+        audio_codecs: &str,
+        max_audio_channels: Option<u32>,
+    ) -> Self {
         Self {
             video_codecs: video_codecs
                 .split(',')
@@ -73,6 +83,7 @@ impl DetectedCodecs {
                 .filter(|s| !s.is_empty())
                 .map(|s| s.to_string())
                 .collect(),
+            max_audio_channels,
         }
     }
 
@@ -87,11 +98,19 @@ impl DetectedCodecs {
     }
 }
 
-/// Public function to get detected codecs (for use in repository layer)
-pub fn get_detected_codecs() -> Option<(String, String)> {
-    DETECTED_CODECS
-        .get()
-        .map(|codecs| (codecs.video_codecs_string(), codecs.audio_codecs_string()))
+/// Public function to get detected codecs (for use in repository layer).
+///
+/// Returns `(video, audio, max_audio_channels)` — the third element is how many
+/// channels the current audio output can actually voice, which bounds what the
+/// server may direct-play.
+pub fn get_detected_codecs() -> Option<(String, String, Option<u32>)> {
+    DETECTED_CODECS.get().map(|codecs| {
+        (
+            codecs.video_codecs_string(),
+            codecs.audio_codecs_string(),
+            codecs.max_audio_channels,
+        )
+    })
 }
 
 /// Trait for handling media commands from Android MediaSession.
@@ -148,6 +167,10 @@ struct ExoPlayerState {
     volume: f32,
     is_loaded: bool,
     current_media: Option<MediaItem>,
+    /// Last applied audio settings. Unlike the fields above (which JNI callbacks
+    /// push *in*), this is commanded *out* — audio settings are never reported
+    /// by the player, so this is the authoritative copy for `audio_settings()`.
+    audio_settings: AudioSettings,
 }
 
 impl ExoPlayerState {
@@ -159,6 +182,7 @@ impl ExoPlayerState {
             volume: 1.0,
             is_loaded: false,
             current_media: None,
+            audio_settings: AudioSettings::default(),
         }
     }
 }
@@ -529,6 +553,56 @@ impl PlayerBackend for ExoPlayerBackend {
         self.shared_state.lock_safe().volume
     }
 
+    /// Apply audio settings to ExoPlayer (equalizer, normalization, gapless).
+    ///
+    /// Sent as JSON rather than a wide JNI signature so new fields do not change
+    /// the method signature — the same approach `load()` uses for subtitles. The
+    /// Kotlin side owns the *mechanics* (attaching AudioEffects to the audio
+    /// session); the canonical band layout and preset curves stay in Rust.
+    ///
+    /// TRACES: UR-027, UR-032, UR-033 | DR-030, DR-035, DR-036
+    fn set_audio_settings(&mut self, settings: &AudioSettings) -> Result<(), PlayerError> {
+        let json = audio_settings_jni_payload(settings).map_err(|e| {
+            PlayerError::playback_failed(format!("Failed to serialize audio settings: {}", e))
+        })?;
+
+        let vm = JAVA_VM
+            .get()
+            .ok_or_else(|| PlayerError::playback_failed("JavaVM not initialized"))?;
+
+        let mut env = vm
+            .attach_current_thread()
+            .map_err(|e| PlayerError::playback_failed(format!("Failed to attach thread: {}", e)))?;
+
+        let json_jstring = env.new_string(&json).map_err(|e| {
+            PlayerError::playback_failed(format!("Failed to create settings string: {}", e))
+        })?;
+
+        env.call_method(
+            &self.player_ref,
+            "setAudioSettings",
+            "(Ljava/lang/String;)V",
+            &[JValue::Object(&json_jstring)],
+        )
+        .map_err(|e| {
+            PlayerError::playback_failed(format!("Failed to call setAudioSettings: {}", e))
+        })?;
+
+        // Store the sanitised form so audio_settings() reflects what was applied,
+        // not what was requested.
+        self.shared_state.lock_safe().audio_settings = settings
+            .clone()
+            .with_crossfade_clamped()
+            .with_equalizer_normalised();
+
+        Ok(())
+    }
+
+    /// TRACES: UR-027, UR-032, UR-033 | DR-030, DR-035, DR-036
+    fn audio_settings(&self) -> AudioSettings {
+        self.shared_state.lock_safe().audio_settings.clone()
+    }
+
     fn set_audio_track(&mut self, stream_index: i32) -> Result<(), PlayerError> {
         let vm = JAVA_VM
             .get()
@@ -858,12 +932,39 @@ pub extern "system" fn Java_com_dtourolle_jellytau_player_JellyTauPlayer_nativeO
                         });
                     }
 
-                    // Start countdown if auto_advance enabled
                     if auto_advance {
+                        // Shared with the frontend-invoked command path
+                        // (player_on_playback_ended) so the two dispatchers cannot
+                        // disagree about how a background audio-only episode
+                        // advances — they did, and the command's copy was missing
+                        // the case entirely. That copy is the one that actually
+                        // decides here: the end reason set at load makes this
+                        // callback's own decision Stop, and the frontend echoes the
+                        // resulting PlaybackEnded back into the command.
                         controller
                             .lock()
                             .await
-                            .start_autoplay_countdown(next_episode, countdown_seconds);
+                            .auto_advance_to_next_episode(next_episode, countdown_seconds)
+                            .await;
+                    }
+                }
+                Ok(AutoplayDecision::ResumeStream { position }) => {
+                    // ExoPlayer reported ENDED because the progressive transcode's
+                    // connection dropped, not because the episode finished. This
+                    // is the arm that matters while backgrounded: it needs no
+                    // frontend echo, so the stream re-opens even with the webview
+                    // suspended — and playback never parks in STATE_ENDED, where
+                    // the next lockscreen/Bluetooth play restarts the item at 0:00.
+                    log::info!(
+                        "[Autoplay] Decision: Resume truncated stream at {:.1}s",
+                        position
+                    );
+                    let ctrl = controller.lock().await;
+                    if let Err(e) = ctrl.resume_stream_at(position).await {
+                        log::error!("[Autoplay] Failed to resume truncated stream: {}", e);
+                        if let Some(emitter) = EVENT_EMITTER.get() {
+                            emitter.emit(PlayerStatusEvent::PlaybackEnded);
+                        }
                     }
                 }
                 Err(e) => {
@@ -910,11 +1011,61 @@ pub extern "system" fn Java_com_dtourolle_jellytau_player_JellyTauPlayer_nativeO
         .get_string(&message)
         .map(|s| s.into())
         .unwrap_or_else(|_| "Unknown error".to_string());
+    let recoverable = recoverable != 0;
+
+    // A background audio-only handoff is an mp3 the device was already decoding,
+    // so a recoverable failure part-way through is the network. Surfacing it as a
+    // player error stops playback for good (the frontend's handler calls
+    // player_stop); re-opening the stream where it died is the "buffer and
+    // resume" this actually is. Everything else keeps reporting the error.
+    if recoverable {
+        if let Some(controller) = PLAYER_CONTROLLER.get() {
+            let controller = controller.clone();
+            let message_str = message_str.clone();
+            tauri::async_runtime::spawn(async move {
+                let resume = controller.lock().await.recoverable_error_resume();
+                let Some((position, delay_secs)) = resume else {
+                    // Declined here, so report it as NOT recoverable: the frontend
+                    // would otherwise echo it into player_recover_stream and ask
+                    // the same question a second time.
+                    if let Some(emitter) = EVENT_EMITTER.get() {
+                        emitter.emit(PlayerStatusEvent::Error {
+                            message: message_str,
+                            recoverable: false,
+                        });
+                    }
+                    return;
+                };
+
+                log::warn!(
+                    "[ExoPlayer] Recoverable stream error ({}) — re-opening at {:.1}s in {}s",
+                    message_str,
+                    position,
+                    delay_secs
+                );
+                // Give a brief outage time to clear before asking the server for
+                // the stream again; retrying instantly just burns the budget.
+                tokio::time::sleep(std::time::Duration::from_secs(delay_secs)).await;
+
+                let ctrl = controller.lock().await;
+                if let Err(e) = ctrl.resume_stream_at(position).await {
+                    log::error!("[ExoPlayer] Failed to resume after error: {}", e);
+                    if let Some(emitter) = EVENT_EMITTER.get() {
+                        emitter.emit(PlayerStatusEvent::Error {
+                            message: message_str,
+                            recoverable: false,
+                        });
+                    }
+                }
+            });
+            return;
+        }
+    }
 
     if let Some(emitter) = EVENT_EMITTER.get() {
         emitter.emit(PlayerStatusEvent::Error {
             message: message_str,
-            recoverable: recoverable != 0,
+            recoverable,
         });
     }
 }
@@ -992,6 +1143,7 @@ pub extern "system" fn Java_com_dtourolle_jellytau_player_JellyTauPlayer_00024Co
     _class: JClass,
     video_codecs: JString,
     audio_codecs: JString,
+    max_audio_channels: jint,
 ) {
     let video_str: String = env
         .get_string(&video_codecs)
@@ -1003,7 +1155,10 @@ pub extern "system" fn Java_com_dtourolle_jellytau_player_JellyTauPlayer_00024Co
         .map(|s| s.into())
         .unwrap_or_default();
 
-    let codecs = DetectedCodecs::from_jni_strings(&video_str, &audio_str);
+    // Kotlin sends 0 when AudioCapabilities had no answer for the current route.
+    let channels = u32::try_from(max_audio_channels).ok().filter(|c| *c > 0);
+
+    let codecs = DetectedCodecs::from_jni_strings(&video_str, &audio_str, channels);
 
     log::info!(
         "[CodecDetection] Detected {} video codecs: {}",
@@ -1015,6 +1170,10 @@ pub extern "system" fn Java_com_dtourolle_jellytau_player_JellyTauPlayer_00024Co
         codecs.audio_codecs.len(),
         codecs.audio_codecs_string()
     );
+    log::info!(
+        "[CodecDetection] Audio route max channels: {:?}",
+        codecs.max_audio_channels
+    );
 
     // Store in global state
     if DETECTED_CODECS.set(codecs).is_err() {
@@ -1315,9 +1474,11 @@ pub fn update_lockscreen_metadata(meta: &LockscreenMetadata) -> Result<(), Strin
     Ok(())
 }
 
-/// Set the base position offset (seconds) on the lockscreen MediaSession.
+/// Set the background-audio handoff base (seconds) on the playback service.
 ///
-/// Calls `JellyTauPlaybackService.setPositionOffset(double)`. No-op if the
+/// The service holds it for `JellyTauPlayer`'s position tick, which is the one
+/// place the relative handoff timeline is converted to the episode's own — see
+/// DR-159. Calls `JellyTauPlaybackService.setHandoffBase(double)`. No-op if the
 /// service isn't running yet, so it's safe to call unconditionally.
 pub fn set_position_offset(offset_seconds: f64) -> Result<(), String> {
     let vm = JAVA_VM.get().ok_or("JavaVM not initialized")?;
@@ -1366,11 +1527,11 @@ pub fn set_position_offset(offset_seconds: f64) -> Result<(), String> {
 
     env.call_method(
         &service_obj,
-        "setPositionOffset",
+        "setHandoffBase",
         "(D)V",
         &[JValue::Double(offset_seconds)],
     )
-    .map_err(|e| format!("Failed to set position offset: {}", e))?;
+    .map_err(|e| format!("Failed to set handoff base: {}", e))?;
 
     Ok(())
 }

src-tauri/src/player/autoplay.rs

@@ -11,6 +11,10 @@ pub enum AutoplayDecision {
     Stop,
     /// Advance to next track in queue (for audio/movies)
     AdvanceToNext,
+    /// The stream ended well short of the item's runtime — the connection
+    /// dropped, not the media. Re-open the same stream at `position` instead of
+    /// running any end-of-item logic (UR-040).
+    ResumeStream { position: f64 },
     /// Show next episode popup with countdown
     ShowNextEpisodePopup {
         current_episode: MediaItem,

src-tauri/src/player/backend.rs

@@ -332,6 +332,7 @@ mod tests {
             gapless_playback: false,
             normalize_volume: true,
             volume_level: VolumeLevel::Loud,
+            ..Default::default()
         };
 
         backend.set_audio_settings(&settings).unwrap();

src-tauri/src/player/events.rs

@@ -156,6 +156,25 @@ pub enum PlayerStatusEvent {
         /// Target position in seconds (only meaningful for "seek").
         position: Option<f64>,
     },
+    /// Ask the frontend webview `<audio>` element to load and play a stream.
+    ///
+    /// Emitted by `WebviewAudioBackend` on platforms with no native audio
+    /// backend (e.g. Windows): audio-only playback is rendered by an `<audio>`
+    /// element in the webview, mirroring how all video already renders through
+    /// the webview `<video>`. The element then reports its state/position back
+    /// through the `player_report_*` commands, so the Rust controller stays the
+    /// single source of truth. Subsequent play/pause/seek/stop reach the element
+    /// via `ControlCommand`.
+    WebviewAudioLoad {
+        /// Stream URL for the `<audio>` element to play.
+        url: String,
+        /// Jellyfin item id, used as the media_id when reporting state back.
+        media_id: Option<String>,
+        /// Resume position in seconds (0 = start from the beginning).
+        position: f64,
+        /// Whether to begin playing immediately after loading.
+        autoplay: bool,
+    },
 }
 
 /// Trait for emitting player events to the frontend.

src-tauri/src/player/media.rs

@@ -23,6 +23,23 @@ pub enum QueueContext {
 }
 
 /// Represents a subtitle track
+///
+/// 🔴 **Do not add `#[serde(rename_all = "camelCase")]` here.** This is the one
+/// struct in the player that deliberately keeps snake_case on the wire, because
+/// the *same* serialization feeds two consumers that both spell `mime_type`:
+///
+/// * the JNI boundary — `player/android/mod.rs` serializes `MediaItem::subtitles`
+///   with `serde_json` and hands the string to `JellyTauPlayer.loadWithMetadata`,
+///   whose parser reads `url`, `language`, `label` and `optString("mime_type")`;
+/// * the IPC boundary — `PlayItemRequest::subtitles` deserializes this same type
+///   from the frontend, and the generated binding (`SubtitleTrack` in
+///   `bindings.ts`) therefore also declares `mime_type`.
+///
+/// Renaming would not break the build and would not fail the IPC: Kotlin's
+/// `optString` would just fall back to its default MIME type for every track, so
+/// the failure would be silent. UT-146 asserts the serialized keys.
+///
+/// TRACES: UR-020 | IR-016, JA-008 | UT-146
 #[derive(specta::Type, Debug, Clone, Serialize, Deserialize, PartialEq)]
 pub struct SubtitleTrack {
     /// Stream index in the media source
@@ -33,7 +50,8 @@ pub struct SubtitleTrack {
     pub language: Option<String>,
     /// Display title
     pub label: Option<String>,
-    /// MIME type (e.g., "text/vtt", "application/x-subrip")
+    /// MIME type (e.g., "text/vtt", "application/x-subrip").
+    /// Snake_case on purpose — see the note on the struct.
     pub mime_type: String,
 }