refactor: delete two orphans, and record why the reparent design changed

`resolveVideoSource` chose between a local file and a remote URL for video
playback. Backend-owned stream selection took that decision into Rust —
`media_local_selection` for a downloaded file, `get_stream_selection` for a
streamed one — and its last caller went with it. What remained was the function
plus sixty lines of tests exercising nothing that ships.

`fittedVideoSize` computed the rendered size of a video letterboxed into its
container. Nothing has ever called it: it arrived with the fix that made the
video fill its viewport and was superseded by `object-fit: contain` in the same
change. There is some irony in a helper that models letterboxing sitting unused
beside a container that was not letterboxing at all — the bug fixed in the
previous commit was CSS, and this function would not have helped.

A survey for exported symbols referenced only by their own tests finds 22 more.
Most are legitimate — test mocks, deliberate reset hooks, public utility APIs —
and the rest are unrelated to this work, so they are left for a cleanup that can
be reviewed on its own terms rather than smuggled into a playback branch.

Also records in the spec why DR-231's design changed. Reparenting Tauri's
webview into a GtkOverlay aborts the process on the first click: Linux calls
`attach_resize_handler` unconditionally (the Windows path guards it with
`is_decorated()`), and its handler walks webview -> GtkBox -> GtkWindow with an
unwrap that an overlay breaks. So the webview is not moved at all — mpv draws
into the default vbox's own `draw` handler via `gdk_cairo_draw_from_gl()`, and
GTK's container-before-children order puts the webview on top for free. No
reparent, one less widget, and nothing a Tauri upgrade can invalidate by
assuming its own layout.

.env.example

@@ -0,0 +1,23 @@
+# Local Android release signing.
+#
+# Copy to `.env` and fill in. `.env` is gitignored and is the single source of
+# truth for local release signing — scripts/write-keystore-properties.sh reads
+# it and regenerates src-tauri/gen/android/keystore.properties before every
+# release build, because `tauri android init` overwrites that file.
+#
+# Only needed for `bun run android:build:release`. Debug builds sign with the
+# local debug keystore and need nothing here.
+#
+# CI does not use this file: build-release.yml reconstructs the keystore from
+# the ANDROID_KEYSTORE_BASE64 secret and writes the same properties itself.
+
+# Key alias inside the keystore.
+ANDROID_KEY_ALIAS=jellytau
+
+# Absolute path to the .jks. Keep it outside the repo, or in the gitignored
+# android-keystore/ directory.
+ANDROID_KEYSTORE_FILE=/absolute/path/to/jellytau-release.jks
+
+# Keystore and key passwords. These are secrets — never commit the filled-in .env.
+ANDROID_KEYSTORE_PASSWORD=
+ANDROID_KEY_PASSWORD=

.gitea/ISSUE_TEMPLATE/bug.yaml

@@ -0,0 +1,103 @@
+name: Bug report
+about: Something behaves incorrectly
+title: ""
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Security vulnerabilities do **not** go here — see
+        [SECURITY.md](../../SECURITY.md).
+
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened
+      description: What you did, what you expected, and what you got instead.
+      placeholder: |
+        1. Opened an album from the Music library
+        2. Tapped the third track
+        3. Playback started from the first track instead
+    validations:
+      required: true
+
+  - type: input
+    id: version
+    attributes:
+      label: JellyTau version
+      description: Settings scrolls to the bottom, or the filename you installed.
+      placeholder: "0.9.1"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: platform
+    attributes:
+      label: Platform
+      options:
+        - Linux (AppImage)
+        - Linux (deb)
+        - Linux (rpm)
+        - Linux (Arch package)
+        - Windows
+        - Android
+    validations:
+      required: true
+
+  - type: markdown
+    attributes:
+      value: |
+        ### Playback questions
+
+        If this involves playback, these three answers decide which of several
+        very different code paths you were on. "I don't know" is a fine answer.
+
+  - type: dropdown
+    id: source
+    attributes:
+      label: Was the media streaming or downloaded?
+      options:
+        - Streaming from the server
+        - Downloaded for offline use
+        - Not playback-related
+    validations:
+      required: true
+
+  - type: dropdown
+    id: transcode
+    attributes:
+      label: Was the server transcoding?
+      description: Jellyfin's dashboard shows this while something is playing.
+      options:
+        - Direct play
+        - Transcoding
+        - Don't know
+        - Not playback-related
+
+  - type: dropdown
+    id: kind
+    attributes:
+      label: Music or video?
+      options:
+        - Music
+        - Video (movie)
+        - Video (TV episode)
+        - Not playback-related
+
+  - type: textarea
+    id: logs
+    attributes:
+      label: Logs
+      description: |
+        Android: `adb logcat | grep -i jellytau`.
+        Linux: run from a terminal, or `RUST_LOG=debug jellytau` for more.
+        In the app, `localStorage.setItem("jellytau:logLevel","debug")` in the
+        webview console turns the frontend up too.
+      render: shell
+
+  - type: textarea
+    id: server
+    attributes:
+      label: Jellyfin server
+      description: Version, and anything unusual about the library layout.
+      placeholder: "10.9.11, series stored without season folders"

.gitea/ISSUE_TEMPLATE/feature.yaml

@@ -0,0 +1,37 @@
+name: Feature request
+about: Suggest something JellyTau should do
+title: ""
+labels: ["enhancement"]
+body:
+  - type: textarea
+    id: problem
+    attributes:
+      label: What are you trying to do?
+      description: |
+        The situation, not the solution. "I listen to albums in a fixed order and
+        lose my place when I switch devices" tells us more than "add a sync
+        button", and often has a better answer than the one you had in mind.
+    validations:
+      required: true
+
+  - type: textarea
+    id: proposal
+    attributes:
+      label: What would you like it to do?
+    validations:
+      required: true
+
+  - type: dropdown
+    id: platform
+    attributes:
+      label: Which platforms does this matter on?
+      multiple: true
+      options:
+        - Linux
+        - Windows
+        - Android
+
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Anything you have tried, or how other clients handle it

.gitea/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,22 @@
+## What and why
+
+<!-- What changes, and the reason. The diff shows the what; the why is what
+     the commit log is for. -->
+
+## How it was verified
+
+<!-- What you actually ran or clicked. "Tests pass" on its own says little;
+     "played a transcoded episode on Android, seeked twice, backgrounded it"
+     says a lot. -->
+
+## Checklist
+
+- [ ] `bun run check`, `bun run test`, `bun run format:check`, `bun run lint`
+- [ ] `cargo fmt`, `cargo clippy --all-targets -- -D warnings`, `cargo test`
+- [ ] `bun run check:boundary` — no Jellyfin taxonomy in the frontend
+- [ ] New requirement-implementing code carries a `TRACES:` comment, and every
+      ID it names exists in `docs/requirements.md` (`bun run traces:validate`)
+- [ ] **Bug fix:** a test that reproduces it was written *first* and observed
+      failing before the fix
+- [ ] Android source edits were made in `src-tauri/android/src` and synced with
+      `scripts/sync-android-sources.sh` (never edit `gen/` directly)

.gitea/workflows/build-and-test.yml

@@ -13,12 +13,22 @@ on:
       - '**/*.md'
   workflow_dispatch:
 
+env:
+  # Incremental state is never reused between CI runs -- pure disk cost.
+  CARGO_INCREMENTAL: 0
+
 jobs:
   test:
     name: Run Tests
+    # A release push triggers build-release.yml on the tag, which runs this exact
+    # test suite itself — and on a single-slot runner the two ~1h workflows would
+    # otherwise serialize/contend. Skip the duplicate for chore(release) commits.
+    # (head_commit is absent on pull_request/workflow_dispatch; startsWith(null,…)
+    # is false there, so those events still run.)
+    if: "!startsWith(github.event.head_commit.message, 'chore(release)')"
     runs-on: linux/amd64
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
 
     steps:
       - name: Checkout repository
@@ -27,13 +37,22 @@ jobs:
       - name: Cache Rust dependencies
         uses: actions/cache@v3
         with:
+          # Registry only -- never src-tauri/target. That directory is ~16 GB and
+          # was cached under five separate keys, which filled the runner's 74 GB
+          # disk at ~1.15 GB/day (23 GB in 20 days, measured Aug 2026).
+          # registry/src is omitted too: cargo re-extracts it for free from
+          # registry/cache (155 MB of .crate tarballs vs 1.1 GB extracted).
           path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            src-tauri/target
-          key: ${{ runner.os }}-cargo-host-${{ hashFiles('**/Cargo.lock') }}
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          # One shared key across every job. The old per-job keys existed to stop
+          # debug/release target artifacts clobbering each other; with target no
+          # longer cached, registry contents are target-independent, so all jobs
+          # want the same crates. First job to finish saves; the rest restore.
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-host-
+            ${{ runner.os }}-cargo-registry-
 
       - name: Cache Node dependencies
         uses: actions/cache@v3
@@ -56,10 +75,99 @@ jobs:
       - name: Check frontend/backend boundary
         run: bash scripts/check-frontend-boundary.sh
 
+      # The docs are the maintained source of truth for architecture and
+      # process, and they cross-reference each other heavily. A rename that
+      # misses a link turns a doc into a dead end silently. Pure shell + git —
+      # no tool is installed at job time.
+      - name: Check documentation links
+        run: bash scripts/check-doc-links.sh
+
+      # Formatting, linting and type-checking were all configured in this repo
+      # and enforced by nothing: .prettierrc described a tree where 199 files did
+      # not match it, eslint.config.js ran in no workflow and in no hook, and
+      # `bun run check` ran only in build-release.yml — i.e. a type error could
+      # sit on master until somebody cut a tag. These three steps are what make
+      # those configs load-bearing. All are project deps installed by
+      # `bun install`; nothing is fetched at job time.
+      # Cheap tripwire for a class of defect this repo kept hitting: tooling on
+      # a rarely-taken path. scripts/build-android.sh ran `npm install` on its
+      # clean-build branch -- in a bun project, ignoring bun.lock and
+      # re-resolving the tree, which is how the Tauri plugin crate/package
+      # versions drifted apart and broke a release build. It survived because
+      # clean builds are rare.
+      - name: Check build tooling
+        run: bash scripts/check-tooling.sh
+
+      - name: Check formatting
+        run: bun run format:check
+
+      # RATCHET — this number only ever goes DOWN. Same policy as MIN_THRESHOLD
+      # in traceability-check.yml and the coverage thresholds in
+      # vitest.config.ts. 159 is what the tree carried when the gate went in; the
+      # backlog is real findings (dead bindings, unkeyed {#each}, `any` at the
+      # IPC boundary) that eslint.config.js documents rule by rule, each parked
+      # at "warn" until its class is cleared and it can be promoted to "error".
+      # Lower this as you clear them. Never raise it to make a build pass.
+      - name: Lint
+        run: bun run lint -- --max-warnings=158
+
+      - name: Check TypeScript
+        run: |
+          bunx svelte-kit sync
+          bun run check
+
+      # Tauri refuses to build when a plugin's Rust crate and npm package are on
+      # different minor versions. Nothing here runs `tauri build` -- that only
+      # happens on a tag -- so a mismatch introduced on master stayed invisible
+      # until the release build, which is where it was found: v0.10.0 prep hit
+      # `tauri-plugin-log (v2.8.0) : @tauri-apps/plugin-log (v2.9.0)`. `cargo
+      # check`, clippy, the tests and svelte-check had all passed.
+      #
+      # `tauri info` performs the same comparison the bundler does, without a
+      # build. Grepping its output is crude, but the alternative is discovering
+      # this at tag time again.
+      - name: Check Tauri plugin versions match
+        run: |
+          set -e
+          if bunx tauri info 2>&1 | tee /tmp/tauri-info.txt | grep -q "version mismatched"; then
+            echo "::error::A Tauri plugin's Rust crate and npm package versions disagree."
+            echo "::error::The release build will refuse to start. Align them in"
+            echo "::error::src-tauri/Cargo.toml and package.json (both are pinned exactly)."
+            grep -A6 "version mismatched" /tmp/tauri-info.txt || true
+            exit 1
+          fi
+          echo "✅ Tauri plugin crate/package versions agree."
+
+      # Coverage rather than a bare `bun run test`: same suite, plus the
+      # thresholds in vitest.config.ts, so a large untested module or a deleted
+      # test fails here instead of being noticed months later.
       - name: Run frontend tests
         run: |
           bunx svelte-kit sync
-          bun run test
+          bun run test:coverage
+
+      # CLAUDE.md has required `cargo fmt` + `cargo clippy` before every commit
+      # for as long as the rule has existed, but nothing in CI checked either,
+      # so the requirement rested entirely on memory. Both components are baked
+      # into the builder image (Dockerfile.builder: `rustup component add
+      # rustfmt clippy`) — nothing is installed at job time.
+      - name: Check Rust formatting
+        run: |
+          cd src-tauri
+          cargo fmt --all -- --check
+
+      # Clippy is a hard gate. It was advisory while the tree carried a warning
+      # backlog; that backlog is gone (0 warnings on 1.97.1, the pinned
+      # toolchain), so a warning here is now new breakage rather than old noise.
+      #
+      # This only means anything because src-tauri/rust-toolchain.toml pins the
+      # compiler: clippy's lint set moves between releases, so an unpinned gate
+      # would fail on whatever the runner happened to install. The pin and this
+      # flag stand or fall together — if you unpin, drop this back to advisory.
+      - name: Run clippy
+        run: |
+          cd src-tauri
+          cargo clippy --all-targets -- -D warnings
 
       - name: Run Rust tests
         run: |
@@ -79,7 +187,7 @@ jobs:
     runs-on: linux/amd64
     needs: test
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
     env:
       ANDROID_HOME: /opt/android-sdk
       ANDROID_SDK_ROOT: /opt/android-sdk
@@ -93,13 +201,22 @@ jobs:
       - name: Cache Rust dependencies
         uses: actions/cache@v3
         with:
+          # Registry only -- never src-tauri/target. That directory is ~16 GB and
+          # was cached under five separate keys, which filled the runner's 74 GB
+          # disk at ~1.15 GB/day (23 GB in 20 days, measured Aug 2026).
+          # registry/src is omitted too: cargo re-extracts it for free from
+          # registry/cache (155 MB of .crate tarballs vs 1.1 GB extracted).
           path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            src-tauri/target
-          key: ${{ runner.os }}-cargo-android-${{ hashFiles('**/Cargo.lock') }}
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          # One shared key across every job. The old per-job keys existed to stop
+          # debug/release target artifacts clobbering each other; with target no
+          # longer cached, registry contents are target-independent, so all jobs
+          # want the same crates. First job to finish saves; the rest restore.
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-android-
+            ${{ runner.os }}-cargo-registry-
 
       - name: Cache Node dependencies
         uses: actions/cache@v3
@@ -122,3 +239,60 @@ jobs:
           export AR_aarch64_linux_android="$TC/llvm-ar"
           cd src-tauri
           cargo check --target aarch64-linux-android --lib
+
+  # Supply-chain gate. Until this job existed the project had no vulnerability
+  # scanning of any kind: nothing checked the ~500-crate Rust graph or the JS
+  # dependencies against a CVE feed, and nothing checked that everything we
+  # redistribute is licence-compatible with shipping JellyTau under MIT.
+  #
+  # The first run of this found eight vulnerabilities and one unsoundness
+  # (bytes, four in rustls-webpki, time, two in quick-xml, rand) — all fixed by
+  # `cargo update`, none of which anybody had reason to run.
+  #
+  # Runs in parallel with android-check rather than after `test`: a dependency
+  # advisory has nothing to do with whether the tests pass, and finding out
+  # sooner is the point.
+  security:
+    name: Supply Chain
+    runs-on: linux/amd64
+    container:
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Cache Rust dependencies
+        uses: actions/cache@v3
+        with:
+          path: |
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
+          restore-keys: |
+            ${{ runner.os }}-cargo-registry-
+
+      # cargo-deny is baked into the builder image. It fetches the RustSec
+      # advisory database at run time — that is *data*, like the crates
+      # `bun install` fetches, not a toolchain install, so the 🔴 rule in
+      # CLAUDE.md is not in play here.
+      #
+      # Config and every documented exception live in src-tauri/deny.toml.
+      # Vulnerabilities and unsoundness are hard failures with no override;
+      # unmaintained transitive crates that have no safe upgrade (Tauri's GTK3
+      # stack, the unic-* tables) are ignored there by ID, each with a reason.
+      - name: cargo-deny (advisories, licences, bans, sources)
+        run: |
+          cd src-tauri
+          cargo deny check
+
+      # Advisory for now, deliberately. The Rust graph was clean after one
+      # update pass, so gating it costs nothing; the JS graph has not been
+      # audited before and a first run that fails the build teaches everyone to
+      # ignore this job. Promote to a hard gate once the output is empty and
+      # stays empty — same approach that got clippy from advisory to -D warnings.
+      - name: bun audit (advisory)
+        run: |
+          bun install
+          bun audit || echo "::warning::bun audit reported findings — advisory for now, see CLAUDE.md"

.gitea/workflows/build-release.yml

@@ -13,13 +13,15 @@ on:
 env:
   RUST_BACKTRACE: 1
   CARGO_TERM_COLOR: always
+  # Incremental state is never reused between CI runs -- pure disk cost.
+  CARGO_INCREMENTAL: 0
 
 jobs:
   test:
     name: Run Tests
     runs-on: linux/amd64
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -27,13 +29,22 @@ jobs:
       - name: Cache Rust dependencies
         uses: actions/cache@v3
         with:
+          # Registry only -- never src-tauri/target. That directory is ~16 GB and
+          # was cached under five separate keys, which filled the runner's 74 GB
+          # disk at ~1.15 GB/day (23 GB in 20 days, measured Aug 2026).
+          # registry/src is omitted too: cargo re-extracts it for free from
+          # registry/cache (155 MB of .crate tarballs vs 1.1 GB extracted).
           path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            src-tauri/target
-          key: ${{ runner.os }}-cargo-host-${{ hashFiles('**/Cargo.lock') }}
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          # One shared key across every job. The old per-job keys existed to stop
+          # debug/release target artifacts clobbering each other; with target no
+          # longer cached, registry contents are target-independent, so all jobs
+          # want the same crates. First job to finish saves; the rest restore.
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-host-
+            ${{ runner.os }}-cargo-registry-
 
       - name: Cache Node dependencies
         uses: actions/cache@v3
@@ -54,6 +65,22 @@ jobs:
           bun run test --run
         continue-on-error: false
 
+      # Same gate as build-and-test.yml. A release must not ship from a tree
+      # that would fail the per-commit checks. rustfmt/clippy come from the
+      # builder image; nothing is installed here.
+      - name: Check Rust formatting
+        run: |
+          cd src-tauri
+          cargo fmt --all -- --check
+        continue-on-error: false
+
+      # Advisory until the ~51 pre-existing warnings are cleared; see the longer
+      # note in build-and-test.yml. Tighten both to `-- -D warnings` together.
+      - name: Run clippy (advisory)
+        run: |
+          cd src-tauri
+          cargo clippy --all-targets
+
       - name: Run Rust tests
         run: bun run test:rust
         continue-on-error: false
@@ -67,7 +94,7 @@ jobs:
     runs-on: linux/amd64
     needs: test
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -75,13 +102,22 @@ jobs:
       - name: Cache Rust dependencies
         uses: actions/cache@v3
         with:
+          # Registry only -- never src-tauri/target. That directory is ~16 GB and
+          # was cached under five separate keys, which filled the runner's 74 GB
+          # disk at ~1.15 GB/day (23 GB in 20 days, measured Aug 2026).
+          # registry/src is omitted too: cargo re-extracts it for free from
+          # registry/cache (155 MB of .crate tarballs vs 1.1 GB extracted).
           path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            src-tauri/target
-          key: ${{ runner.os }}-cargo-host-${{ hashFiles('**/Cargo.lock') }}
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          # One shared key across every job. The old per-job keys existed to stop
+          # debug/release target artifacts clobbering each other; with target no
+          # longer cached, registry contents are target-independent, so all jobs
+          # want the same crates. First job to finish saves; the rest restore.
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-host-
+            ${{ runner.os }}-cargo-registry-
 
       - name: Cache Node dependencies
         uses: actions/cache@v3
@@ -96,21 +132,91 @@ 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')
+
+      # TAURI_SKIP_UPDATER is gone: it was suppressing the updater artifacts
+      # (.AppImage.tar.gz + .sig) that the update manifest points at, back when
+      # there was no updater to feed. With the signing key present, `tauri build`
+      # emits and signs them.
+      #
+      # If TAURI_SIGNING_PRIVATE_KEY is ever absent the build fails loudly rather
+      # than quietly shipping an unsigned release that no client will accept --
+      # which is the behaviour we want.
+      # Same hazard as the Windows job: the bundle directory is never cleaned by
+      # cargo and the runner reuses src-tauri/target, while the copy step below
+      # globs bundle/deb/*.deb and friends. Windows is where this actually bit
+      # (v0.8.2 shipped thirteen stale installers), but only because Linux
+      # packaging is newer -- the glob is identical. Remove the directory so a
+      # stale artifact cannot exist to be copied.
+      - name: Clear previous bundle output
+        run: rm -rf src-tauri/target/release/bundle
+
       - name: Build for Linux
         run: bun run tauri build
         env:
-          TAURI_SKIP_UPDATER: true
+          # linuxdeploy's bundled `strip` cannot parse the `.relr.dyn` section
+          # modern toolchains emit, and fails on every bundled library:
+          #   strip: libzstd.so.1: unknown type [0x13] section `.relr.dyn'
+          #   failed to bundle project `failed to run linuxdeploy`
+          # Ubuntu 23.10+ links with -z pack-relative-relocs by default, so this
+          # image hits it. Skipping strip is linuxdeploy's documented escape
+          # hatch; the cost is a larger AppImage. Found by building the target
+          # locally before tagging -- nothing in CI builds the app, so a release
+          # would have been the first time anyone discovered the AppImage target
+          # does not work.
+          NO_STRIP: "true"
+          TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
+          TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
 
       - name: Prepare Linux artifacts
         run: |
           mkdir -p dist/linux
-          # Copy AppImage
-          if [ -f "src-tauri/target/release/bundle/appimage/jellytau_"*.AppImage ]; then
-            cp src-tauri/target/release/bundle/appimage/jellytau_*.AppImage dist/linux/
+          # Match by extension, not by product name. Bundle filenames follow
+          # `productName`, so renaming the app (jellytau -> JellyTau) made the
+          # old `jellytau_*.deb` glob match nothing — and because the copy was
+          # wrapped in `if [ -f ... ]`, the artifact simply vanished from the
+          # release with no error. Each bundle directory holds one file.
+          #
+          # `if [ -f "dir/"*.ext ]` was also wrong on its own terms: with more
+          # than one match `test` gets extra arguments and fails.
+          #
+          # No `shopt -s nullglob` here: the runner executes `run:` blocks with
+          # POSIX sh, where shopt does not exist -- it exited 127 and killed the
+          # step (which is why v0.9.0 and v0.9.1 built but never published).
+          # Without nullglob an unmatched pattern stays literal, so test each
+          # candidate instead. Same POSIX-only rule as traceability-check.yml.
+          #
+          # Tauri v2 signs the .AppImage ITSELF and writes <name>.AppImage.sig
+          # beside it -- there is no .AppImage.tar.gz unless
+          # bundle.createUpdaterArtifacts is set to "v1Compatible". The updater
+          # downloads the same AppImage a human does and verifies that .sig, so
+          # both files must ship or the manifest points at a signature nobody
+          # can fetch.
+          for bundle in \
+            src-tauri/target/release/bundle/appimage/*.AppImage \
+            src-tauri/target/release/bundle/appimage/*.AppImage.sig \
+            src-tauri/target/release/bundle/deb/*.deb \
+            src-tauri/target/release/bundle/rpm/*.rpm; do
+            [ -e "$bundle" ] || continue
+            cp -v "$bundle" dist/linux/
+          done
+
+          # An AppImage that did not build means no updater artifact either, and
+          # the release notes have advertised an AppImage for months. Fail rather
+          # than publish a release whose manifest points at nothing.
+          if ! ls dist/linux/*.AppImage >/dev/null 2>&1; then
+            echo "::error::No AppImage produced -- check bundle.targets in tauri.conf.json"
+            exit 1
           fi
-          # Copy .deb if built
-          if [ -f "src-tauri/target/release/bundle/deb/jellytau_"*.deb ]; then
-            cp src-tauri/target/release/bundle/deb/jellytau_*.deb dist/linux/
+
+          # A release with no Linux package is a failure, not a quiet success.
+          if [ -z "$(ls -A dist/linux/)" ]; then
+            echo "::error::No Linux bundles found under src-tauri/target/release/bundle/"
+            exit 1
           fi
           ls -lah dist/linux/
 
@@ -119,7 +225,7 @@ jobs:
         with:
           name: jellytau-linux
           path: dist/linux/
-          retention-days: 30
+          retention-days: 7
 
   build-windows:
     name: Build Windows
@@ -129,7 +235,7 @@ jobs:
     # 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
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -137,14 +243,31 @@ jobs:
       - name: Cache Rust dependencies
         uses: actions/cache@v3
         with:
+          # Registry only -- never src-tauri/target. That directory is ~16 GB and
+          # was cached under five separate keys, which filled the runner's 74 GB
+          # disk at ~1.15 GB/day (23 GB in 20 days, measured Aug 2026).
+          # registry/src is omitted too: cargo re-extracts it for free from
+          # registry/cache (155 MB of .crate tarballs vs 1.1 GB extracted).
           path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            ~/.cache/cargo-xwin
-            src-tauri/target
-          key: ${{ runner.os }}-cargo-windows-${{ hashFiles('**/Cargo.lock') }}
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          # One shared key across every job. The old per-job keys existed to stop
+          # debug/release target artifacts clobbering each other; with target no
+          # longer cached, registry contents are target-independent, so all jobs
+          # want the same crates. First job to finish saves; the rest restore.
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-windows-
+            ${{ runner.os }}-cargo-registry-
+
+      - name: Cache Windows CRT/SDK (cargo-xwin)
+        uses: actions/cache@v3
+        with:
+          path: ~/.cache/cargo-xwin
+          # Contents track the xwin version baked into the builder image, not our
+          # lockfile -- keying this on Cargo.lock re-downloaded the whole SDK on
+          # every release bump. Bump the suffix by hand if the image's xwin moves.
+          key: ${{ runner.os }}-cargo-xwin-v1
 
       - name: Cache Node dependencies
         uses: actions/cache@v3
@@ -156,18 +279,19 @@ jobs:
           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: |
-          # On a tag build the tag is the single source of truth for the version.
-          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: Build Windows (NSIS installer + exe)
         run: OUTPUT_DIR="$PWD/dist/windows" WIN_BUNDLES=nsis ./scripts/build-windows-cross.sh
+        env:
+          TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
+          TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
 
       - name: List Windows artifacts
         run: ls -lah dist/windows/
@@ -177,14 +301,14 @@ jobs:
         with:
           name: jellytau-windows
           path: dist/windows/
-          retention-days: 30
+          retention-days: 7
 
   build-android:
     name: Build Android
     runs-on: linux/amd64
     needs: test
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
     env:
       ANDROID_HOME: /opt/android-sdk
       ANDROID_SDK_ROOT: /opt/android-sdk
@@ -196,13 +320,22 @@ jobs:
       - name: Cache Rust dependencies
         uses: actions/cache@v3
         with:
+          # Registry only -- never src-tauri/target. That directory is ~16 GB and
+          # was cached under five separate keys, which filled the runner's 74 GB
+          # disk at ~1.15 GB/day (23 GB in 20 days, measured Aug 2026).
+          # registry/src is omitted too: cargo re-extracts it for free from
+          # registry/cache (155 MB of .crate tarballs vs 1.1 GB extracted).
           path: |
-            ~/.cargo/registry
-            ~/.cargo/git
-            src-tauri/target
-          key: ${{ runner.os }}-cargo-android-${{ hashFiles('**/Cargo.lock') }}
+            ~/.cargo/registry/index
+            ~/.cargo/registry/cache
+            ~/.cargo/git/db
+          # One shared key across every job. The old per-job keys existed to stop
+          # debug/release target artifacts clobbering each other; with target no
+          # longer cached, registry contents are target-independent, so all jobs
+          # want the same crates. First job to finish saves; the rest restore.
+          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
           restore-keys: |
-            ${{ runner.os }}-cargo-android-
+            ${{ runner.os }}-cargo-registry-
 
       - name: Cache Node dependencies
         uses: actions/cache@v3
@@ -217,48 +350,22 @@ 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
@@ -273,8 +380,12 @@ jobs:
           keyPassword=${{ secrets.ANDROID_KEY_PASSWORD }}
           EOF
 
+      # `--apk` is a boolean flag, not `--apk true`. tauri-cli took a value here
+      # until 2.10; from 2.11 the stray `true` is parsed as a positional and the
+      # command fails with "unexpected argument 'true' found" before building.
+      # This line and scripts/build-android.sh must agree.
       - name: Build signed Android APK
-        run: bun run tauri android build --apk true --target aarch64
+        run: bun run tauri android build --apk --target aarch64
 
       - name: Collect & verify signed APK
         run: |
@@ -292,7 +403,7 @@ jobs:
         with:
           name: jellytau-android
           path: dist/android/
-          retention-days: 30
+          retention-days: 7
 
   create-release:
     name: Create Release
@@ -300,7 +411,7 @@ jobs:
     needs: [build-linux, build-windows, build-android]
     if: startsWith(github.ref, 'refs/tags/v')
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -329,63 +440,228 @@ jobs:
           name: jellytau-android
           path: artifacts/android/
 
+      # Runs before the SBOM, the checksums and the upload -- everything
+      # downstream describes this set of files, so a stale artifact must be
+      # caught before it gets hashed into SHA256SUMS and published as though it
+      # belonged to this release.
+      #
+      # See the script for the eight months of releases that shipped their
+      # predecessors' Windows installers.
+      - name: Verify artifacts belong to this release
+        run: |
+          ./scripts/check-release-artifacts.sh \
+            "${{ steps.tag_name.outputs.VERSION }}" \
+            artifacts/linux artifacts/windows artifacts/android
+
+      # Software Bill of Materials, one per half of the app. Without it there is
+      # no answer to "does this release contain <vulnerable crate>?" other than
+      # rebuilding the tag and re-resolving it. cargo-cyclonedx is in the builder
+      # image; the JS side is read straight from the lockfile bun install used.
+      - name: Generate SBOM
+        run: |
+          set -e
+          mkdir -p artifacts/sbom
+          cd src-tauri
+          cargo cyclonedx --format json
+          find . -maxdepth 2 -name "*.cdx.json" -exec cp -v {} ../artifacts/sbom/ \;
+          cd ..
+          bun install --frozen-lockfile
+          bun pm ls --all > artifacts/sbom/frontend-dependencies.txt
+          ls -lah artifacts/sbom/
+
+      # Checksums over everything being published. A release of unsigned Linux
+      # and Windows binaries with no checksum gives a user no way at all to tell
+      # a corrupted or substituted download from a good one — and the AppImage
+      # and NSIS installer are both fetched over plain HTTP redirects.
+      #
+      # Written with paths relative to the asset directory so `sha256sum -c
+      # SHA256SUMS` works in the directory a user downloaded into.
+      # The update manifest. Built before the checksums so latest.json is not
+      # itself hashed into SHA256SUMS (it is metadata about the release, not a
+      # download), and after the artifacts exist so the signatures can be read.
+      #
+      # Why a dedicated `updater` branch and a raw-file URL: this Gitea serves
+      # /releases/download/<tag>/<asset> but returns 404 for
+      # /releases/latest/download/<asset>, so there is no stable "latest release"
+      # URL to point a client at. The gitea-pages branch is force-pushed whole by
+      # publish-docs.yml, so hosting the manifest there would delete it on the
+      # next docs build. An orphan branch that only ever contains latest.json is
+      # the one location both stable and ours.
+      - name: Build update manifest (latest.json)
+        id: manifest
+        run: |
+          set -e
+          VERSION="${{ steps.tag_name.outputs.VERSION }}"
+          # The manifest carries the bare version; the tag carries the v prefix.
+          PLAIN="${VERSION#v}"
+          BASE="${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/releases/download/${VERSION}"
+
+          # Tauri matches on "<os>-<arch>". We ship one desktop arch today.
+          APPIMAGE_SIG=""
+          NSIS_SIG=""
+          APPIMAGE_URL=""
+          NSIS_URL=""
+
+          # Tauri v2 signs the AppImage itself; <name>.AppImage.sig sits beside
+          # it. Verified against a real signed build before tagging -- the
+          # v1-style .AppImage.tar.gz is never produced with
+          # createUpdaterArtifacts: true.
+          for f in artifacts/linux/*.AppImage; do
+            [ -e "$f" ] || continue
+            case "$f" in *.sig) continue;; esac
+            APPIMAGE_URL="${BASE}/$(basename "$f")"
+            [ -e "$f.sig" ] && APPIMAGE_SIG="$(cat "$f.sig")"
+          done
+
+          for f in artifacts/windows/*-setup.exe; do
+            [ -e "$f" ] || continue
+            NSIS_URL="${BASE}/$(basename "$f")"
+            [ -e "$f.sig" ] && NSIS_SIG="$(cat "$f.sig")"
+          done
+
+          # A manifest with an empty signature is worse than no manifest: the
+          # client rejects it after downloading the whole payload.
+          if [ -z "$APPIMAGE_SIG" ] || [ -z "$NSIS_SIG" ]; then
+            echo "::error::Missing updater signature (appimage='$APPIMAGE_SIG' nsis='$NSIS_SIG')."
+            echo "::error::Check that TAURI_SIGNING_PRIVATE_KEY reached both desktop build jobs."
+            exit 1
+          fi
+
+          # What the in-app update prompt shows. Same reviewed source as the
+          # release body -- the CHANGELOG section for this version, not the
+          # traceability draft.
+          NOTES="$(awk -v ver="## $VERSION" '$0==ver{f=1;next} /^## /{if(f)exit} f' CHANGELOG.md | head -c 4000)"
+          [ -n "$NOTES" ] || NOTES="See the release page for details."
+
+          jq -n \
+            --arg version "$PLAIN" \
+            --arg notes "$NOTES" \
+            --arg pub_date "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
+            --arg lin_sig "$APPIMAGE_SIG" --arg lin_url "$APPIMAGE_URL" \
+            --arg win_sig "$NSIS_SIG"     --arg win_url "$NSIS_URL" \
+            '{
+              version: $version,
+              notes: $notes,
+              pub_date: $pub_date,
+              platforms: {
+                "linux-x86_64":   { signature: $lin_sig, url: $lin_url },
+                "windows-x86_64": { signature: $win_sig, url: $win_url }
+              }
+            }' > latest.json
+
+          echo "📄 latest.json:"
+          cat latest.json
+
+      - name: Publish latest.json to the updater branch
+        env:
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN }}
+          AUTO_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -e
+          TOKEN="${GITEA_TOKEN:-$AUTO_TOKEN}"
+          HOST="$(echo "$GITHUB_SERVER_URL" | sed -E 's#^https?://##')"
+          REMOTE="https://oauth2:${TOKEN}@${HOST}/${GITHUB_REPOSITORY}.git"
+
+          # Built in a scratch repo, NOT by switching branches in the checkout.
+          # `git checkout --orphan` here would leave every later step standing on
+          # a one-commit branch -- and the next step but one runs
+          # `bun run release:notes`, which resolves a commit range against the
+          # real history and would silently produce nothing.
+          WORK="$RUNNER_TEMP/updater-branch"
+          rm -rf "$WORK"
+          mkdir -p "$WORK"
+          cp latest.json "$WORK/latest.json"
+          cd "$WORK"
+          git init -q
+          git config user.email "ci@jellytau"
+          git config user.name "JellyTau CI"
+          git add latest.json
+          git commit -qm "chore(updater): manifest for ${{ steps.tag_name.outputs.VERSION }}"
+          echo "🚀 Force-pushing update manifest to the updater branch"
+          # Force-push: the branch holds exactly one file and no history worth
+          # keeping, same shape as publish-docs.yml's gitea-pages.
+          git push -f "$REMOTE" HEAD:refs/heads/updater
+          echo "✅ Served at ${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/raw/branch/updater/latest.json"
+
+      - name: Generate SHA256SUMS
+        run: |
+          set -e
+          mkdir -p artifacts/release
+          find artifacts/linux artifacts/windows artifacts/android -type f -exec cp -v {} artifacts/release/ \;
+          cd artifacts/release
+          sha256sum * > SHA256SUMS
+          echo "🔐 Published checksums:"
+          cat SHA256SUMS
+          # Verify what we just wrote, so a broken checksum file fails the
+          # release rather than shipping and failing for users.
+          sha256sum -c SHA256SUMS
+
+      # The published body is the hand-written CHANGELOG.md section for this
+      # version. `bun run release:notes` is printed into the job log as a
+      # drafting aid, but is NOT published: CLAUDE.md is explicit that its
+      # output is "a reviewed draft, not a final changelog", and publishing it
+      # unreviewed proved the point -- a range containing a repo-wide prettier
+      # sweep resolved to nearly the whole requirement matrix and produced notes
+      # claiming one release had added the entire application.
+      #
+      # A missing CHANGELOG section fails the release. A release whose notes say
+      # nothing is worse than one that waits for a maintainer to write two
+      # sentences, and the checklist already requires that entry.
       - name: Prepare release notes
         id: release_notes
         run: |
+          set -e
           VERSION="${{ steps.tag_name.outputs.VERSION }}"
-          echo "## JellyTau $VERSION Release" > release_notes.md
-          echo "" >> release_notes.md
-          echo "### Downloads" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "#### Linux" >> release_notes.md
-          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
-          echo "" >> release_notes.md
-          echo "### What's New" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "See [CHANGELOG.md](CHANGELOG.md) for detailed changes." >> release_notes.md
-          echo "" >> release_notes.md
-          echo "### Installation" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "#### Linux (AppImage)" >> release_notes.md
-          echo "\`\`\`bash" >> release_notes.md
-          echo "chmod +x jellytau_*.AppImage" >> release_notes.md
-          echo "./jellytau_*.AppImage" >> release_notes.md
-          echo "\`\`\`" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "#### Linux (DEB)" >> release_notes.md
-          echo "\`\`\`bash" >> release_notes.md
-          echo "sudo dpkg -i jellytau_*.deb" >> release_notes.md
-          echo "jellytau" >> release_notes.md
-          echo "\`\`\`" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "#### Android" >> release_notes.md
-          echo "- Sideload: Download APK and install via file manager or ADB" >> release_notes.md
-          echo "- Play Store: Coming soon" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "### Known Issues" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "See [GitHub Issues](../../issues) for reported bugs." >> release_notes.md
-          echo "" >> release_notes.md
-          echo "### Requirements" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "**Linux:**" >> release_notes.md
-          echo "- 64-bit Linux system" >> release_notes.md
-          echo "- GLIBC 2.29+" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "**Android:**" >> release_notes.md
-          echo "- Android 8.0 or higher" >> release_notes.md
-          echo "- 50MB free storage" >> release_notes.md
-          echo "" >> release_notes.md
-          echo "---" >> release_notes.md
-          echo "Built with Tauri, SvelteKit, and Rust" >> release_notes.md
+
+          echo "📋 Traceability draft (for reference; not published):"
+          bun run release:notes 2>/dev/null || echo "(could not derive a draft)"
+          echo ""
+
+          # The section between this version's heading and the next one.
+          CHANGES=$(awk -v ver="## $VERSION" '$0==ver{f=1;next} /^## /{if(f)exit} f' CHANGELOG.md)
+          if [ -z "$(echo "$CHANGES" | tr -d '[:space:]')" ]; then
+            echo "::error::CHANGELOG.md has no '## $VERSION' section."
+            echo "::error::Add the entry for this version and re-tag; see docs/release-checklist.md."
+            exit 1
+          fi
+
+          {
+            echo "$CHANGES"
+            echo ""
+            echo "### Downloads"
+            echo ""
+            echo "| Platform | File |"
+            echo "|---|---|"
+            echo "| Linux (portable) | \`*.AppImage\` — \`chmod +x\` and run |"
+            echo "| Linux (Debian/Ubuntu) | \`*.deb\` — \`sudo dpkg -i\` |"
+            echo "| Linux (Fedora/openSUSE) | \`*.rpm\` — \`sudo rpm -i\` |"
+            echo "| Windows | \`*-setup.exe\` (NSIS). Unsigned — SmartScreen may warn on first run. |"
+            echo "| Android | \`*.apk\` sideload, or \`*.aab\` for Play Console |"
+            echo ""
+            echo "Desktop builds check for updates from here and can install a new"
+            echo "version in place, verifying its signature first."
+            echo ""
+            echo "### Verifying your download"
+            echo ""
+            echo "\`\`\`bash"
+            echo "sha256sum -c SHA256SUMS"
+            echo "\`\`\`"
+            echo ""
+            echo "\`SHA256SUMS\` covers every file in this release. An SBOM"
+            echo "(\`*.cdx.json\`, \`frontend-dependencies.txt\`) lists what went into it."
+            echo ""
+            echo "### Requirements"
+            echo ""
+            echo "- **Linux:** 64-bit, GLIBC 2.29+"
+            echo "- **Windows:** 64-bit Windows 10 or later"
+            echo "- **Android:** 8.0 or later, ~50 MB free"
+            echo ""
+            echo "---"
+            echo "Report a problem: ${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/issues"
+          } > release_notes.md
+
+          echo "📝 Release notes:"
+          cat release_notes.md
 
       - name: Publish Gitea release & upload assets
         env:
@@ -425,7 +701,10 @@ jobs:
           fi
           echo "Release id=$RELEASE_ID"
 
-          for f in artifacts/android/* artifacts/linux/* artifacts/windows/*; do
+          # artifacts/release/ holds a copy of every platform artifact plus the
+          # SHA256SUMS generated over exactly that set, so the checksums describe
+          # precisely what is uploaded. artifacts/sbom/ rides along.
+          for f in artifacts/release/* artifacts/sbom/*; do
             [ -f "$f" ] || continue
             echo "⬆️  Uploading $(basename "$f")"
             curl -fsS -X POST \

.gitea/workflows/publish-docs.yml

@@ -21,7 +21,7 @@ jobs:
     name: Build & publish docs to gitea-pages
     runs-on: linux/amd64
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
 
     steps:
       - name: Checkout code
@@ -34,14 +34,13 @@ jobs:
       - name: Install dependencies
         run: bun install
 
-      - name: Install mdBook
-        run: |
-          set -e
-          MDBOOK_VERSION=v0.4.40
-          URL="https://github.com/rust-lang/mdBook/releases/download/${MDBOOK_VERSION}/mdbook-${MDBOOK_VERSION}-x86_64-unknown-linux-gnu.tar.gz"
-          echo "⬇️  Downloading mdBook ${MDBOOK_VERSION}"
-          curl -fsSL "$URL" | tar -xz -C /usr/local/bin
-          mdbook --version
+      # mdBook is baked into jellytau-builder (Dockerfile.builder, MDBOOK_VERSION).
+      # It used to be curl'd from GitHub releases straight into /usr/local/bin
+      # right here, which was a toolchain install at job time — the exact thing
+      # CLAUDE.md's 🔴 rule forbids — and made every docs publish depend on
+      # GitHub's CDN answering. To move the version, bump it in the image.
+      - name: Confirm mdBook is present
+        run: mdbook --version
 
       - name: Regenerate traceability matrix (keep published copy current)
         run: bun run traces:markdown

.gitea/workflows/traceability-check.yml

@@ -17,7 +17,7 @@ jobs:
     runs-on: linux/amd64
     name: Check Requirement Traces
     container:
-      image: gitea.tourolle.paris/dtourolle/jellytau-builder:latest
+      image: gitea.tourolle.paris/dtourolle/jellytau-builder:2026.08.1
 
     steps:
       - name: Checkout repository
@@ -42,32 +42,59 @@ 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/traceability-ci.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 ""
 
-          # Check minimum threshold
-          MIN_THRESHOLD=50
+          # 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: $ORPHANED"
+            exit 1
+          fi
+
+          # Minimum coverage. RATCHET POLICY: this number only ever goes UP.
+          #
+          # It sits a few points under the coverage actually achieved, so a real
+          # regression trips it. It was 50 while true coverage was 86%, which
+          # meant nearly half the matrix could rot before CI said a word — a
+          # gate that cannot fail is not a gate.
+          #
+          # When coverage rises durably, raise this to just under the new figure
+          # (`bun run traces:coverage` prints it). Never lower it to make a red
+          # build pass — add the missing TRACES comments instead.
+          #
+          # Keep in sync with MIN_COVERAGE_PERCENT in scripts/extract-traces.ts;
+          # scripts/extract-traces.test.ts fails if the two drift apart.
+          MIN_THRESHOLD=89
           if [ "$COVERAGE" -lt "$MIN_THRESHOLD" ]; then
             echo "❌ ERROR: Coverage ($COVERAGE%) is below minimum threshold ($MIN_THRESHOLD%)"
             exit 1
@@ -75,6 +102,15 @@ jobs:
 
           echo "✅ Coverage is acceptable ($COVERAGE% >= $MIN_THRESHOLD%)"
 
+      # Every ID named by a TRACES comment must be defined as a table row in
+      # docs/requirements.md. The extractor used to accept any well-formed ID
+      # silently, so a typo or a rename that missed a call site passed CI
+      # unnoticed (DR-189 and UT-188 lived in three source files, defined
+      # nowhere, for months). This covers UT/IT too, which the coverage
+      # orphan list above deliberately ignores.
+      - name: Validate requirement IDs
+        run: bun run traces:validate
+
       - name: Check modified files
         if: github.event_name == 'pull_request'
         run: |

.gitignore

@@ -30,11 +30,6 @@ coverage
 .nyc_output
 *.lcov
 
-# WebdriverIO E2E tests
-e2e/logs/
-e2e/screenshots/
-wdio-*.log
-
 # Vitest
 .vitest
 
@@ -64,3 +59,9 @@ 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

.prettierignore

@@ -0,0 +1,35 @@
+# Dependencies & build output
+node_modules/
+.svelte-kit/
+
+# Scratch worktrees (git-ignored) — full checkouts of this repo
+.claude/
+build/
+dist/
+coverage/
+/package/
+
+# Rust backend (rustfmt owns this tree)
+src-tauri/
+
+# Generated by tauri-specta — regenerated on every Rust build, never hand-edited
+src/lib/api/bindings.ts
+
+# Lockfiles and generated data
+bun.lock
+*.lcov
+
+# Generated docs (built by the publish-docs CI job)
+docs/SUMMARY.md
+docs/README.md
+docs/api-redirect.md
+docs-site/book/
+
+# Hand-maintained Markdown (docs/, CHANGELOG.md, README.md, ...). Prettier
+# reflows tables and wrapped prose, which would swamp real doc diffs and fight
+# the hand-tuned layout of docs/requirements.md and docs/traceability.md
+# (the latter is generated by scripts/extract-traces.ts).
+**/*.md
+
+# CI workflow YAML — formatting churn here would obscure real pipeline diffs.
+.gitea/

.prettierrc

@@ -0,0 +1,20 @@
+{
+  "$schema": "https://json.schemastore.org/prettierrc",
+  "printWidth": 100,
+  "tabWidth": 2,
+  "useTabs": false,
+  "semi": true,
+  "singleQuote": false,
+  "quoteProps": "as-needed",
+  "trailingComma": "all",
+  "bracketSpacing": true,
+  "arrowParens": "always",
+  "endOfLine": "lf",
+  "plugins": ["prettier-plugin-svelte"],
+  "overrides": [
+    {
+      "files": "*.svelte",
+      "options": { "parser": "svelte" }
+    }
+  ]
+}

CLAUDE.md

@@ -21,7 +21,8 @@ bun run check                # svelte-check (types)
 bun run test                 # vitest (frontend unit/integration)
 bun run test:rust            # cargo test (scripts/test-rust.sh)
 bun run test:all             # full suite (scripts/test-all.sh)
-bun run test:e2e             # webdriverio e2e
+bun run lint                 # eslint (src/, scripts/, root configs)
+bun run format:check         # prettier
 
 # Android — canonical entry points (see scripts/):
 bun run android:build            # debug APK
@@ -31,6 +32,16 @@ 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`.
@@ -51,7 +62,8 @@ only against the mirror if one exists; the canonical remote is
 
 ## Before Committing
 
-- Frontend: `bun run check` and `bun run test` must pass.
+- Frontend: `bun run check`, `bun run test`, `bun run format:check` and
+  `bun run lint` (0 errors; the warning count is a CI ratchet) must pass.
 - Rust: `cd src-tauri && cargo fmt` then `cargo clippy`, plus `bun run test:rust`.
 - **Boundary**: `bun run check:boundary` must pass — no domain taxonomy (Jellyfin
   item-type category sets) leaked into the frontend. See below.
@@ -91,14 +103,24 @@ Tooling:
 bun run traces            # extract traces (default format)
 bun run traces:json       # JSON — e.g. | jq '.byType' or '.requirements."UR-005"'
 bun run traces:markdown   # regenerate docs/traceability.md
+bun run traces:coverage   # coverage gate — exits non-zero below the threshold
+bun run traces:validate   # dangling-ID gate — every traced ID must be defined
 git diff --name-only | xargs grep -L "TRACES:"   # find untraced changed files
 ```
 
+Every ID a `TRACES:` comment names must exist as a table row in
+`docs/requirements.md` — `traces:validate` fails otherwise, so a typo or a
+rename that missed a call site can no longer pass silently.
+
 **CI is Gitea Actions** (`.gitea/workflows/`, remote `gitea.tourolle.paris`), not
 GitHub. `traceability-check.yml` fails the build if coverage drops below
-**50%** (`MIN_THRESHOLD`); `build-and-test.yml` runs frontend + Rust tests and an
-Android `cargo check`. See [docs/traceability-ci.md](docs/traceability-ci.md) and
-[docs/traces-quick-ref.md](docs/traces-quick-ref.md).
+**89%** (`MIN_THRESHOLD`, a *ratchet* — raise it as coverage climbs, never lower
+it to make a build pass) or if any traced ID is undefined; `build-and-test.yml`
+runs frontend tests **with coverage thresholds**, `bun run check`, `format:check`,
+a `--max-warnings` eslint ratchet, Rust tests, `cargo fmt --check`, `cargo clippy
+-D warnings`, and an Android `cargo check`. See
+[docs/traceability-ci.md](docs/traceability-ci.md)
+and [docs/traces-quick-ref.md](docs/traces-quick-ref.md).
 
 ### Traces drive release notes
 
@@ -152,7 +174,7 @@ canonical, maintained source; this file only summarizes. See
 | [09-security.md](docs/architecture/09-security.md) | Token storage, secure storage, network security |
 
 Release process lives in [docs/release-checklist.md](docs/release-checklist.md)
-and [docs/build-release.md](docs/build-release.md).
+and [docs/build/build-release.md](docs/build/build-release.md).
 
 ### Core principles (from the architecture docs)
 
@@ -183,12 +205,20 @@ 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
 
-New feature specs go in [docs/specs/](docs/specs/). **Start from
+New feature specs go in [docs/specs/](docs/specs/) — see its
+[README](docs/specs/README.md) for the index and what is already built.
+**Start from
 [SPEC-TEMPLATE.md](docs/specs/SPEC-TEMPLATE.md)** — its "Layer assignment" section
 forces each piece of *logic* to be placed in the correct layer (Rust = domain,
 frontend = presentation) *with a reason*, which is what prevents boundary leaks.
@@ -197,6 +227,34 @@ Before accepting a spec, run it past
 a spec around "no Rust changes required" — correct layer placement is the goal,
 not minimal backend churn.
 
+### 🔴 A spec becomes an architecture doc when it ships
+
+`docs/specs/` holds **only work that has not shipped**. There is no "Implemented"
+resting state for a spec file: when the last acceptance criterion is met, fold
+the design into [docs/architecture/](docs/architecture/README.md) and **delete
+the spec in the same commit**.
+
+This is not tidying. A directory that mixes promises with descriptions makes both
+unreliable — you cannot tell from a file whether it describes the build or
+proposes a change to it, and stale specs then quietly disagree with the code
+while reading as authority.
+
+- **Every spec names its destination up front** — the template's "Destination on
+  completion" line. Deciding at spec time which architecture doc will absorb it
+  is a design check in itself: a feature that fits no existing doc is usually a
+  feature whose layer assignment is unclear.
+- **Carry the reasoning, not the plan.** The architecture doc gets the *why* a
+  future change still needs — invariants, rejected alternatives that would be
+  re-attempted, the defect a piece of code exists to prevent. Acceptance
+  criteria, phase breakdowns and migration steps die with the spec; git history
+  keeps them.
+- **Deferred work outlives its spec.** Anything the spec listed as out-of-scope
+  and still worth doing goes beside the code it concerns, not into the void.
+- **Rewrite inbound references before deleting** — source comments and CI
+  scripts cite spec paths, and `check-doc-links` only sees markdown.
+- **Partially implemented is a real status.** A spec stays until *all* of it
+  ships, with the header naming what is left.
+
 ## Conventions
 
 ### Rust Backend

CODE_OF_CONDUCT.md

@@ -0,0 +1,47 @@
+# Code of Conduct
+
+## The short version
+
+Be decent to people. Assume the person you are talking to is acting in good
+faith and knows things you do not.
+
+## What that means here
+
+**Expected:**
+
+- Criticise code, decisions and ideas — not the people who wrote them.
+- Accept that "no" is a complete answer. This is a small project with a
+  maintainer who has finite time; a declined feature request is not a slight.
+- Give people room to be new. Everyone was once confused by Tauri's IPC.
+- Assume a bug report is someone trying to help, even when it arrives terse or
+  frustrated.
+
+**Not accepted:**
+
+- Harassment, personal attacks, or demeaning remarks — including about someone's
+  identity, background, or level of experience.
+- Sexualised language or imagery, and unwelcome attention of any kind.
+- Publishing someone's private information without their permission.
+- Persistently derailing discussions, or badgering people who have already
+  answered you.
+
+## Scope
+
+This applies in the issue tracker, pull requests, commit messages and any other
+project space, and to anyone taking part — maintainer included.
+
+## Reporting
+
+Email **duncan@tourolle.paris**. Reports are read by the maintainer and handled
+privately.
+
+Responses range from a quiet word through to removing comments or blocking an
+account, depending on what happened. If a report concerns the maintainer, and
+that makes reporting to them pointless, you are free to say so publicly — a
+project this size has no separate committee to appeal to, and pretending
+otherwise would be dishonest.
+
+## Attribution
+
+Adapted in spirit from the [Contributor Covenant](https://www.contributor-covenant.org),
+shortened to what a single-maintainer project can actually honour.

CONTRIBUTING.md

@@ -0,0 +1,123 @@
+# Contributing to JellyTau
+
+Thanks for looking. This file is the short version of how the project is built
+and what has to be true before a change lands. The long version lives in
+[CLAUDE.md](CLAUDE.md) and [docs/architecture/](docs/architecture/README.md),
+which are maintained rather than decorative — read them before a structural
+change.
+
+## Getting set up
+
+Package manager is **bun**. You will also need a Rust toolchain (the exact
+version is pinned in [src-tauri/rust-toolchain.toml](src-tauri/rust-toolchain.toml)
+— rustup honours it automatically) and the Tauri Linux dependencies.
+
+```bash
+bun install
+bun run hooks:install    # do this once: it enables the pre-commit gates
+bun run tauri dev
+```
+
+`hooks:install` points `core.hooksPath` at [scripts/hooks/](scripts/hooks/), so
+hook updates arrive with a `git pull` instead of needing a re-install.
+
+## What has to pass
+
+Everything below runs in CI, and the fast half runs in the pre-commit hook. None
+of it is advisory:
+
+```bash
+bun run check          # svelte-check — 0 errors
+bun run test           # vitest
+bun run format:check   # prettier
+bun run lint           # eslint — 0 errors; the warning count is a ratchet
+bun run check:boundary # no Jellyfin taxonomy in the frontend
+bun run test:rust      # cargo test
+cd src-tauri && cargo fmt --all && cargo clippy --all-targets -- -D warnings
+cd src-tauri && cargo deny check          # advisories, licences, bans, sources
+```
+
+`bun run test:all` runs the whole set.
+
+Several of these are **ratchets** — a number that only ever moves in the
+improving direction:
+
+| Ratchet | Where | Rule |
+|---|---|---|
+| eslint `--max-warnings` | [.gitea/workflows/build-and-test.yml](.gitea/workflows/build-and-test.yml) | only goes down |
+| Coverage thresholds | [vitest.config.ts](vitest.config.ts) | only go up |
+| Traceability coverage | [.gitea/workflows/traceability-check.yml](.gitea/workflows/traceability-check.yml) | only goes up |
+
+Never relax one to make a build pass. Fix the thing it caught.
+
+## The two rules that surprise people
+
+**1. Bug fixes start with a failing test.** Write a test that reproduces the bug
+and *watch it fail* before you touch the fix. A test written against
+already-fixed code can pass for the wrong reason and guards nothing. If the logic
+is trapped in a component, extract the pure part into a plain `.ts` module and
+test that — see `episodeStrip.ts` or `TrackList.logic.ts` for the pattern.
+
+**2. Domain vocabulary lives in Rust.** The frontend is presentation-only. It
+must not encode Jellyfin's *taxonomy* — for example, the set of item types that
+makes up a category like "Music". Send an opaque scope across the IPC boundary
+and let the backend expand it. `bun run check:boundary` is a tripwire, not a
+proof: it only flags item-type array literals, so a green run does not mean you
+are clear. [docs/specs/scoped-search-boundary.md](docs/specs/scoped-search-boundary.md)
+describes the leak that made this a rule.
+
+## Traceability
+
+Code that implements a requirement carries a `TRACES:` comment naming the
+requirement IDs, and a tool builds the matrix from those comments:
+
+```rust
+/// TRACES: UR-005 | DR-001
+```
+
+Every ID must exist as a row in [docs/requirements.md](docs/requirements.md) —
+`bun run traces:validate` fails on a typo or a stale rename. Internal helpers and
+requirement-less code stay untraced; do not sprinkle IDs to raise the number.
+
+If you add a requirement, add its row. If you implement one, tag the code.
+
+## Commits and pull requests
+
+- Conventional-commit subjects: `fix(player): …`, `feat(updater): …`, `ci: …`.
+- Explain **why** in the body, not what the diff already shows. The commit log
+  is the main record of why things are the way they are here, and it is used to
+  draft release notes.
+- One concern per commit. A formatting sweep and a behaviour change in the same
+  commit is unreviewable.
+- Rebase rather than merge-commit onto `master`.
+
+## Specs
+
+New features start from [docs/specs/SPEC-TEMPLATE.md](docs/specs/SPEC-TEMPLATE.md).
+Its "Layer assignment" section is the point: each piece of logic gets placed in
+Rust or the frontend *with a reason*. Review against
+[docs/specs/SPEC-REVIEW-CHECKLIST.md](docs/specs/SPEC-REVIEW-CHECKLIST.md). Do
+not frame a spec around "no Rust changes required" — correct placement is the
+goal, not minimal backend churn.
+
+## CI
+
+CI is **Gitea Actions** (`.gitea/workflows/`), not GitHub.
+
+🔴 **CI installs no system tools.** Every build, test and packaging tool must
+already be in the Docker builder image. If a job needs a tool the image lacks,
+add it to [Dockerfile.builder](Dockerfile.builder), rebuild and push the image,
+and pin the new tag — do not `apt-get` it at job time. Details in
+[docs/build/ci-operations.md](docs/build/ci-operations.md).
+
+Fetching the project's own declared dependencies (`bun install`, cargo crates,
+an advisory database) is not a toolchain install and is fine.
+
+## Reporting bugs
+
+Use the issue templates. For anything involving playback, include what the
+platform was, whether the media was streaming or downloaded, and whether it was
+transcoding — those three answers determine which of several code paths you were
+actually on.
+
+Security issues go to [SECURITY.md](SECURITY.md), not the tracker.

Dockerfile

@@ -117,9 +117,13 @@ RUN cd src-tauri && cargo fetch && cd .. && \
 
 # 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, mingw-w64 + nsis + the x86_64-pc-windows-gnu rust target
-# for Windows). ONE source of dependency truth, shared with CI — no per-stage
-# apt/rustup here.
+# (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

Dockerfile.builder

@@ -52,13 +52,34 @@ RUN curl -fsSL https://deb.nodesource.com/setup_20.x | bash - && \
 RUN curl -fsSL https://bun.sh/install | bash && \
     ln -s /root/.bun/bin/bun /usr/local/bin/bun
 
-# Install Rust using rustup
-RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && \
+# Install Rust using rustup, pinned to an exact release.
+#
+# 🔴 RUST_VERSION must equal `channel` in src-tauri/rust-toolchain.toml.
+#
+# The two are a pair. rust-toolchain.toml is what makes a developer's `cargo
+# clippy` agree with CI's; this line is what makes the image already contain that
+# toolchain. If they drift, rustup silently downloads the pinned version the
+# first time cargo runs inside a job — a toolchain install at job time, which
+# CLAUDE.md's "🔴 CI installs no system tools" rule forbids (and which costs
+# ~1min plus a network dependency on every build).
+#
+# 🔴 Changing this line does NOT change CI on its own: the image must be
+# rebuilt and pushed (`scripts/build-builder-image.sh`) before the new pin is
+# authoritative. Bump rust-toolchain.toml and this line together, rebuild, push,
+# then merge.
+#
+# Was: `sh -s -- -y` (latest stable, whatever it happened to be on rebuild day).
+ENV RUST_VERSION=1.97.1
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | \
+        sh -s -- -y --profile minimal --default-toolchain "$RUST_VERSION" && \
     . $HOME/.cargo/env && \
+    rustup default "$RUST_VERSION" && \
     rustup target add aarch64-linux-android && \
     rustup target add armv7-linux-androideabi && \
     rustup target add x86_64-linux-android && \
-    rustup component add rustfmt clippy
+    rustup component add rustfmt clippy && \
+    rustc --version && \
+    cargo clippy --version
 
 # Setup Android SDK
 RUN mkdir -p $ANDROID_HOME && \
@@ -87,6 +108,22 @@ 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
@@ -104,6 +141,17 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     lld \
     llvm \
     nsis \
+    # AppImage bundling. linuxdeploy embeds xdg-open into the AppImage and
+    # aborts the whole bundle if it is missing:
+    #   failed to bundle project: xdg-open binary not found
+    # It is present on most desktop distros, which is why the AppImage built on
+    # a developer machine and failed here. desktop-file-utils and zsync are the
+    # other two linuxdeploy commonly wants (desktop-file-validate, and zsync for
+    # delta updates), added together so a missing one does not cost another
+    # image rebuild and another failed release build.
+    xdg-utils \
+    desktop-file-utils \
+    zsync \
     && 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,
@@ -115,6 +163,29 @@ RUN . $HOME/.cargo/env && \
     rustup target add x86_64-pc-windows-msvc && \
     cargo install --locked cargo-xwin
 
+# ---------------------------------------------------------------------------
+# Supply-chain and docs tooling.
+#
+# cargo-deny     — advisories/licences/bans/sources gate (src-tauri/deny.toml),
+#                  run by the `security` job. It fetches the RustSec advisory
+#                  database at run time; that is *data*, not a toolchain, so it
+#                  does not breach the no-installs-in-CI rule.
+# cargo-cyclonedx — SBOM for the Rust half of a release.
+# mdbook         — builds the docs site. It used to be curl'd from GitHub
+#                  releases *inside* the job (publish-docs.yml), which was both a
+#                  breach of that rule and a hard dependency on GitHub's CDN
+#                  being up at publish time. Pinned to the version that job used.
+ENV MDBOOK_VERSION=v0.4.40
+RUN . $HOME/.cargo/env && \
+    cargo install --locked cargo-deny cargo-cyclonedx && \
+    wget -q "https://github.com/rust-lang/mdBook/releases/download/${MDBOOK_VERSION}/mdbook-${MDBOOK_VERSION}-x86_64-unknown-linux-gnu.tar.gz" \
+        -O /tmp/mdbook.tar.gz && \
+    tar -xzf /tmp/mdbook.tar.gz -C /usr/local/bin && \
+    rm /tmp/mdbook.tar.gz && \
+    cargo deny --version && \
+    cargo cyclonedx --version && \
+    mdbook --version
+
 WORKDIR /app
 
 ENTRYPOINT ["/bin/bash"]

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Duncan Tourolle
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -42,11 +42,34 @@ For the full set of build, test, and Android helper scripts, see
 |-------|----------|
 | Architecture overview & subsystem docs | [docs/architecture/](docs/architecture/) |
 | Requirements, traceability & technical debt | [docs/requirements.md](docs/requirements.md) |
-| Build & release process | [docs/build-release.md](docs/build-release.md) |
+| Build & release process | [docs/build/build-release.md](docs/build/build-release.md) |
 | Docker builds | [docs/build/docker.md](docs/build/docker.md) |
 | Traceability tooling & CI | [docs/traceability.md](docs/traceability.md), [docs/traceability-ci.md](docs/traceability-ci.md) |
 | Release checklist | [docs/release-checklist.md](docs/release-checklist.md) |
 | UX flows | [docs/ux-flows.md](docs/ux-flows.md) |
+| CI operations (builder image, secrets, runner) | [docs/build/ci-operations.md](docs/build/ci-operations.md) |
+
+## Contributing
+
+[CONTRIBUTING.md](CONTRIBUTING.md) covers the setup, the gates a change has to
+pass, and the two rules that catch people out (bug fixes start with a failing
+test; Jellyfin's taxonomy stays in Rust). Please also read the
+[Code of Conduct](CODE_OF_CONDUCT.md).
+
+Found a security problem? Do not open an issue — see [SECURITY.md](SECURITY.md).
+
+## Verifying a download
+
+Every release publishes `SHA256SUMS` covering all of its artifacts, plus an SBOM
+of what went into the build:
+
+```bash
+sha256sum -c SHA256SUMS
+```
+
+Desktop builds update themselves from Settings → Updates, verifying each payload
+against JellyTau's signing key before installing. Android installs are handled by
+the system installer, so the app links to the releases page instead.
 
 ## Recommended IDE Setup
 

SECURITY.md

@@ -0,0 +1,60 @@
+# Security Policy
+
+## Reporting a vulnerability
+
+Email **duncan@tourolle.paris** with `[JellyTau security]` in the subject.
+Please do **not** open a public issue for a vulnerability — JellyTau handles
+Jellyfin credentials and media, and an unfixed issue in a public tracker is an
+advisory for everyone running it.
+
+Include what you have: what the problem is, how to reproduce it, the version and
+platform, and what you think an attacker could do with it. A rough report is
+worth more than a polished one that never gets sent.
+
+You can expect an acknowledgement within a week. If a fix is warranted it will
+ship in the next release, and you will be credited in the release notes unless
+you would rather not be.
+
+## Supported versions
+
+JellyTau is a single-maintainer project without long-term support branches.
+**Only the latest release receives fixes.** Desktop builds can update themselves
+(Settings → Updates); on Android, install the latest APK from the releases page.
+
+## What is in scope
+
+The application and its build pipeline:
+
+- The Tauri backend (`src-tauri/`) and the Svelte frontend (`src/`)
+- Credential storage — the system keyring and its encrypted-file fallback
+- The Android player service and its JNI bridge
+- The loopback media server used for downloaded playback
+- The release pipeline: artifact signing, the update manifest, the builder image
+
+**Out of scope:** vulnerabilities in Jellyfin itself (report those to the
+Jellyfin project), and issues that require an already-compromised device or a
+malicious server the user deliberately configured and trusted.
+
+## What the project already does
+
+Not a guarantee, but so you know what has been considered:
+
+- **Credentials** never go in plaintext config: the system keyring is used where
+  available, with an AES-GCM encrypted file as fallback (see
+  [docs/architecture/09-security.md](docs/architecture/09-security.md)).
+- **The webview runs under a restrictive CSP**, and the asset protocol is scoped
+  to the thumbnail cache directory only.
+- **Path confinement** is enforced on the cache and download roots — a
+  server-supplied id cannot decide where a file lands (DR-210, DR-211).
+- **Queries and URLs bind or encode their inputs** rather than interpolating
+  them (DR-212).
+- **Dependencies are scanned on every build** by `cargo deny` against the RustSec
+  advisory database, and licence-checked against an allow-list (DR-216).
+- **Releases carry `SHA256SUMS` and an SBOM**, so you can verify a download and
+  find out what went into it.
+- **Desktop updates are signature-verified** against a key held only in CI before
+  anything is installed (DR-217).
+
+Windows installers are **not** Authenticode-signed — SmartScreen will warn on
+first run. That is a cost and identity problem, not an oversight; verify the
+download against `SHA256SUMS` instead.

docker-compose.yml

@@ -1,4 +1,4 @@
-version: '3.8'
+version: "3.8"
 
 services:
   # Test service - runs tests only
@@ -31,7 +31,7 @@ services:
     depends_on:
       - test
     ports:
-      - "5172:5172"  # In case you want to run dev server
+      - "5172:5172" # In case you want to run dev server
 
   # Linux desktop packages - deb + rpm + pacman into ./dist
   desktop-linux-build:

docs-site/SUMMARY.md

@@ -22,15 +22,35 @@
   - [Database Design](architecture/08-database-design.md)
   - [Security](architecture/09-security.md)
 
-# UX & Specs
+# UX
 
 - [UX Flows](ux-flows.md)
-- [Video Background Audio](specs/video-background-audio.md)
+
+# Specs — Pending Work
+
+- [Specs Index](specs/README.md)
+- [Spec Template](specs/SPEC-TEMPLATE.md)
+- [Spec Review Checklist](specs/SPEC-REVIEW-CHECKLIST.md)
+- [Playback Backend Unification](specs/playback-backend-unification.md)
+- [Linux Native Video Spike](specs/linux-native-video-spike.md)
+- [Player Facade Enforcement](specs/player-facade-enforcement.md)
+- [Windows Native Audio Backend](specs/windows-native-audio-backend.md)
+- [libmpv2 Migration](specs/libmpv2-migration.md)
+- [Read-Through Media Cache](specs/read-through-media-cache.md)
+- [Scoped Search](specs/scoped-search.md)
+- [Scoped Search Boundary](specs/scoped-search-boundary.md)
+- [Scoped Search Boundary — Implementation](specs/scoped-search-boundary-implementation.md)
+- [Frontend Domain Model](specs/frontend-domain-model.md)
+- [Desktop Native Video](specs/desktop-native-video.md)
+- [Build Provenance](specs/build-provenance.md)
 
 # Build & Release
 
-- [Build & Release](build-release.md)
+- [Build & Release](build/build-release.md)
 - [Release Checklist](release-checklist.md)
+- [Desktop Packaging](build/build-desktop-packages.md)
+- [Windows Build](build/build-windows.md)
+- [Defect Windows](defect-windows.md)
 - [Docker](build/docker.md)
 - [Builder Image](build/build-builder-image.md)
 

docs/architecture/01-rust-backend.md

@@ -376,57 +376,77 @@ flowchart TB
 ## Favorites System
 
 **Location**:
-- Service: `src/lib/services/favorites.ts`
-- Component: `src/lib/components/FavoriteButton.svelte`
-- Backend: `src-tauri/src/commands/storage.rs`
+- Commands: `src-tauri/src/commands/favorites.rs` (offline drain),
+  `src-tauri/src/commands/repository.rs` (query + toggle),
+  `src-tauri/src/commands/storage/` (local `user_data` writes)
+- Repository: `get_favorites` on the trait, implemented by `online.rs`,
+  `offline.rs` and `hybrid.rs`
+- Frontend: `src/lib/services/favorites.ts`,
+  `src/lib/components/FavoriteButton.svelte`, `/library/favorites`
 
-The favorites system implements optimistic updates with server synchronization:
+Favouriting has two halves that are easy to confuse: **marking** an item, which
+has existed since UR-017, and **browsing** what was marked, which arrived with
+UR-067…069 (DR-113 … DR-120). Both go through the repository, not around it.
+
+### Marking
+
+Optimistic local write, then server sync:
 
 ```mermaid
 flowchart TB
     UI[FavoriteButton] -->|Click| Service[toggleFavorite]
-    Service -->|1. Optimistic| LocalDB[(SQLite user_data)]
-    Service -->|2. Sync| JellyfinAPI[Jellyfin API]
-    Service -->|3. Mark Synced| LocalDB
-
-    JellyfinAPI -->|POST| MarkFav["/Users/{id}/FavoriteItems/{itemId}"]
-    JellyfinAPI -->|DELETE| UnmarkFav["/Users/{id}/FavoriteItems/{itemId}"]
-
-    LocalDB -->|is_favorite<br/>pending_sync| UserData[user_data table]
+    Service -->|"1. Optimistic"| LocalDB[("SQLite user_data<br/>is_favorite, pending_sync")]
+    Service -->|"2. Sync"| Repo[Repository]
+    Repo -->|POST / DELETE| JellyfinAPI["/Users/{id}/FavoriteItems/{itemId}"]
+    Service -->|"3. Mark synced"| LocalDB
+    Drain["spawn_favorites_drain<br/>(background task)"] -->|"pending_sync = 1"| Repo
 ```
 
-**Flow**:
-1. User clicks heart button in UI (MiniPlayer, AudioPlayer, or detail pages)
-2. `toggleFavorite()` service function handles the logic:
-   - Updates local SQLite database immediately (optimistic update)
-   - Attempts to sync with Jellyfin server
-   - Marks as synced if successful, otherwise leaves `pending_sync = 1`
-3. UI reflects the change immediately without waiting for server response
+1. The local row is updated immediately, so the heart fills without a round trip.
+2. The repository is asked to mark or unmark on the server.
+3. On success `pending_sync` is cleared; on failure the row stays pending.
+4. A **background drain** (`spawn_favorites_drain`, started in `lib.rs` setup)
+   retries pending rows, so a favourite marked offline still reaches the server
+   (DR-120). This is the same pattern as the sync-queue drain — see
+   [Background workers](#background-workers).
 
-**Components**:
+### Browsing
 
-- **FavoriteButton.svelte**: Reusable heart button component
-  - Configurable size (sm/md/lg)
-  - Red when favorited, gray when not
-  - Loading state during toggle
-  - Bindable `isFavorite` prop for two-way binding
+`get_favorites(scope, options)` answers "what did this user favourite", across
+libraries, with the **scope owned by Rust** — the frontend sends a
+[`SearchScope`](#search-scope-and-the-taxonomy-boundary) variant and never names
+an item type. `HybridRepository` splits it the same way it splits every query:
 
-- **Integration Points**:
-  - MiniPlayer: Shows favorite button for audio tracks (hidden on small screens)
-  - Full AudioPlayer: Shows favorite button (planned)
-  - Album/Artist detail pages: Shows favorite button (planned)
+| Method | Used for |
+|--------|----------|
+| `get_favorites_cache_only` | The instant leg — the local `user_data` join |
+| `get_favorites_server_only` | The reconciliation leg |
+| `get_favorites` | Cache-first with server merge, per the repository's usual policy |
 
-**Database Schema**:
-- `user_data.is_favorite`: Boolean flag (stored as INTEGER 0/1)
-- `user_data.pending_sync`: Indicates if local changes need syncing
+`GetItemsOptions.favorites_only` is the other entry point: it filters an
+*existing* library listing rather than starting a cross-library query (DR-116),
+which is what a library page's favourites filter uses.
 
-**Tauri Commands**:
-- `storage_toggle_favorite`: Updates favorite status in local database
-- `storage_mark_synced`: Clears pending_sync flag after successful sync
+Server favourite state is mirrored into the local `user_data` table on catalog
+sync (DR-113/DR-114), so a favourite marked in another Jellyfin client shows up
+here — before this, `MediaItem.user_data` was left empty and no query anywhere
+asked for favourites.
 
-**API Methods**:
-- `LibraryApi.markFavorite(itemId)`: POST to Jellyfin
-- `LibraryApi.unmarkFavorite(itemId)`: DELETE from Jellyfin
+**Tauri commands**:
+
+| Command | Description |
+|---------|-------------|
+| `repository_get_favorites` | Cross-library favourites for a scope |
+| `repository_mark_favorite` / `repository_unmark_favorite` | Toggle on the server, through the repository |
+| `storage_toggle_favorite` | Local optimistic write (`is_favorite`, `pending_sync`) |
+| `storage_mark_synced` | Clear `pending_sync` after a successful server write |
+
+**Frontend surfaces** (DR-117 … DR-119): the `/library/favorites` page with a
+scope selector, favourite rows on home (`favoriteMovies` / `favoriteShows` /
+`favoriteMusic` in `stores/home.ts`), a favourites tile per category in the
+library mosaic, and `FavoriteButton` mounted wherever a whole item is shown —
+movie, series, episode, album, artist and playlist detail views as well as the
+mini player.
 
 ## Player Backend Trait
 
@@ -569,3 +589,269 @@ async fn move_playlist_item(&self, playlist_id: &str, item_id: &str, new_index:
 | `player_set_autoplay_settings` | `settings: AutoplaySettings` | `AutoplaySettings` |
 | `player_get_autoplay_settings` | - | `AutoplaySettings` |
 | `player_on_playback_ended` | - | `()` |
+
+## Domain Vocabulary Owned by Rust
+
+The frontend is presentation-only and must not encode Jellyfin's *taxonomy* — the
+rule in [CLAUDE.md](../../CLAUDE.md) and
+[scoped-search-boundary.md](../specs/scoped-search-boundary.md). These are the
+places where that vocabulary actually lives.
+
+### Search scope and the taxonomy boundary
+
+**Location**: `src-tauri/src/repository/types.rs`
+
+`SearchScope` is the canonical example the boundary rule is taught from. The
+frontend sends an opaque variant; Rust expands it into Jellyfin item types:
+
+```rust
+pub enum SearchScope { All, Music, Movies, Tv }
+
+impl SearchScope {
+    /// The Jellyfin item types this scope requests, or `None` for `All`.
+    pub fn item_types(self) -> Option<Vec<String>> { … }
+
+    /// The scope a library of this Jellyfin `CollectionType` belongs to.
+    pub fn for_collection_type(collection_type: &str) -> Option<SearchScope> { … }
+}
+```
+
+Two details that are load-bearing:
+
+- `All` returns `None`, **not** the union of every listed type. An explicit
+  `includeItemTypes` list filters out anything not named in it, so a union would
+  silently drop People, folders, and any type nobody enumerated. Callers must
+  omit the filter entirely on `None`.
+- `for_collection_type` maps a Jellyfin `CollectionType` to a favourites
+  category (DR-175). It changes when *Jellyfin* renames a collection type, not
+  when the library page is redesigned — which is the test for whether something
+  belongs on this side of the boundary.
+
+⚠️ **The result side has not moved yet.** `GROUP_ITEM_TYPES` in
+`src/lib/utils/searchScope.ts` still maps result groups to item types in the
+frontend, and `check:boundary` does not match its shape. Tracked as Stage 2 of
+[scoped-search-boundary-implementation.md](../specs/scoped-search-boundary-implementation.md).
+
+### Library exclusions
+
+**Location**: `src-tauri/src/repository/exclusions.rs` (TRACES: UR-076 | DR-209)
+
+Folders the user has chosen to keep out of music browsing — a "Podcasts" folder
+inside a music library being the canonical case. Excluded **by item id**, not by
+name, in a process-wide `RwLock<Vec<String>>` restored from the database at
+startup, and applied by the repository layer to every music query (libraries,
+artists, albums, genres, search, home rows).
+
+The id is normalised (`trim`, strip `-`, lowercase) because Jellyfin writes the
+same GUID both dashed and undashed depending on the endpoint. The predecessor was
+a frontend filter matching the English string "Podcasts" — wrong in three ways at
+once, and the reason this lives in the repository.
+
+The set is process-wide rather than a field on a repository for the same reason
+as `online::STREAMING_QUALITY`: it is a preference about *this user's browsing*,
+not about a server session, so it must survive a repository being rebuilt on
+re-login.
+
+### Streaming quality ladder
+
+**Location**: `src-tauri/src/settings.rs` (TRACES: UR-074 | DR-162)
+
+`StreamingQuality` is a bandwidth ladder (`Original`, 20/10/8/4/2/1 Mbps,
+720 kbps), not a resolution picker: it exists to fit a connection, and the
+resolution cap is chosen *from* the bitrate so the encoder does not spend a small
+budget on pixels it cannot afford.
+
+| Method | Answers |
+|--------|---------|
+| `max_bitrate()` | Total bits/s (video + audio), `None` for `Original` |
+| `audio_bitrate()` | The audio share — shrinks down the ladder, so 384 kbps is not a third of the budget at the bottom |
+| `video_bitrate()` | Total minus audio, so the two together honour the ceiling |
+| `max_height()` | Resolution ceiling that suits the bitrate |
+
+The ceiling goes to `PlaybackInfo` as `MaxStreamingBitrate` **and** into the
+device profile. Sending it there — not just on the transcode URL — is what makes
+the cap real: a stream the server decides to *direct play* is served at the
+source file's own bitrate, and no URL parameter afterwards can reduce it.
+
+#### Two levels of ceiling
+
+**Location**: `src-tauri/src/repository/online.rs` (TRACES: UR-074, UR-079 | DR-226)
+
+There are two, and they are not the same thing:
+
+| | Set by | Lives until | Read via |
+|---|---|---|---|
+| **Device default** | Settings (`player_set_video_settings`) | Persisted; restored at startup | `streaming_quality()` |
+| **Per-playback override** | The in-player picker (`player_set_stream_quality`) | The next item starts playing | `playback_quality_override()` |
+
+`effective_streaming_quality()` resolves the pair — override first, else default —
+and **is the only thing stream construction may read**. Every URL builder and the
+`PlaybackInfo` negotiation go through it, for the reason the process-wide static
+existed in the first place: if the negotiation and the URL builder disagree, the
+cap leaks — the negotiation authorises a direct play the builder then never gets
+to constrain, or the reverse.
+
+> The override exists because a single global cannot express "this 4K remux needs
+> a ceiling, that podcast does not". The picker had documented itself as a "this
+> film, this connection" control since it was written, but was implemented by
+> writing the *default* — so dropping one awkward film to 2 Mbps silently capped
+> every video played afterwards for the rest of the process, with Settings still
+> showing the old value. It is cleared on every `player_play_item` /
+> `player_play_queue` / `player_play_tracks`, which is what stops it surviving
+> into an autoplayed next episode where nobody would reopen the picker.
+
+### Stream selection
+
+**Location**: `src-tauri/src/repository/stream_selection.rs`,
+`OnlineRepository::get_stream_selection` (TRACES: UR-070, UR-079 | DR-225, DR-227, DR-228)
+
+**Rust decides *what stream*. The player decides *how to deliver it*.** That line
+is the whole design. A backend with genuine adaptive selection (ExoPlayer over a
+multi-variant playlist) is left to do it; Rust chooses what to request and never
+paces bytes.
+
+`get_stream_selection` returns one self-describing `StreamSelection` in place of
+the bare URL `get_video_stream_url` used to hand out:
+
+| Field | Carries |
+|---|---|
+| `url` | What to open |
+| `transport` | `Hls` / `Progressive` / `LocalFile` — how to fetch it |
+| `playback_kind` | `DirectPlay` / `DirectStream` / `Transcode` — what the server is doing to the source |
+| `rendition` | The negotiated ceiling and codecs; `None` for a direct play, which *is* the source |
+| `available` | The quality ladder as it applies to this media source (DR-227) |
+| `needs_transcoding` | Derived from `playback_kind`, so the rule is answered once |
+
+Both enums are serde-tagged (`{"type":"hls"}`) so the frontend matches a
+discriminant rather than comparing text.
+
+> **Why `transport` exists.** `VideoPlayer.svelte` chose its loader with
+> `url.includes(".m3u8")`, in two places. Rust *built* that URL and knows exactly
+> what it is; re-deriving it downstream by substring match is a domain fact
+> reconstructed in the presentation layer — the same class of error as leaking
+> item-type taxonomy, and one that fails silently in **both** directions: a
+> progressive file served from a path containing the substring gets an HLS
+> loader, and a playlist served from a path without it does not.
+>
+> The paths that never negotiate get the same shape from Rust rather than letting
+> a caller assemble one — `media_local_selection` for a downloaded file,
+> `LiveStreamInfo.transport` for a live channel — so there is no second place
+> where a transport is decided.
+
+#### The playback-kind decision
+
+`decide_playback_kind` is a free function and pure, so every branch is testable
+from `PlaybackInfo` fixtures without a server. Order matters — the two
+client-side overrides come first, because each describes a case where the
+server's answer is right about the *file* and wrong about what this app will do
+with it:
+
+1. **Undecodable audio → `Transcode`.** Jellyfin 10.11.5 honours a
+   DirectPlayProfile's container and video codec but *ignores its audio codec*,
+   so it offers direct play for an E-AC-3 track the webview renders in silence.
+   A silent direct play is worse than a transcode.
+2. **A pinned audio track → `Transcode`.** Not a defect in the server's answer, a
+   different question: the file has one default track and the viewer asked for
+   another.
+3. Otherwise `supports_direct_play` → `DirectPlay`, else `supports_direct_stream`
+   → `DirectStream`, else `Transcode`.
+
+A direct **stream** is a remux — codecs copied, container repackaged. It is cheap
+and is deliberately *not* counted as transcoding; conflating the two would report
+a free passthrough as a server-side re-encode.
+
+> **What this is worth, measured.** Against the development server (Jellyfin
+> 10.11.5), 400 items sampled for codec mix and 40 put through a real negotiation
+> per profile:
+>
+> | Profile | Direct play |
+> |---|---|
+> | Linux / WebKitGTK (`h264` only, 2ch) | 3/40 — **7%** |
+> | Android / ExoPlayer (`h264,hevc,vp8,vp9,av1,mpeg4` + `ac3,eac3`, 6ch) | 34/40 — **85%** |
+>
+> The library is ~80% hevc (`hevc+eac3` alone is a third of it), which is why the
+> two diverge so hard.
+>
+> **Read that 85% as a ceiling, not a result.** It was measured with a profile
+> containing `ac3,eac3`. The Android device this was later run on reports neither
+> in its `MediaCodecList` — no Dolby licence, which is normal for a tablet — so
+> eac3 content, about a third of the sampled library, correctly transcodes there.
+> What any given device achieves depends on its own codec list, and on the
+> profile being derived from the renderer at all (DR-234), which it was not when
+> the figure was taken.
+>
+> **The payoff is still overwhelmingly Android**, because that is where a real
+> decoder is already doing the work. Linux stays near 7% until libmpv decodes the
+> picture — the h264-only profile is a WebKitGTK constraint, not a JellyTau
+> choice, and is what `linux-native-video-spike.md` exists to remove. A reviewer
+> should not expect this code to fix Linux on its own.
+
+#### The quality ladder per source
+
+`quality_options_for_source(source_bitrate)` returns every rung, each marked with
+`exceeds_source`: true when that rung's ceiling is at or above what the source
+itself carries, so selecting it produces the same bytes as `Original`. The
+frontend draws the list and drops the redundant rungs; it does not decide which
+they are.
+
+- `Original` is never marked — it *is* the source.
+- An unreported source bitrate (some containers have none; the sampled library
+  has `avi` files with no bitrate at all) marks **nothing** redundant, keeping
+  every rung offered. That is the safe direction: the viewer keeps every choice.
+
+#### No adaptive ladder to preserve
+
+**TRACES: UR-079 | DR-229 (Won't Do)**
+
+Mid-playback re-negotiation on throughput was scoped and dropped on measurement.
+A master playlist from this server carries exactly **one** `EXT-X-STREAM-INF`:
+Jellyfin builds it from the single rendition the request asked for rather than
+publishing a ladder. So there is no adaptation for hls.js to be preserving and
+none that mpv would lose — the claim that there was is recorded in
+`playback-backend-unification.md` and does not hold. "Adapt mid-stream" collapses
+into "pick well at open", which is what the two levels of ceiling and the
+per-source ladder already are.
+
+Kept here because it is a measurement, not an opinion: a server that *does*
+publish a ladder would change the answer, and the re-negotiation path below is
+the hook that work would build on.
+
+#### Re-negotiation
+
+One mechanism, not two. `player_seek_video`, `player_switch_audio_track` and
+`player_set_stream_quality` all return a tagged `strategy` saying who reloads —
+the backend handles a native backend itself and hands the webview a
+`StreamSelection` for `reloadSource`. Note the wire wart: tauri-specta keeps
+these response fields snake_case (`seek_offset`), while the `strategy` tag itself
+is camelCase.
+
+The frontend names a variant and nothing else; the labels the picker shows are
+served over IPC — from `available` on the selection, or
+`player_get_streaming_qualities` for the Settings list.
+
+## Background workers
+
+Three long-lived tasks are spawned from the Tauri `setup` hook in `lib.rs`. All
+three exist because *when* something happens is a backend policy, not something
+a page load should decide.
+
+| Worker | Location | Responsibility |
+|--------|----------|----------------|
+| `spawn_catalog_indexer` | `commands/catalog.rs` | Keeps the local FTS5 catalog fresh (DR-109, IR-030) |
+| `spawn_favorites_drain` | `commands/favorites.rs` | Retries favourite toggles made while offline (DR-120) |
+| `spawn_sync_queue_drain` | `commands/sync_drain.rs` | Drains the offline mutation queue (DR-131) |
+
+### Catalog indexer
+
+Replaces the frontend's startup-only `syncCatalog()` call. It ticks on
+`CATALOG_INDEX_TICK` and runs a pass when three things hold: a repository exists,
+the server is reachable, and the index is due per `index_is_due`. A tick is
+nearly free — one indexed `app_settings` lookup — which is what makes it
+responsive to events it cannot subscribe to, such as signing in: a fresh install
+would otherwise sit unindexed until the next scheduled pass.
+
+`index_is_due` treats both "never indexed" and an unparseable stored timestamp as
+due; a corrupt timestamp should trigger a re-index, not silently freeze the
+catalog. A failed pass is never fatal — it leaves the existing index in place and
+warns. Progress is emitted on `CATALOG_INDEX_EVENT` for the staleness hint in the
+UI.

docs/architecture/02-svelte-frontend.md

@@ -538,6 +538,14 @@ sequenceDiagram
 
 ## Auto-Play Episode Limit
 
+> ⚠️ **Autoplay is season-bounded.** `player/mod.rs:fetch_next_episode_for_item`
+> does not cross a season boundary, so autoplay stops at the end of a season even
+> though the "More Episodes" strip runs past it. 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](../../CLAUDE.md)) — its own change, not a drive-by.
+
+
 **Location**: `src-tauri/src/player/mod.rs`, `src-tauri/src/player/autoplay.rs`, `src-tauri/src/settings.rs`
 
 **TRACES**: UR-023 | DR-049
@@ -657,3 +665,219 @@ The playlist UI provides full CRUD operations for Jellyfin playlists with offlin
 All playlist mutations are queued for offline sync:
 - `queuePlaylistCreate`, `queuePlaylistDelete`, `queuePlaylistRename`
 - `queuePlaylistAddItems`, `queuePlaylistRemoveItems`, `queuePlaylistReorderItem`
+
+## App Shell and Chrome
+
+**Location**: `src/lib/utils/layoutShell.ts` (pure rules),
+`src/lib/components/AppHeader.svelte`,
+`src/lib/components/account/AccountMenu.svelte`, `BottomUi.svelte`
+**TRACES**: UR-054 | DR-075, DR-076, DR-077
+
+Account actions used to be reachable **only from `/library/*`** — the header
+that hosted them belonged to the library layout, the bottom nav offered Home /
+Search / Library, and the desktop username was inert text. From `/`, `/search`
+or `/downloads` there was no route to Settings or Sign out at all. The header is
+now shared and rendered from the root layout.
+
+### Visibility rules
+
+All four rules are pure functions in `layoutShell.ts`, so the contract is
+unit-testable rather than a scattering of `$derived` booleans that drift per
+route and platform (which is what they were):
+
+| Function | Rule |
+|----------|------|
+| `showBottomNav` | Every authenticated route except `/player/*` and `/login` |
+| `showGlobalMiniPlayer` | Everything except `/player/*`, `/login`, `/settings`. **Not** gated on platform or `/library` — the root owns the mini player everywhere, so the library route must never render a second one |
+| `routeOwnsLayout` | `/library`, `/player/`, `/login` render their own full-height flex column; everything else renders into the root scroller |
+| `showGlobalHeader` | Authenticated, not a layout-owning route, not `/settings` (the user is already there) |
+
+### The structural fix worth not undoing
+
+The "last row hidden behind the nav" bug is solved **structurally, not by
+measurement**: the bottom UI is an in-flow flex child *below* the scroller
+(`BottomUi.svelte`), so the scroller is physically bounded above it and cannot
+render behind it. There is no measurement and no reserved padding. If you
+restructure the shell, preserve the scroll containment — reintroducing padding
+math reintroduces the bug.
+
+### AccountMenu
+
+One component for both breakpoints, anchored to the username/avatar (a real
+button with `aria-expanded`, not a bare three-dot icon). Fixed item order:
+identity block (user + server) → Downloads, Settings, Display → divider → Sign
+out, destructive and last. Dismissal is backdrop click, `Escape`, and focus
+return to the trigger.
+
+The identity block falls back to the bare host of the server URL when the server
+has no human-readable name, so it always shows *something* server-identifying.
+
+Settings' Display section and the library page-header toggle are two views onto
+the **same** persisted `viewMode` store (`jellytau-view-mode`) — no second state,
+no migration, and they stay in sync for free.
+
+## Library Mosaic
+
+**Location**: `src/lib/components/library/libraryMosaic.ts` (pure),
+`MosaicGrid.svelte`, `MosaicTile.svelte`
+**TRACES**: UR-075, UR-067 | DR-174, DR-175
+
+The library overview and the home "Your Libraries" strip are a **mosaic**, not a
+grid: rows share one height and each tile is as wide as its own artwork is, so 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 a
+grid picked.
+
+`libraryMosaic.ts` is deliberately pure — it takes the libraries and returns the
+tiles to draw, so ordering and de-duplication are unit-testable rather than
+buried in markup. Tiles start at an *assumed* aspect (square, 16:9) and a
+measured image overrides it in `MosaicGrid`.
+
+Note what this file does **not** decide: which favourites category a library
+belongs to. That is Jellyfin vocabulary and arrives on the library itself as
+`favoritesScope`, from `SearchScope::for_collection_type` in Rust (see
+[01-rust-backend.md](01-rust-backend.md#search-scope-and-the-taxonomy-boundary)).
+The frontend only decides what to *call* it and where to put it.
+
+## Series and Episode Navigation
+
+**Location**: `src/lib/components/library/` — `SeasonSection.svelte`,
+`EpisodeFocusView.svelte`, `episodeStrip.ts` (pure)
+**TRACES**: UR-062 … UR-064 | DR-101 … DR-107
+
+Opening a series lands the viewer where they actually are in it. **"Where is this
+viewer in this series" is resolved in Rust** (DR-101), not by the page: the
+series detail page asks the repository and anchors on the answer — the current
+season expanded, the current episode highlighted and scrolled into view, and a
+hero button labelled `Resume S2E4` / `Play S1E1`.
+
+A season is not a destination: `/library/<seasonId>` redirects to its series
+(DR-103). Video library routes collapse to one per library (DR-105).
+
+`episodeStrip.ts` holds the pure logic for the "More Episodes" strip, extracted
+from the component because it had three distinct bugs that markup made
+untestable: the strip collapsing to just the current episode while real siblings
+existed, number-less episodes all matching as "current" (`undefined ===
+undefined`), and the window dead-ending at a season boundary instead of running
+past it. It matches by id first and only falls back to season+episode number when
+both numbers are known on both sides.
+
+## Downloaded Browse
+
+**Location**: `src/lib/services/downloadedCatalog.ts`,
+`src/lib/components/downloads/DownloadedBrowse.svelte`
+**TRACES**: UR-055, UR-056 | DR-081 … DR-085
+
+`/downloads` is two views: **Downloaded** (the default) — the library filtered to
+what is on the device, reusing the same grids, cards and detail pages as online
+browsing — and **Transfers**, the in-flight progress rows demoted to a secondary
+tab.
+
+`downloadedCatalog` reads the **offline-only** browse path on the repository,
+never the hybrid merge. That is the point: an empty result means "nothing
+downloaded here", never "server unreachable", so the view is authoritative
+regardless of connectivity. It also owns disk usage — a per-item/container byte
+map plus the device total, aggregated by the backend from `downloads.file_size`
+(DR-085).
+
+## Safe-area Insets
+
+**Location**: `src/app.css`, `WindowInsetsBridge.kt`
+**TRACES**: UR-066 | DR-112, IR-031
+
+The Android WebView does not reliably report system-bar insets through
+`env(safe-area-inset-*)`. Native `WindowInsets` (`systemBars() |
+displayCutout()`) are therefore pushed in as CSS custom properties, and every
+edge takes the larger of the two sources:
+
+```css
+--safe-top: max(env(safe-area-inset-top, 0px), var(--jt-inset-top, 0px));
+```
+
+Two rules keep this from going wrong: **one owner per edge** (two components both
+padding the top edge double-pads it), and **no nested `h-screen`** — a full-height
+child inside a full-height parent that has already consumed the inset overflows
+by exactly the inset.
+
+Unlike `addJavascriptInterface`, the inset push only writes CSS properties, so it
+can safely be re-sent on resume.
+
+## Stream Transport
+
+**Location**: `src/lib/player/streamTransport.ts`
+**TRACES**: UR-079 | DR-225 | UT-214
+
+`videoLoaderFor(selection, capabilities)` picks the loader for the webview
+`<video>` element — `hlsjs`, `nativeHls`, or `direct` — from the backend's tagged
+`selection.transport`. `elementSrcFor` is its template companion: the element's
+`src` is emptied only when hls.js is driving it.
+
+The split is the point. **The transport is the stream's property and comes from
+Rust; whether a given loader exists is the browser's, and is the only thing
+decided here.**
+
+> This replaced `currentStreamUrl.includes(".m3u8")`, which appeared twice in
+> `VideoPlayer.svelte` — once in the HLS `$effect` and once inline in the
+> template's `src`. Rust builds that URL and knows what it is; re-deriving it
+> here by substring match was a domain fact reconstructed in the presentation
+> layer, and it fails silently in both directions. The two tests that pin it are
+> the ones that failed against the old implementation: a `progressive` stream
+> whose URL contains `.m3u8` must **not** get an HLS loader, and an `hls` stream
+> whose URL contains no `.m3u8` must.
+>
+> Logic lives in a plain `.ts` module rather than in the component for the usual
+> reason — it is testable there. Same pattern as `episodeStrip.ts`.
+
+`VideoPlayer` holds a `currentSelection`, not a URL string; `currentStreamUrl` is
+derived from it. A reload replaces the selection **wholesale** (the adapter's
+bridge takes a `StreamSelection`, not a URL), so transport and URL can never
+drift apart. The background-audio handoff states the transport it is moving to —
+progressive mp3 out, HLS back — via `selectionAt()`, rather than leaving it to be
+inferred.
+
+The quality picker is filled from `selection.available` (DR-227): rungs the
+backend marked `exceedsSource` are not drawn, because they produce the same bytes
+as `Original`. Nothing is optimistically assigned when the viewer picks a rung —
+what the menu shows comes from the selection the backend hands back, since a
+ceiling above the source bitrate *is* the source.
+
+## Native Video Store
+
+**Location**: `src/lib/stores/nativeVideo.ts`
+**TRACES**: UR-003, UR-004 | DR-188
+
+Two separate concerns live here, deliberately:
+
+- `experimentalNativeVideo` — the user-facing opt-in flag, **defaulting to on**.
+  Rust already decides *which backend this platform has* (`useHtml5Element` from
+  `player_play_item`); this flag only *suppresses* that decision. It never turns
+  native on where Rust says HTML5. An explicit stored choice wins in both
+  directions, so someone who opted out is not re-enabled by a default flip —
+  hence the `null` check rather than a bare `=== "true"`.
+- `nativeVideoActive` — whether a native surface is on screen *right now*.
+  Setting it toggles `data-native-video` on `<html>`, which is what the CSS in
+  `app.css` keys off to clear the app's opaque backgrounds. It is deliberately
+  **not** derived from the flag: the backgrounds must come back the moment the
+  player unmounts.
+
+See [05-platform-backends.md](05-platform-backends.md#native-video-compositing-android)
+for what is behind the WebView.
+
+## Logging
+
+**Location**: `src/lib/utils/logger.ts`
+**TRACES**: DR-204
+
+The frontend's equivalent of the Rust `log` crate: four levels
+(`debug < info < warn < error`), a compile-environment default (dev → `debug`,
+production → `warn`), and a runtime override that is the moral equivalent of
+`RUST_LOG`. Scoped loggers carry the subsystem in the message, so a filtered
+console stays usable while a player, a download worker and a store are all
+talking.
+
+Production deliberately keeps **warn and error**: this is a client talking to a
+server that may or may not be there, and a silent failure is worse to support
+than a noisy console. Only the chatter is suppressed.
+
+`no-console` is an ESLint **error**, with the sink module itself the only
+exception, so a raw `console.*` cannot re-appear.

docs/architecture/03-data-flow.md

@@ -49,6 +49,61 @@ sequenceDiagram
 - Background cache updates (planned)
 - **Connectivity side-effect**: each server request feeds the `ConnectivityMonitor`, which is the source of truth for the offline/online banner (see [07-connectivity.md](07-connectivity.md)). A server-answered error (401/404/5xx) still counts as *reachable* — only network failures, sustained past a debounce window, flip the app to offline.
 
+## Search Flow (Locally Indexed)
+
+**TRACES**: UR-065 | DR-108 … DR-111, IR-030
+
+Search does not depend on a per-keystroke round trip to Jellyfin. The instant leg
+reads the **local SQLite catalog**, which is already synced and already
+FTS5-indexed, so results appear as fast as SQLite can answer — online or offline.
+The server query stays, demoted to a background reconciliation that merges in
+late results.
+
+```mermaid
+sequenceDiagram
+    participant UI as Search UI
+    participant Rust as repository_search
+    participant Cache as Local catalog (FTS5)
+    participant Server as Jellyfin
+    participant Indexer as spawn_catalog_indexer
+
+    UI->>Rust: search(query, scope)
+    Rust->>Cache: FTS5 query, scope expanded by SearchScope::item_types()
+    Cache-->>UI: instant results
+    Rust->>Server: reconciliation query (background)
+    Server-->>UI: search-event with late/merged results
+    Note over Indexer,Cache: Independent of any query:<br/>scheduled crawl keeps the index fresh,<br/>prunes items deleted on the server
+```
+
+**Key points:**
+
+- The **scope is opaque on the wire**. The frontend sends a `SearchScope`
+  variant; Rust expands it to item types
+  ([01-rust-backend.md](01-rust-backend.md#search-scope-and-the-taxonomy-boundary)).
+- **Index freshness is a Rust policy**, not a frontend startup call — a scheduled
+  background pass, not "whatever was synced when the app last launched"
+  (DR-109). See
+  [Background workers](01-rust-backend.md#background-workers).
+- **Index hygiene matters as much as freshness**: the catalog save path uses
+  `INSERT OR REPLACE` and the crawl prunes rows for content deleted on the
+  server, or search keeps returning items that no longer exist (DR-110).
+- The index covers **exactly the types the result groups render** (DR-111) —
+  including Artists, which the crawl must reach or the Artists group is silently
+  always empty.
+
+**Deliberately not done, with reasons:**
+
+- **Incremental indexing** (Jellyfin's `MinDateLastSaved`). A *full* crawl is
+  what makes the deletion sweep sound — it yields the authoritative id set per
+  library, and an incremental pass cannot detect deletions. Worth revisiting if
+  full crawls prove slow on large libraries; measure first.
+- **Removing the server leg.** The reconciliation query stays.
+
+> ⚠️ Two dead search implementations still exist: `storage_search_items`
+> (`commands/storage/mod.rs`) and `offline_search` (`commands/offline.rs`). Both
+> are registered in `lib.rs` and exported to `bindings.ts`; neither is called
+> from the frontend. Deleting them is correct and unclaimed.
+
 ## Playback Initiation Flow
 
 ```mermaid
@@ -77,6 +132,55 @@ sequenceDiagram
     Note over Store: UI updates reactively
 ```
 
+## Video Stream Selection Flow
+
+**TRACES: UR-070, UR-079 | DR-225, DR-227, DR-228**
+
+Before a video plays, Rust decides *what stream* — direct play, remux or
+transcode, over which transport — and hands the player one self-describing
+`StreamSelection`. The page no longer inspects the URL to work any of this out.
+
+```mermaid
+sequenceDiagram
+    participant Page as player/[id]/+page.svelte
+    participant Repo as HybridRepository
+    participant Online as OnlineRepository
+    participant Server as Jellyfin
+    participant VP as VideoPlayer.svelte
+
+    Page->>Repo: playerLocalMediaPath(id)
+    alt a completed download exists
+        Page->>Repo: mediaLocalSelection(path)
+        Note over Page: LocalFile / DirectPlay, no ladder —<br/>nothing about a file on disk re-negotiates
+    else stream from the server
+        Page->>Repo: getStreamSelection(id, mediaSourceId)
+        Repo->>Online: get_stream_selection()
+        Online->>Online: effective_streaming_quality()
+        Note over Online: per-playback override, else device default
+        Online->>Server: POST /Items/{id}/PlaybackInfo<br/>(device profile + ceiling)
+        Server-->>Online: MediaSource {supportsDirectPlay,<br/>supportsDirectStream, transcodingUrl, bitrate}
+        Online->>Online: decide_playback_kind()
+        alt Transcode
+            Online->>Online: adopt/stop prior play session,<br/>build HLS URL
+            Note over Online: Transport::Hls
+        else DirectPlay / DirectStream
+            Online->>Online: /Videos/{id}/stream?static=true
+            Note over Online: Transport::Progressive,<br/>rendition = None (it IS the source)
+        end
+        Online->>Online: quality_options_for_source(bitrate)
+        Online-->>Page: StreamSelection
+    end
+    Page->>VP: selection
+    VP->>VP: videoLoaderFor(selection, caps)
+    Note over VP: hls.js / native HLS / direct —<br/>from the tag, never from the URL
+```
+
+The selection travels with the stream from then on. A reload — a quality change,
+an audio-track switch, a transcoded seek — returns a *new* selection through the
+same tagged `strategy` response, so transport and URL can never disagree; and the
+queue item carries the transport so `player_seek_video` picks its seek strategy
+from the backend's decision rather than from the URL string.
+
 ## Playback Mode Transfer Flow
 
 ```mermaid

docs/architecture/05-platform-backends.md

@@ -247,6 +247,184 @@ pub extern "system" fn Java_com_dtourolle_jellytau_player_JellyTauPlayer_nativeO
 }
 ```
 
+### Audio settings on ExoPlayer
+
+**TRACES**: UR-027, UR-032, UR-033 | DR-030, DR-035, DR-036
+
+`PlayerBackend` declares `set_audio_settings` with a default `Ok(())` body. For a
+long time `ExoPlayerBackend` took that default, so Settings › Audio rendered
+controls that silently did nothing on Android — the parity gap recorded in
+[requirements.md](../requirements.md#platform-playback-backend-parity-linux-vs-android),
+now closed.
+
+The settings cross to Kotlin as **JSON over JNI**, not as a wide signature, so new
+fields do not change the method signature — the same approach `load()` uses for
+subtitles:
+
+```rust
+fn set_audio_settings(&mut self, settings: &AudioSettings) -> Result<(), PlayerError> {
+    let json = audio_settings_jni_payload(settings)?;
+    env.call_method(&self.player_ref, "setAudioSettings", "(Ljava/lang/String;)V", …)?;
+    // Store the sanitised form, so audio_settings() reflects what was applied.
+    self.shared_state.lock_safe().audio_settings =
+        settings.clone().with_crossfade_clamped().with_equalizer_normalised();
+}
+```
+
+Kotlin owns the *mechanics* — attaching `AudioEffect`s to the audio session — while
+the canonical band layout and preset curves stay in Rust:
+
+| Feature | Android mechanism | Notes |
+|---------|-------------------|-------|
+| Gapless | `pauseAtEndOfMediaItems` | |
+| Volume normalization | `LoudnessEnhancer` | A gain stage — approximate next to MPV's `dynaudnorm` |
+| Equalizer | `android.media.audiofx.Equalizer` | The canonical 10 bands are resampled onto the device's own band centres |
+| Crossfade | — | Unimplemented on **every** platform (DR-034), architecturally blocked on MPV. Building it on Android alone would invert the parity gap |
+
+Two things are deliberately still open: the effects are **not yet verified on a
+physical device** (`AudioEffect` availability and band layouts are device-specific),
+and the trait default is still a silent `Ok(())` rather than an error, so a backend
+that omits the method still reports success. Flipping that default waits on the
+device verification.
+
+### The equalizer, and where its vocabulary lives
+
+**TRACES**: UR-027 | DR-030, IR-020
+
+The canonical band layout (`EQ_BANDS`) and the preset curves live in
+`settings.rs`, **not** in either backend and not in the UI: a preset *is* a gain
+curve defined by the band layout, and the layout is a property of the audio
+engine rather than of the picker that renders it. Presets are Flat, Rock, Pop,
+Jazz, Classical, Bass Boost, Treble Boost and Vocal, all conservative (within
+±8 dB) so they stack safely with volume normalization.
+
+| Platform | Mechanism |
+|----------|-----------|
+| Linux | One ffmpeg two-pole peaking `equalizer` filter per band, composed by `build_af_filter` into MPV's `af` property alongside the normalization filter: `equalizer=f=31:width_type=o:width=1:g=5` |
+| Android | `android.media.audiofx.Equalizer`, with the canonical 10 bands **resampled onto whatever band centres the device actually has** |
+
+Gains are normalised (`with_equalizer_normalised`) before use, and bands beyond
+`EQ_BANDS` are ignored, so a malformed settings payload cannot produce a filter
+chain of unbounded length.
+
+## Background Audio Handoff (Android)
+
+**TRACES**: UR-040 | IR-025, DR-051, DR-052, DR-178 … DR-180, DR-196, DR-203
+
+Keeping a video's **audio** alive when the app is backgrounded or the screen
+locks, while video decode stops. Two verified facts drive the whole design:
+
+1. An Android WebView `<video>` **does not** keep playing audio once the app is
+   backgrounded — the system throttles the WebView and media pauses.
+2. Keeping audio alive in the background requires a **native foreground media
+   service**, which already exists for music (`JellyTauPlaybackService` +
+   `JellyTauPlayer` + `MediaSessionCompat`).
+
+So this is a **handoff**, not "keep the WebView alive": on background, tear down
+the current renderer and play the same item audio-only through the native
+service; on foreground, hand back. In the project's one-directional playback
+model this is a change of *which player is authoritative*, and the position must
+transfer cleanly across it.
+
+```mermaid
+sequenceDiagram
+    participant App as App backgrounded
+    participant FE as VideoPlayer
+    participant Rust as player_enter/exit_background_audio
+    participant Exo as Native audio service
+
+    App-->>FE: jellytau-background (DOM CustomEvent)
+    FE->>Rust: enter(item, position, audioStreamIndex)
+    Rust->>Exo: play audio-only at position
+    Note over Exo: lockscreen + notification, existing MediaSession
+    App-->>FE: jellytau-foreground
+    FE->>Rust: exit() -> final position
+    Rust-->>FE: position
+    FE->>FE: restart the renderer that is on screen
+```
+
+Details that were each a shipped defect:
+
+- **Position is absolute.** Transcoded HLS tracks time as
+  `videoElement.currentTime + seekOffset` (the element resets to 0 after each
+  transcode reload). `computeHandoffPosition` sums both terms; using the element
+  time alone rewinds by the offset.
+- **A downloaded episode takes no base URL and an ordinary seek** (DR-180); a
+  stream takes the base and no seek; a handoff at 0:00 takes neither.
+- **The return must restart the renderer that is actually on screen** (DR-196).
+  The two paths resume by different means — the webview `<video>` reloads off its
+  stream URL, watched by an `$effect`; ExoPlayer owns no element and nothing
+  watches the URL for it, so it needs an explicit re-issue. Doing only the URL
+  assignment restarted nothing on the native path and left a black screen with a
+  play button that did nothing.
+- **`wasPlaying` is captured on the way out** so play/pause survives the round
+  trip, and the handoff does not silently rewind (DR-203).
+- **Mutually exclusive with PiP.** Toggle on → `setAutoEnterEnabled(false)`;
+  toggle off → PiP on background, the status quo. The frontend re-asserts the
+  value whenever the toggle changes and on unmount, so a stale setting cannot
+  leak into the next player.
+- The pure arithmetic and state transitions live in
+  `backgroundAudioHandoff.ts`, free of Svelte and the DOM, so they are testable
+  without mounting the player.
+
+Native signals background/foreground to the frontend as DOM CustomEvents
+(`jellytau-background` / `jellytau-foreground`); the frontend carries the toggle
+state to native through the `AndroidBackgroundAudio` bridge. No-op on every
+non-Android platform.
+
+## Native Video Compositing (Android)
+
+**TRACES**: UR-003, UR-004 | DR-150 … DR-152, DR-182 … DR-196
+
+Android can render video on the **native ExoPlayer surface behind a transparent
+Tauri WebView**, with the Svelte controls drawn over it. This is on by default;
+the HTML5 `<video>` path remains the fallback and is not being removed. The
+default has been flipped and reverted twice and each revert has a named cause —
+the per-defect record is in `requirements.md` (DR-150 … DR-196).
+
+```mermaid
+flowchart TB
+    subgraph Window["One Android window"]
+        Texture["TextureView (index 0)<br/>ExoPlayer video"]
+        WebView["Tauri WebView (above)<br/>transparent, Svelte controls"]
+    end
+    Rust["ExoPlayerBackend"] -->|JNI| Player["JellyTauPlayer"]
+    Player --> Texture
+    MainActivity -->|"setTransparent(true)"| WebView
+    VideoOverlayManager -->|"attach / detach"| Texture
+```
+
+Load-bearing details, each of which was a shipped defect:
+
+- **TextureView, not SurfaceView** (DR-192). A SurfaceView renders on its own
+  layer *outside* the app window and punches a transparent hole through it;
+  everything drawn above that hole — for us the whole UI — depends on that
+  composition path, which Android's own documentation says does not reliably
+  work. A TextureView makes "behind" ordinary view z-order within one window.
+- **Attached at index 0** by `VideoOverlayManager`, and **detached when the video
+  goes** (DR-184) — a surface left in the hierarchy outlives its player.
+- **Bridges are installed before the page that uses them** (DR-183).
+  `addJavascriptInterface` must run once per WebView instance and a call that
+  lands after the page has loaded never reaches it, so `setTransparent(true)`
+  could be dropped entirely.
+- **The app shell stops painting over the surface** (DR-185). `app.css` clears
+  its opaque backgrounds off `[data-native-video]`; before that, a CSS rule
+  targeted an attribute nothing ever set, so the fix looked applied and was not.
+- **The poster card can lift on a path with no `<video>` element** (DR-182) — the
+  native reveal fires on a `playing` state or a position tick carrying a position
+  or duration, and on nothing else.
+- **Letterbox bars are painted**, not left holding whatever was last in the
+  framebuffer (DR-194).
+- There is deliberately **no audio-focus bridge**: manual focus requests from the
+  WebView competed with Chromium's `AudioFocusDelegate` and with ExoPlayer, and
+  the resulting `AUDIOFOCUS_LOSS` paused playback.
+
+Related Kotlin pieces in the same window: `PictureInPictureManager` (DR-160/161),
+`ScreenWakeManager` (DR-202 — Android counts its display timeout from touch
+events, which a playing video does not generate), `ImmersiveModeBridge` and
+`WindowInsetsBridge` (IR-031/DR-112 — see
+[02-svelte-frontend.md](02-svelte-frontend.md#safe-area-insets)).
+
 ## Android MediaSession & Remote Volume Control
 
 **Location**: `JellyTauPlaybackService.kt`

docs/architecture/06-downloads-and-offline.md

@@ -139,9 +139,56 @@ flowchart TB
     CheckStorage -->|"OK"| Download["Queue Download"]
 ```
 
+## One Storage Model: Cache Entries Are Downloads
+
+**TRACES**: UR-071 | DR-126, DR-127
+
+A cache entry **is** a download with a shorter life: the same `downloads` row and
+the same file handling, distinguished by `download_source` plus an expiry. There
+is one storage model rather than a cache and a download library that can
+disagree about what is on disk.
+
+| `download_source` | Life | Reclaimed by |
+|-------------------|------|--------------|
+| `'auto'` (temporary) | Expiry, or eviction under space pressure | Both |
+| `'user'` (permanent) | No expiry | Neither |
+
+**Eviction only reclaims the temporary tier.** `evict_lru_async` originally
+selected every completed download ordered by `completed_at ASC` with no source
+filter, so hitting the 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 is load-bearing: rows predating the
+migration 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".
+
+A temporary row can be **promoted** to permanent when the user chooses to keep
+it. That only clears the expiry and flips the source; the bytes never move.
+
+## Offline Catalog Visibility
+
+**TRACES**: UR-052 | DR-078, DR-079, DR-080
+
+Offline, a library page shows **only media on the device**. A "Show all server
+media" toggle additionally reveals the cached server catalog, greyed out and
+queueable for download on reconnect.
+
+The gate is a process-global `INCLUDE_CATALOG_BROWSE` in
+`repository/offline.rs`, written by the `set_show_server_catalog` command. It
+gates the synced-catalog leg of `get_items`; without it the toggle rendered but
+every server item still appeared, which is the defect the spec was written for.
+`isConnected` derives from backend-reported reachability alone (DR-079) — see
+[07-connectivity.md](07-connectivity.md).
+
+Per-item disk usage comes from `repository_get_download_disk_usage`
+(`DownloadDiskUsage`), aggregated from `downloads.file_size` — used by the
+Downloaded browse cards, detail pages, the device total and the remove
+confirmation (DR-085).
+
 ## Download Commands
 
-**Location**: `src-tauri/src/commands/download.rs`
+**Location**: `src-tauri/src/commands/download/` — `mod.rs` (the commands below), `pinning.rs`, `smart_cache.rs`
 
 | Command | Parameters | Description |
 |---------|------------|-------------|
@@ -152,6 +199,11 @@ flowchart TB
 | `resume_download` | `download_id` | Resume paused download |
 | `cancel_download` | `download_id` | Cancel and delete partial |
 | `delete_download` | `download_id` | Delete completed download |
+| `download_video` / `download_series` / `download_season` | item ids | Queue video content |
+| `get_download_storage_stats` | `user_id` | Device totals for the downloads screen |
+| `delete_album_downloads` / `delete_downloads_under` / `delete_all_downloads` | container id | Bulk removal |
+| `pin_item` / `unpin_item` / `is_item_pinned` | `item_id` | Protect metadata from a cache clear |
+| `set_max_concurrent_downloads` | `max` | Worker concurrency (3 by default) |
 
 ## Offline Commands
 

docs/architecture/09-security.md

@@ -50,6 +50,68 @@ pub struct EncryptedFileStorage; // AES-256-GCM fallback
 | Certificate Validation | System CA store (configurable for self-signed) |
 | Token Transmission | Bearer token in `Authorization` header only |
 | Token Refresh | Handled by Jellyfin server (long-lived tokens) |
+| Android cleartext | `res/xml/network_security_config.xml` blocks cleartext everywhere except `127.0.0.1` (the loopback media server, DR-137/DR-138). The manifest's `usesCleartextTraffic` is ignored once the config is present, so the config is the single authority |
+| Android WebView | `mixedContentMode = COMPATIBILITY` with `allowFileAccess`/`allowContentAccess` both `false` (DR-199). These are the second half of the cleartext policy: `ALWAYS_ALLOW` re-opened by hand what the network security config closes. Change the two together |
+
+## Webview Content Security Policy
+
+`app.security.csp` in `tauri.conf.json` (TRACES: UR-012, UR-071 | DR-198). It was
+`null` — CSP disabled — which meant any script that reached the web layer
+inherited the full IPC surface. Tauri computes the header from this value when it
+serves the embedded HTML, injecting a nonce for SvelteKit's inline bootstrap
+script, so `script-src` needs no `'unsafe-inline'`.
+
+```
+default-src 'self';
+script-src  'self';
+style-src   'self' 'unsafe-inline';
+font-src    'self' data:;
+img-src     'self' data: blob: asset: http://asset.localhost http: https:;
+media-src   'self' blob: asset: http://asset.localhost http://127.0.0.1:* http: https:;
+connect-src 'self' ipc: http://ipc.localhost http: https:;
+worker-src  'self' blob:;
+object-src 'none'; frame-src 'none'; base-uri 'self'; form-action 'self'; frame-ancestors 'none'
+```
+
+| Directive | Why |
+|-----------|-----|
+| `default-src 'self'` | Everything not named below is same-origin only. |
+| `script-src 'self'` | The genuinely restrictive half. Bundled JS only; Tauri's build-time nonce covers the one inline `<script>` in `index.html`. Adding `'unsafe-inline'` here would silently do nothing anyway — a nonce in a directive voids it. |
+| `style-src 'self' 'unsafe-inline'` | Svelte compiles `style="…"` attributes into markup, including `app.html`'s `display: contents` wrapper, and CSP treats a style *attribute* as inline. Safe only while no `<style>` **element** survives into `index.html`: Tauri would nonce it, and the nonce would then void `'unsafe-inline'`. The production build extracts all CSS to files, so it currently has none. |
+| `img-src` | Thumbnails come from two places: the asset protocol (`asset://localhost/…` on Linux/macOS, `http://asset.localhost/…` on Windows/Android — the same protocol, named differently by `convertFileSrc`) and, on a cache miss, straight from the Jellyfin server. `data:`/`blob:` cover inline and generated images. |
+| `media-src` | `<video>`/`<audio>` sources: HLS transcodes and progressive streams from the server, the token-guarded loopback media server on `http://127.0.0.1:<random port>` (DR-137), and `blob:` for the MSE object URL hls.js attaches. |
+| `connect-src` | `ipc:` / `http://ipc.localhost` is Tauri's `invoke` transport (custom scheme on Linux/macOS, `http` host on Windows/Android) — without it every command is blocked. `http:`/`https:` is hls.js fetching manifests and segments; ordinary API traffic goes through Rust and is not subject to CSP. |
+| `worker-src 'self' blob:` | hls.js runs its demuxer in a worker built from a blob (`enableWorker: true`). Without `blob:` it falls back to main-thread demuxing — playback survives but costs more CPU. |
+| `object-src`, `frame-src` = `'none'` | No plugins, no iframes; both are classic injection sinks. |
+| `base-uri 'self'`, `form-action 'self'`, `frame-ancestors 'none'` | Block `<base>` hijacking, form exfiltration and framing. `frame-ancestors` is only honoured when the policy is delivered as a header, which is platform-dependent; it is harmless where it is not. |
+
+**`img-src`/`media-src`/`connect-src` are deliberately permissive.** The Jellyfin
+origin is typed in by the user at run time and is routinely plain `http` on a
+LAN, so it cannot be enumerated at build time. `http: https:` is a wide grant for
+*data* — but it still bars `file:`, `filesystem:` and scripting schemes, and it
+does not touch `script-src`, which is where an injected origin would actually
+hurt. A run-time policy naming the server exactly was considered and rejected:
+Tauri derives the header from immutable config at the moment it serves the HTML,
+so it would mean rebuilding the config and reloading the webview whenever the
+user adds or switches a server, to constrain a destination the user chooses
+anyway.
+
+`devCsp` mirrors the policy with `'unsafe-inline' 'unsafe-eval'` on `script-src`
+and `ws:`/`wss:` on `connect-src`, because the Vite dev server injects styles and
+code and drives HMR over a websocket. It applies only to `tauri dev`.
+
+### Asset protocol scope
+
+`app.security.assetProtocol.scope` is `$APPDATA/thumbnails/**` — not the storage
+root. `imageCache.ts` is the only `convertFileSrc` caller left in the frontend:
+downloaded media moved to the loopback media server in DR-137, and downloaded
+audio is opened by MPV/ExoPlayer directly from its path. The old `$APPDATA/**`
+grant let the webview read the SQLite database and the encrypted-token fallback
+file alongside the thumbnails it actually needs.
+
+If a new feature hands the webview a local file, widen this scope to that
+subdirectory specifically; a path outside it resolves to nothing and the webview
+reports `NETWORK_NO_SOURCE` (which is exactly how DR-134's failure presented).
 
 ## Local Data Protection
 
@@ -60,6 +122,41 @@ pub struct EncryptedFileStorage; // AES-256-GCM fallback
 | Downloaded Media | Filesystem permissions only |
 | Cached Thumbnails | Filesystem permissions only |
 
+## Path Confinement and Input Binding
+
+Two classes of defect, both of the same *shape*: a value that arrived from
+outside decided something it should not, at a site whose neighbours a few lines
+away already did it correctly.
+
+### Filesystem path confinement
+
+| Surface | Rule | TRACES |
+|---------|------|--------|
+| Thumbnail cache | The filename is built from `item_id`, `image_type` and `tag`; all three are sanitised (non-alphanumerics → `_`), and the resolved path is checked with `starts_with(cache_dir)` **at the point of use** | DR-210 |
+| Downloads | `file_path` and `target_dir` are sanitised inside `download_item` itself, not only in `download_item_and_start` — the latter is what made the existing guard bypassable rather than absent | DR-211 |
+
+Two mechanics worth remembering, because both are easy to get subtly wrong:
+
+- `Path::join` **neither folds `..` nor keeps the base when handed an absolute
+  path**. Confinement therefore has to be checked *after* the join, not before.
+- Sanitising is **per path component**. Whole-string sanitising would rewrite
+  `downloads/x.mp3` to `downloads_x.mp3` and relocate every existing download.
+
+The database keeps both the raw key and the resolved path, so lookups still match
+and pre-existing rows still resolve.
+
+### Query and URL construction
+
+Caller-supplied values are **bound or encoded**, never interpolated (DR-212):
+
+- The offline `get_items` item-type filter uses parameter placeholders rather
+  than formatting `IN ('a','b')`.
+- `build_get_items_endpoint` encodes `ParentId` / `IncludeItemTypes` / `SortBy` /
+  `SortOrder`. Encoding is **per element** and list separators stay unencoded,
+  because Jellyfin splits these parameters on the comma.
+- `player_set_volume` clamps at the command boundary — it previously accepted
+  NaN and out-of-range floats even though every backend clamps internally.
+
 ## Security Considerations
 
 1. **No Secrets in SQLite**: The database contains only non-sensitive metadata
@@ -67,3 +164,4 @@ pub struct EncryptedFileStorage; // AES-256-GCM fallback
 3. **Logout Cleanup**: Token deletion from secure storage on logout
 4. **No Token Logging**: Tokens are never written to logs or debug output
 5. **IPC Security**: Tauri's IPC uses structured commands, not arbitrary code execution
+6. **Webview Containment**: A restrictive `script-src` keeps injected script off the IPC surface; the asset protocol is scoped to the thumbnail cache only (see above)

docs/architecture/README.md

@@ -95,15 +95,15 @@ Each major subsystem is documented in its own file in this directory:
 
 | Document | Contents |
 |----------|----------|
-| [01 - Rust Backend](01-rust-backend.md) | Media session state machine, player state machine, playback mode, media items, queue manager, favorites, player backend trait, player controller, playlist system, Tauri commands |
-| [02 - Svelte Frontend](02-svelte-frontend.md) | Store structure, music library navigation, playback reporting, repository architecture, playback mode system, database service abstraction, component hierarchy, MiniPlayer, sleep timer, auto-play, navigation guard, playlist management UI |
-| [03 - Data Flow](03-data-flow.md) | Repository query flow (cache-first), playback initiation, playback mode transfer, queue navigation, volume control |
+| [01 - Rust Backend](01-rust-backend.md) | Media session state machine, player state machine, playback mode, media items, queue manager, favorites (marking + browsing), player backend trait, player controller, playlist system, **domain vocabulary owned by Rust** (search scope, library exclusions, streaming quality ladder), **background workers** (catalog indexer, drains), Tauri commands |
+| [02 - Svelte Frontend](02-svelte-frontend.md) | Store structure, music library navigation, playback reporting, repository architecture, playback mode system, database service abstraction, component hierarchy, MiniPlayer, sleep timer, auto-play, navigation guard, playlist management UI, library mosaic, series/episode navigation, downloaded browse, safe-area insets, native-video store, logging |
+| [03 - Data Flow](03-data-flow.md) | Repository query flow (cache-first), locally-indexed search, playback initiation, playback mode transfer, queue navigation, volume control |
 | [04 - Type Sync & Threading](04-type-sync-and-threading.md) | Rust/TypeScript type synchronization, Tauri v2 IPC parameter naming convention, thread safety patterns |
-| [05 - Platform Backends](05-platform-backends.md) | Player events system, MpvBackend (Linux), ExoPlayerBackend (Android), MediaSession & remote volume, album art caching, backend initialization |
-| [06 - Downloads & Offline](06-downloads-and-offline.md) | Download manager, download worker, smart caching engine, download/offline commands, player integration, frontend store, UI components |
+| [05 - Platform Backends](05-platform-backends.md) | Player events system, HTML5 video adapter, MpvBackend (Linux), ExoPlayerBackend (Android) incl. audio settings parity, **native video compositing**, MediaSession & remote volume, album art caching, backend initialization |
+| [06 - Downloads & Offline](06-downloads-and-offline.md) | Download manager, download worker, smart caching engine, **one storage model (cache entries are downloads)**, offline catalog visibility, download/offline commands, player integration, frontend store, UI components |
 | [07 - Connectivity](07-connectivity.md) | HTTP client with retry logic, connectivity monitor, network resilience architecture |
 | [08 - Database Design](08-database-design.md) | Entity relationships, all table definitions (servers, users, libraries, items, user_data, downloads, media_streams, sync_queue, thumbnails, playlists), key queries, data flow diagrams, storage estimates |
-| [09 - Security](09-security.md) | Authentication token storage, secure storage module, network security, local data protection |
+| [09 - Security](09-security.md) | Authentication token storage, secure storage module, network security, webview CSP + asset-protocol scope, **path confinement and input binding**, local data protection |
 
 ---
 
@@ -112,17 +112,24 @@ Each major subsystem is documented in its own file in this directory:
 ```
 src-tauri/src/
 ├── lib.rs                    # Tauri app setup, state initialization
-├── commands/                 # Tauri command handlers (90+ commands)
+├── commands/                 # Tauri command handlers (~245 #[tauri::command] fns)
 │   ├── mod.rs               # Command exports
-│   ├── player.rs            # 16 player commands
-│   ├── repository.rs        # 27 repository commands
-│   ├── playlist.rs          # 7 playlist commands
-│   ├── playback_mode.rs     # 5 playback mode commands
-│   ├── connectivity.rs      # 7 connectivity commands
-│   ├── storage.rs           # Storage & database commands
-│   ├── download.rs          # 7 download commands
-│   ├── offline.rs           # 3 offline commands
-│   └── sync.rs              # Sync queue commands
+│   ├── player/              # Player commands: queue, remote, session, settings, timers
+│   ├── repository.rs        # Repository commands (items, search, favourites, disk usage)
+│   ├── catalog.rs           # Catalog sync + the background index pass
+│   ├── favorites.rs         # Offline favourite drain
+│   ├── library.rs           # Library listing + folder exclusions
+│   ├── playlist.rs          # Playlist commands
+│   ├── playback_mode.rs     # Local/remote transfer
+│   ├── playback_reporting.rs
+│   ├── connectivity.rs      # Connectivity commands
+│   ├── storage/             # Storage & database commands: people, series_prefs, thumbnails
+│   ├── download/            # Download commands: mod, pinning, smart_cache
+│   ├── offline.rs           # Offline commands
+│   ├── device.rs            # Device id / capabilities
+│   ├── sessions.rs          # Remote sessions
+│   ├── sync.rs              # Sync queue commands
+│   └── sync_drain.rs        # Background sync-queue drain
 ├── repository/              # Repository pattern implementation
 │   ├── mod.rs               # MediaRepository trait, handle management
 │   ├── types.rs             # RepoError, Library, MediaItem, etc.
@@ -207,4 +214,9 @@ src/lib/
 
 The frontend is genuinely UI-heavy; business decisions live in Rust, but the UI owns layout, navigation, and interaction state.
 
-**Total Commands:** 90+ Tauri commands across 14 command modules
+**Total Commands:** ~245 `#[tauri::command]` functions across 17 command modules
+(~58k lines of Rust, ~37k non-test lines of TypeScript/Svelte).
+
+> Counts and line totals in this file are periodic snapshots, not gates — the
+> authority is the tree. Regenerate with
+> `grep -rc '#\[tauri::command\]' src-tauri/src` and `wc -l`.

docs/build/build-desktop-packages.md

@@ -6,14 +6,14 @@ 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) →
+([../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
+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.
@@ -28,7 +28,7 @@ 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))
+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? |
@@ -40,7 +40,7 @@ because `makepkg` is Arch-specific — it is not part of the unified builder.
 ## Linux: deb + rpm
 
 Tauri's bundler produces these natively. The build runs on the existing Ubuntu
-builder image ([../Dockerfile](../Dockerfile), `desktop-linux-build` stage):
+builder image ([../Dockerfile](../../Dockerfile), `desktop-linux-build` stage):
 
 ```bash
 bun run docker:build:linux            # deb + rpm -> ./dist
@@ -58,8 +58,8 @@ transcoded video). The deb/rpm declare these.
 
 **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)):
+[../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

docs/build/build-release.md

@@ -116,17 +116,30 @@ Runs after both builds succeed (only on version tags):
 - **Use:** Run directly on any Linux distro
 - **Installation:**
   ```bash
-  chmod +x jellytau_*.AppImage
-  ./jellytau_*.AppImage
+  chmod +x JellyTau_*.AppImage
+  ./JellyTau_*.AppImage
   ```
 
 #### DEB Package
-- **File:** `jellytau_*.deb`
+- **File:** `JellyTau_*.deb`
 - **Size:** ~80-120 MB
 - **Use:** Install on Debian/Ubuntu/similar
 - **Installation:**
   ```bash
-  sudo dpkg -i jellytau_*.deb
+  sudo dpkg -i JellyTau_*.deb
+  jellytau
+  ```
+- **Note:** the Debian package is named `jelly-tau` (Tauri kebab-cases
+  `productName`), while the command stays `jellytau`. The package declares
+  `Replaces`/`Conflicts`/`Provides: jellytau`, so upgrading from a release built
+  before the rename replaces it rather than installing a second copy.
+
+#### RPM Package
+- **File:** `JellyTau-*.rpm`
+- **Use:** Install on Fedora/openSUSE/similar
+- **Installation:**
+  ```bash
+  sudo rpm -i JellyTau-*.rpm
   jellytau
   ```
 
@@ -294,8 +307,8 @@ bun run tauri build   # Local build test
 ```
 
 ### Documentation
-1. Update [CHANGELOG.md](../CHANGELOG.md) with changes
-2. Update [README.md](../README.md) with new features
+1. Update [CHANGELOG.md](../../CHANGELOG.md) with changes
+2. Update [README.md](../../README.md) with new features
 3. Document breaking changes
 4. Add migration guide if needed
 

docs/build/build-windows.md

@@ -12,10 +12,10 @@ job / SMTC lockscreen), but it runs and plays media.
   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)
+  `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)),
+  [../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.
 
@@ -33,7 +33,7 @@ Tauri CLI bundle the **NSIS installer from a Linux host**.
 > `--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
+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`.
 

docs/build/ci-operations.md

@@ -0,0 +1,212 @@
+# CI operations
+
+How the pipeline is kept working: the builder image, the secrets it needs, the
+gates that must stay required, and the things that only a human with access to
+the Gitea instance can do.
+
+CI is **Gitea Actions** (`.gitea/workflows/`) on `gitea.tourolle.paris`, not
+GitHub.
+
+## The workflows
+
+| Workflow | Trigger | What it protects |
+|---|---|---|
+| [build-and-test.yml](../../.gitea/workflows/build-and-test.yml) | push/PR to `master` | Frontend + Rust gates, Android compile check, supply chain |
+| [traceability-check.yml](../../.gitea/workflows/traceability-check.yml) | push/PR | Requirement coverage ratchet, dangling IDs |
+| [build-release.yml](../../.gitea/workflows/build-release.yml) | tag `v*` | Builds, signs, publishes, and writes the update manifest |
+| [publish-docs.yml](../../.gitea/workflows/publish-docs.yml) | push to `master` | Docs site on the `gitea-pages` branch |
+
+## 🔴 CI installs no system tools
+
+Every build, test and packaging **tool** lives in the Docker image the job runs
+in. Never add `apt-get`, `rustup`, `sdkmanager`, or a `curl | tar -xz` of a
+binary to a workflow step.
+
+Fetching the project's *own declared dependencies* is not a toolchain install and
+is fine: `bun install`, cargo pulling crates from the lockfile, `cargo deny`
+fetching the RustSec advisory database. The distinction is tool versus data.
+
+This rule has been broken twice, both times invisibly until something else
+failed. `publish-docs.yml` downloaded mdBook from GitHub releases into
+`/usr/local/bin` at job time — a hard dependency on GitHub's CDN being up
+whenever docs were published. Both mdBook and the supply-chain tools are in the
+image now.
+
+## The builder image
+
+`Dockerfile.builder` → `gitea.tourolle.paris/dtourolle/jellytau-builder`.
+It carries: the pinned Rust toolchain plus rustfmt/clippy and the Android,
+Windows-MSVC targets; bun and Node; the Android SDK/NDK and a local Gradle
+distribution; Linux desktop and packaging deps (WebKitGTK, libmpv, rpm, NSIS,
+cargo-xwin); and the tooling — `cargo-deny`, `cargo-cyclonedx`, `mdbook`.
+
+Arch packages build in a separate `Dockerfile.arch`, because `makepkg` is
+Arch-specific.
+
+### Tags are pinned, and why
+
+Workflows name an **immutable dated tag** (`:2026.08`), never `:latest`. While
+every job said `:latest`, rebuilding the image silently changed what every build
+compiled against — including a rebuild of an old release tag, which is the
+opposite of reproducible.
+
+`:latest` is still pushed alongside, for local `docker compose` runs and manual
+pulls.
+
+Date tags rather than per-commit SHA tags on purpose: the runner shares a 74 GB
+disk with two other projects, and SHA-tagged images accumulated there until it
+filled. Keep a couple of dated tags live and prune the rest.
+
+### Changing the image
+
+The order matters — CI breaks if the workflow lands before the image exists.
+
+A caveat learned the hard way: the *trailing* layer is only fast for `cargo
+install` tools. Adding an **apt** package invalidates the packaging layer, which
+sits above the `cargo-xwin`/`cargo-deny` installs, so those recompile too — a
+~20 minute rebuild rather than ~2.
+
+```bash
+# 1. Edit Dockerfile.builder. Put new tools in the TRAILING layer: it exists so
+#    a tool change is a ~2 min rebuild instead of ~15.
+# 2. Build and push, tagged with the new month:
+./scripts/build-builder-image.sh 2026.09
+# 3. Repoint every workflow at the new tag, in the same commit as whatever
+#    needed the new tool:
+sed -i 's|jellytau-builder:2026.08|jellytau-builder:2026.09|g' .gitea/workflows/*.yml
+# 4. Verify the tools are actually in it:
+docker run --rm gitea.tourolle.paris/dtourolle/jellytau-builder:2026.09 \
+  -c "cargo deny --version; mdbook --version"
+```
+
+🔴 The Rust version is pinned in **two** places that must agree:
+`RUST_VERSION` in `Dockerfile.builder` and `channel` in
+`src-tauri/rust-toolchain.toml`. If they drift, rustup downloads the pinned
+toolchain inside the job — a toolchain install in CI. Bump both, rebuild, push,
+then merge.
+
+## Tauri plugin versions are pinned in pairs
+
+Every Tauri plugin exists twice: a Rust crate in `src-tauri/Cargo.toml` and an
+npm package in `package.json`. **The Tauri CLI refuses to build when the two are
+on different minor versions** — not a warning, a hard stop before compilation.
+
+Both sides are therefore pinned *exactly* (`"2.8.0"`, not `"^2.8.0"`). A caret
+range is what let them drift apart in the first place: `bun add` took the latest
+npm package while cargo held an older crate, and nothing noticed until a release
+build refused to start.
+
+Nothing in `build-and-test.yml` runs `tauri build` — that happens only on a tag —
+so this class of breakage used to be invisible until release day. The
+`Check Tauri plugin versions match` step runs `tauri info`, which performs the
+same comparison without building.
+
+To upgrade a plugin, move **both** sides together and re-run that step. Expect
+the Rust side to be the constraint: a newer plugin crate may pull a large
+transitive upgrade (bumping `tauri-plugin-log` to 2.9.0 also moved `wry`,
+`wasm-bindgen`, `web-sys` and `webkit2gtk`), which touches the webview and
+therefore video playback. That is a change to make deliberately, with a full
+build and a playback check — not one to slip into a release.
+
+## AppImage needs more than the Rust toolchain
+
+`linuxdeploy` (which Tauri downloads at build time to assemble the AppImage)
+shells out to distro tools that a minimal server image does not have. It aborts
+the whole bundle on the first one missing:
+
+```
+failed to bundle project: xdg-open binary not found
+```
+
+The image therefore carries `xdg-utils`, `desktop-file-utils` and `zsync`. This
+is a class of failure that **cannot be caught by building locally**: a developer
+machine is a desktop and has all three, so the AppImage builds there and fails in
+CI. It cost one release build to find.
+
+Tauri's AppImage bundler also downloads `linuxdeploy`, `AppRun` and two plugin
+scripts from GitHub during the build. That is Tauri's behaviour, not ours, but it
+means an AppImage build depends on GitHub being reachable from the runner.
+
+## Secrets
+
+Managed with the `tea` CLI (`tea actions secrets list`) or the repo settings UI.
+
+| Secret | Used by | Notes |
+|---|---|---|
+| `ANDROID_KEYSTORE_BASE64` | release | Base64 of the release keystore |
+| `ANDROID_KEYSTORE_PASSWORD` | release | |
+| `ANDROID_KEY_ALIAS` | release | |
+| `ANDROID_KEY_PASSWORD` | release | |
+| `TAURI_SIGNING_PRIVATE_KEY` | release | minisign key for the desktop updater |
+| `TAURI_SIGNING_PRIVATE_KEY_PASSWORD` | release | |
+| `GITEA_TOKEN` | release, docs | PAT; falls back to the auto-provided token |
+
+The updater keypair's public half is committed in `src-tauri/tauri.conf.json` —
+that one is meant to be public; it is what clients verify against. The private
+half exists in the Gitea secret and in the maintainer's local `.env` (which is
+gitignored) and at `~/.tauri/jellytau.key`.
+
+**Losing the private key means losing the ability to ship updates to installed
+desktop clients**, because they will only accept payloads signed by the key
+matching the public key they were built with. Recovering means generating a new
+pair, shipping a build carrying the new public key, and telling everyone on an
+older build to reinstall by hand. Back it up.
+
+## Required status checks
+
+Gitea → repo Settings → Branches → protect `master`, requiring:
+
+- `Run Tests`
+- `Android Compile Check`
+- `Supply Chain`
+- the traceability job
+
+Without branch protection, every gate in this document is advisory: a push
+straight to `master` lands whether or not CI is red. That is the state the repo
+was in for its whole history before this was set up.
+
+## The runner
+
+One self-hosted runner, one ~74 GB disk shared with two other projects. It fills,
+and when it does the symptoms are misleading: cargo dying mid-link, docker
+refusing to pull, `actions/cache` quietly not saving — anything except an obvious
+out-of-space error. Check the disk first.
+
+There is deliberately no scheduled job watching this. On a single-slot runner a
+daily job occupies the slot and pulls the builder image to run `df`, and `df`
+inside a container does not reliably describe the host's disk anyway — it would
+cost real build capacity to report a number that might be wrong. Check it by hand
+on the runner:
+
+```bash
+df -h /
+docker system df -v
+```
+
+When it does fill:
+
+```bash
+docker image prune -a
+docker volume prune -a     # the -a matters: without it, NAMED volumes are kept,
+                           # which is exactly how this filled up unnoticed
+```
+
+Never cache `src-tauri/target` — it is ~16 GB, and caching it under several keys
+is what filled the disk at ~1.15 GB/day. The workflows cache only the cargo
+registry index and `.crate` tarballs; cargo re-extracts `registry/src` for free.
+
+## Release verification
+
+The steps that catch a broken release before users do are in
+[release-checklist.md](../release-checklist.md) — in particular the update path:
+`latest.json` must be live on the `updater` branch, both platform entries must
+carry a non-empty signature, and the previous release should be installed and
+asked to update to the new one.
+
+## Bus factor
+
+The Gitea instance holds the canonical remote, the signing secrets, the container
+registry and the CI runner. **It is not backed up as part of this repository, and
+nothing in this repository can restore it.** That is the largest single risk to
+the project — larger than any gate in this document — and the backup lives
+outside it.

docs/defect-windows.md

@@ -0,0 +1,152 @@
+# 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 |
+| Hero banner auto-rotation never restarted after a manual swipe (DR-038) | v0.0.1 | **v0.9.1** | ~8.5 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) |
+| Length-less handoff transcode left to the player's own load-error retry, which can only restart it (DR-203) | v0.0.16 | **v0.8.2** | feature (the handoff's progressive-mp3 choice) |
+
+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/release-checklist.md

@@ -109,8 +109,9 @@ git push origin v1.2.0
 ## After Release (Workflow Complete)
 
 - [ ] Download artifacts from release page:
-  - [ ] `jellytau_*.AppImage` (Linux)
-  - [ ] `jellytau_*.deb` (Linux)
+  - [ ] `JellyTau_*.AppImage` (Linux)
+  - [ ] `JellyTau_*.deb` (Linux)
+  - [ ] `JellyTau-*.rpm` (Linux)
   - [ ] `jellytau-release.apk` (Android)
   - [ ] `jellytau-release.aab` (Android)
 
@@ -125,6 +126,24 @@ git push origin v1.2.0
   - [ ] All artifacts are uploaded
   - [ ] Release type is correct (prerelease vs release)
 
+- [ ] Verify integrity metadata (DR-216):
+  - [ ] `SHA256SUMS` is present, and `sha256sum -c SHA256SUMS` passes in the
+        directory you downloaded into
+  - [ ] SBOM files are present (`*.cdx.json`, `frontend-dependencies.txt`)
+
+- [ ] Verify the update path (DR-217) — this is the step that catches a broken
+      updater *before* users hit it, because a bad manifest fails only on their
+      machine:
+  - [ ] `latest.json` is live and names this version:
+        `curl -s https://gitea.tourolle.paris/dtourolle/jellytau/raw/branch/updater/latest.json | jq .version`
+  - [ ] Both platform entries carry a non-empty `signature`
+  - [ ] The `.AppImage.tar.gz`, its `.sig`, and the NSIS `.sig` are among the
+        release assets — the manifest points at them
+  - [ ] Install the **previous** release, launch it, and use Settings → Updates:
+        it should offer this version, install it, and relaunch
+  - [ ] On Android, Settings → Updates offers the releases page rather than an
+        install button (the updater plugin is not compiled for that target)
+
 - [ ] Announce release:
   - [ ] Post to relevant channels/communities
   - [ ] Update website/docs
@@ -255,9 +274,8 @@ First build takes longer (cache warming). Subsequent releases are faster due to
 **Android:** 8.0+
 
 ### 🔗 Links
-- [Changelog](../../CHANGELOG.md)
-- [Issues](../../issues)
-- [Discussion](../../discussions)
+- [Changelog](https://gitea.tourolle.paris/dtourolle/jellytau/src/branch/master/CHANGELOG.md)
+- [Issues](https://gitea.tourolle.paris/dtourolle/jellytau/issues)
 
 ---
 Built with Tauri, SvelteKit, and Rust 🦀

docs/requirements.md

@@ -16,7 +16,7 @@ For a narrative overview of the system design, see
 | UR-003 | Play videos | High | Done |
 | UR-004 | Play audio uninterrupted | High | Done |
 | UR-005 | Control media playback (pause, play, skip, scrub) | High | Done |
-| UR-006 | Control media when device is on lock screen or via BLE headsets | Medium | Done |
+| UR-006 | Control media when device is on lock screen or via BLE headsets | Medium | Done (Android); **not implemented on Linux** — see IR-005 |
 | UR-007 | Navigate media in library | High | Done |
 | UR-008 | Search media across libraries | High | Done |
 | UR-009 | Connect to Jellyfin to access media | High | Done |
@@ -37,13 +37,13 @@ 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 | Done (Linux only) |
+| UR-027 | Audio equalizer for sound customization | Low | Done (Linux; Android pending device verification) |
 | 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-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-031 | Crossfade between audio tracks | Low | Not implemented (blocked — see DR-034) |
+| UR-032 | Gapless playback for seamless album listening | Medium | Done (Linux; Android pending device verification) |
+| UR-033 | Volume normalization to prevent volume jumps between tracks | Low | Done (Linux; Android pending device verification) |
 | UR-034 | Rich home screen with hero banners, carousels, and personalized sections | High | Done |
 | UR-035 | View cast/crew (actors, directors) on movie/show detail pages | High | Done |
 | UR-036 | Navigate to actor/person page showing their filmography | Medium | Done |
@@ -70,6 +70,27 @@ For a narrative overview of the system design, see
 | 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-076 | Music browsing shows only what the listener considers music. A Jellyfin server commonly keeps podcasts, audiobooks, sound effects or sample packs in their own folders inside a music library; those folders can be **excluded by choice**, once, and every music surface — library grids, artist and album listings, genre rows, search and the home screen — then agrees on what is in scope. The choice is by folder, not by a name the app happens to recognise, so a folder called anything at all can be excluded and an item is never dropped because its title matched a word | Medium | Done |
+| UR-077 | The app can update itself, or tell the user how. Somebody who installed an AppImage or ran the Windows installer had no upgrade path at all: nothing in the app ever mentioned that a newer version existed, and the release notes were the only announcement. On Linux and Windows the app checks a signed manifest, offers the new version with its notes, and installs and relaunches on request — the signature check is the point, since it is what stops a substituted download from being installed by the app itself. Android cannot do this (an app may not overwrite its own APK; that is the package installer's job) and is given the honest alternative, a link to the releases page, rather than a button that would throw | Medium | Done |
+| UR-078 | JellyTau keeps a record of what it did, and can hand it over. The app forgot everything the moment it exited: the backend logged to stdout only — which a user launching from a desktop icon never sees, and which on Android is not logcat, so the Rust half was invisible on the platform carrying the hardest bugs. A crash left nothing at all. Logs are now written to a size-capped rotating file, a panic is recorded before the process dies, the frontend's messages land in the same timeline as the backend's, and Settings exports the lot as one file to attach to a bug report. Nothing is transmitted anywhere — the user attaches it themselves, which is also what keeps this from being telemetry. Access tokens and passwords never reach the file | Medium | Done |
+| UR-079 | The app decides *what stream to play* and says so. Playing a video used to mean asking the server to re-encode it, always — a decision made nowhere, written down nowhere, and re-derived downstream by whoever needed it: the player worked out whether it had been handed a playlist by looking for `.m3u8` in the URL. So a viewer paid for a transcode of a file their device could have played untouched, and the app could not tell them which it was. Now one negotiation produces one self-describing answer — direct play, remux, or transcode; over a playlist, a plain HTTP file, or a local one — and every renderer consumes that same answer instead of guessing from a string. On Android, where the player decodes almost everything the library holds, this stops around 85% of plays from starting a transcode nobody needed | Medium | Done |
+| UR-080 | Video on the desktop plays as itself. The picture was drawn by a webview `<video>` element, which decodes little beyond h264 — so the app told the server it could accept only h264, and the server re-encoded almost everything before sending it. That was never a statement about the machine: the same machine already runs mpv for audio, which decodes essentially the whole library. Measured against a real library, 93% of desktop playback was a transcode nobody needed, against 15% on Android where a real decoder does the work. mpv now draws the picture, the app claims what it can genuinely decode, and video is sent as it was stored wherever that is possible — sparing the server the work, the network the bitrate, and the picture a generation of re-encoding | Medium | Proposed |
+| 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 |
 
 ---
 
@@ -85,7 +106,7 @@ External system integrations and platform-specific implementations.
 | IR-002 | Build scripts for Android and Linux | Build | UR-001 | Done |
 | IR-003 | Integration of libmpv for Linux playback | Playback | UR-003, UR-004 | Done |
 | IR-004 | Integration of ExoPlayer for Android playback | Playback | UR-003, UR-004 | In Progress (basic playback works, audio settings missing) |
-| IR-005 | MPRIS D-Bus integration for Linux lockscreen/media controls | Platform | UR-006 | Planned |
+| IR-005 | MPRIS D-Bus integration for Linux lockscreen/media controls | Platform | UR-006 | Planned — genuinely absent: no `mpris`/`souvlaki`/`zbus`/`dbus` code or dependency in the project (`zbus` appears in `Cargo.lock` only transitively, via `tauri-plugin-opener`), and no `navigator.mediaSession` use in the frontend. `player::update_lockscreen_metadata` is a no-op off Android. UR-006 is therefore Android-only |
 | IR-006 | Android MediaSession integration for lockscreen controls | Platform | UR-006 | Done |
 | IR-007 | Bluetooth AVRCP integration via system media session | Platform | UR-006 | Planned |
 | IR-008 | Android audio focus handling (pause on call) | Platform | UR-004, UR-006 | Done |
@@ -99,9 +120,9 @@ External system integrations and platform-specific implementations.
 | IR-015 | Jellyfin API client for playback progress reporting | API | UR-019, UR-025 | Done |
 | IR-016 | Jellyfin API client for subtitle/audio track info | API | UR-020, UR-021 | Done |
 | 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 | Done (Linux/MPV; Android parity pending) |
+| IR-018 | Subtitle rendering and selection in the **video** playback backends: ExoPlayer sideloads each track as a `MediaItem.SubtitleConfiguration` and selects by text-track-group position (Android), and the WebKitGTK HTML5 `<video>` element renders `<track kind="subtitles">` children carrying `data-stream-index` (Linux). **Originally scoped to libmpv, which never implemented it**: `MpvBackend` is the audio-only backend here and does not override `PlayerBackend::set_subtitle_track`, so the default `not_implemented()` still stands there. UR-020 is satisfied by the two paths above rather than by MPV | Playback | UR-020 | Done |
+| IR-019 | Audio track selection in the **video** playback backends: ExoPlayer switches track by index natively (Android), while the HTML5 `<video>` path cannot switch a track in the element and instead re-opens the stream at the chosen `AudioStreamIndex` and resumes at the same position (Linux) — the two outcomes `AudioTrackSwitchResponse` distinguishes. **Originally scoped to libmpv, which never implemented it**: `MpvBackend` does not override `PlayerBackend::set_audio_track`, so the default `not_implemented()` still stands there. UR-021 is satisfied by the two paths above rather than by MPV | Playback | UR-021 | Done |
+| IR-020 | libmpv/ExoPlayer equalizer integration | Playback | UR-027 | Done (Linux/MPV and Android/`audiofx.Equalizer`; Android pending device verification) |
 | 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 |
@@ -110,6 +131,30 @@ 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 |
+| IR-033 | libmpv render-API integration for video: `vo=libmpv` driving an OpenGL FBO bound by the host toolkit, with GL entry points resolved through libepoxy. Note that libepoxy exports them as *data* symbols — there is no `glFoo` function, only an `epoxy_glFoo` variable holding a lazily-resolving pointer — so `get_proc_address` must return the pointer stored **at** that symbol; returning the symbol's own address makes mpv jump into non-executable data and take SIGSEGV on the first GL call. The `epoxy` crate resolves this correctly but is unusable, its `gl_generator` dependency pulling a yanked `xml-rs` | Playback | UR-080 | Proposed |
+
+> **Where a UR is met by a different mechanism than its IR anticipated.** Several
+> integration requirements were written when libmpv was expected to be the single
+> playback backend. It is not: `MpvBackend` is the **audio-only** backend, Linux
+> plays video through a WebKitGTK HTML5 `<video>` element (HLS/h264), and Android
+> plays through ExoPlayer. So:
+>
+> * **UR-020 / UR-021** (subtitle and audio track selection) are Done, but not by
+>   MPV — `MpvBackend` overrides neither `PlayerBackend::set_subtitle_track` nor
+>   `set_audio_track`, leaving the trait's `not_implemented()` default. IR-018 and
+>   IR-019 have been **re-scoped to the backends that actually deliver them**
+>   (ExoPlayer sideloaded `SubtitleConfiguration`s and native track switching;
+>   HTML5 `<track>` children and stream re-open at the chosen `AudioStreamIndex`)
+>   and marked Done on that basis. IT-008 / IT-009 were re-worded to match.
+> * **UR-006** (lockscreen / BLE headset control) is Done **on Android only**, via
+>   `MediaSessionCompat` (IR-006) and ExoPlayer/`AudioManager` focus (IR-008).
+>   IR-005 (MPRIS) remains Planned because it genuinely does not exist — there is
+>   no MPRIS/D-Bus code or dependency in the project, and
+>   `player::update_lockscreen_metadata` is a no-op off Android. UR-006's status
+>   was corrected rather than IR-005's.
 
 ### 2.2 Jellyfin API Requirements
 
@@ -149,6 +194,10 @@ 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 |
+| JA-036 | Query next-up episodes excluding in-progress ones (`/Shows/NextUp` with `EnableResumable=false`) | Shows | UR-059 | Done |
 
 ### 2.3 Development Requirements
 
@@ -191,9 +240,9 @@ Internal architecture, components, and application logic.
 | 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-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-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 via MPV; Android via `pauseAtEndOfMediaItems` — pending device verification) |
+| DR-036 | Volume normalization with preset levels (Loud/Normal/Quiet) | Player | UR-033 | Done (Linux via MPV `dynaudnorm`; Android via `LoudnessEnhancer` — pending device verification) |
 | DR-037 | Remote session browser and control UI | UI | UR-010 | Done |
 | DR-038 | Home screen with hero banner carousel (featured/continue watching) | UI | UR-034 | Done |
 | DR-039 | Home screen horizontal carousels (recently added, recommendations) | UI | UR-034, UR-024 | Done |
@@ -218,10 +267,10 @@ 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 |
@@ -242,6 +291,148 @@ Internal architecture, components, and application logic.
 | 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 was `$APPDATA/**` — the storage root under which the database, `downloads/` and the thumbnail cache all live — rather than an unrestricted grant; DR-198 narrows it further to `$APPDATA/thumbnails/**`, since DR-137 moved downloaded media off this protocol and thumbnails are all it still serves | 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 on** since DR-194/DR-196) |
+| 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: the background-audio handoff could only *return* through the HTML5 element, so coming back from the lockscreen left playback dead, and the flip waited for that rather than shipping a verified sub-path over an unverified one as DR-161 had. **The default is now on.** The two defects holding it back are fixed and device-verified — DR-196 (the handoff return restarts the renderer that is actually on screen) and DR-194 (the letterbox bars are painted rather than retaining stale framebuffer content) — with the evidence this default has been held to since DR-161: an audio handoff at 69:54 returning to video playing at 70:18, and clean bars across playback, the control bar and a rotation round-trip. An explicit stored choice still wins in both directions, so an opt-out survives the flip (the stored value is null-checked rather than compared to "true", which would have silently re-enabled it for everyone who turned it off) | Android | UR-003, UR-004 | Done |
+| DR-189 | The control bar comes down on a touchscreen. Its hide timer was armed from exactly one place — the player container's `onmousemove` — and a touchscreen never fires `mousemove`, so on Android the bar was never scheduled to hide and sat over the video for the whole film. It went unnoticed for as long as the native video surface was itself invisible (DR-172/DR-185): with nothing behind it to obscure, a permanent control bar reads as the UI rather than as a defect. Two changes, because there were two faults. `revealControls()` replaces `handleMouseMove` and is called on entry and on every touch interaction as well as on mouse movement, so touch arms the countdown. And the countdown became an `$effect` over the state rather than a one-shot timer armed by the input event: the first attempt armed a timer on entry, three seconds later playback had not started, `shouldHideControls` correctly declined, and nothing ever re-armed it — the timer has to follow the conditions that *permit* hiding, which arrive on their own schedule. The decision itself is `shouldHideControls` in `controlsVisibility.ts`, pure and separated from the clock and the DOM, because what was wrong here was the conditions and not the `setTimeout`: the bar stays up while paused (a user who paused by tapping the surface has no other way back), mid-seek (the position readout is the point of the bar then), and while any track/subtitle/quality menu is open (the menus are anchored to the bar, so hiding it would take the open menu with it) | UI | UR-003, UR-066 | Done |
+| 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-196 | Returning from background audio brings the picture back on the **native** path, because the return now restarts the renderer that is actually on screen. The two paths resume by different means: the webview `<video>` reloads off its stream URL, watched by an `$effect` that reinitialises HLS and lets `canplay` drive the seek — while ExoPlayer owns no element and nothing watches the URL on its behalf, so its playback is only ever started by an explicit `player_play_item` + adapter load, issued once from `onMount`. `exitBackgroundAudioHandoff` did only the URL assignment, for both paths, so on the native path it restarted nothing: `player_exit_background_audio` had already stopped the handoff's audio player, leaving the backend holding no item at all. The symptom is a black screen with a play overlay pinned at 0:00, a seek bar at zero, and a play button that does nothing — the process alive and the frontend still logging, since nothing crashed; the transition was simply dropped. The branch is decided by `planHandoffReturn` (pure, in `backgroundAudioHandoff.ts`), which also folds in `shouldResumeOnForeground` so a lockscreen pause during the handoff still wins over the snapshot taken on the way out. Subtitle configurations are reused from the ones resolved at mount, since ExoPlayer sideloads them as `MediaItem.SubtitleConfiguration`s and cannot accept one after `prepare()`. Verified on device: handoff to audio at 69:54, return restored video playing at 70:18 | Playback | UR-040, UR-003 | Done |
+| DR-197 | Continue Watching and Next Up stop showing the same episode. Jellyfin's `/Shows/NextUp` defaults `EnableResumable=true`, which returns a partially-watched episode as its own series' next up — precisely the episode `/Items/Resume` already returns — so the Home "Next Episode" row and the TV landing's Next Up row duplicated Continue Watching card for card. `build_next_up_endpoint` sends `EnableResumable=false`, and because servers predating that parameter ignore it, `filterInProgressNextUpItems` also drops any next-up entry whose id appears in the resume list. It is the mirror of DR-089 and lives beside it: same presentation-layer de-duplication over two lists the frontend already holds, no Jellyfin taxonomy involved. The resume filter still reads its frontier from the *unfiltered* Next Up list, so removing in-progress entries cannot resurrect a stale resume card. The division is then exact: Continue Watching offers episodes the viewer has started and not finished, Next Up offers the episode after the ones they finished | Repository | UR-059 | Done |
+| DR-200 | The lockscreen notification is exempt from `POST_NOTIFICATIONS`, because of the **session token**, not because it belongs to a foreground service — and the difference is what the code now records. `POST_NOTIFICATIONS` was declared in the manifest and requested nowhere, so on Android 13+ it sat permanently denied; an audit read that as a threat to UR-006, since the media notification is what carries the lockscreen transport controls. It is not. Android's own wording is that the permission covers "non-exempt (including Foreground Services (FGS)) notifications", with denied users seeing FGS notices "in the Task Manager but [not] in the notification drawer" — so an FGS notification is explicitly *not* exempt — while separately "Notifications related to media sessions are exempt from this behavior change". The platform predicate is `Notification.isMediaNotification()`, which requires `MediaStyle` **and** a non-null `EXTRA_MEDIA_SESSION`, and it is byte-identical across API 33–36. `NotificationManagerService` uses it to decide whether to drop the post, and SystemUI's media carousel (`MediaDataProcessor.onNotificationAdded`) is gated on the *same* predicate — so a token-less notification is not merely absent from the shade, it never reaches the notification listener and the lockscreen/Quick-Settings controls do not exist at all. Confirmed on device (HONOR ROD2-W09, Android 16 / SDK 36): appops `POST_NOTIFICATION: ignore`, `granted=false`, and the service simultaneously `isForeground=true` with `foregroundNoti=Notification(category=transport actions=3 vis=PUBLIC)`. So **no runtime permission request is added** — a prompt the app does not need is a prompt that can be permanently denied for nothing — and no `checkSelfPermission` gate is placed on `startForeground`, which would trade a cosmetic problem for the "did not then call Service.startForeground()" kill. What is added is the guard that matches the real precondition: `mediaSessionCompat?.sessionToken` is a null-safe call, and the exemption hangs entirely on it, so both builders now bind the token once and log an error if it is ever null while the permission is denied — converting a failure that is invisible unless the tester happened to deny the permission (most grant it reflexively) into a logcat line. The manifest declaration is *kept*, unrequested, and documented: media3 does not need it (media3-session declares no permissions and the `MediaSessionService` guide asks only for the two `FOREGROUND_SERVICE` ones), but the exemption covers media and self-managed-call notifications only, so a download-completion notice (UR-011) would be an ordinary notification and silently dropped — keeping the declaration is what makes adding one a one-file change | Android | UR-006 | Done |
+| DR-201 | A lockscreen skip means different things depending on what is playing, and the backend decides which. `onSkipToNext`/`onSkipToPrevious` forwarded a bare `"next"`/`"previous"` to Rust, which always advanced the queue — correct for music, wrong for a video whose audio is running through a background-audio handoff (UR-040), where the buttons should scrub. Pressing skip to re-hear a line jumped to the next *episode* instead. `resolve_skip_action` in `player/seek.rs` maps the command to either `Advance` or `SeekTo`, and `is_background_audio_active()` is the whole test: the handoff exists only for video, and an episode played through it reports `MediaType::Audio`, so media type cannot distinguish the case. Forward jumps 30s, back 10s — asymmetric because the back button replays dialogue just missed rather than travels — and both clamp to `[0, duration]`, since a negative offset is rejected by backends and a seek past the end reads as EOF and would advance, the very outcome being prevented. Routed through the same spawn-then-`seek_absolute` path as the scrubber, because a handoff seek re-opens the stream and must not run under the blocking lock (DR-159). The Kotlin keeps sending the same opaque command; only the `PlaybackStateCompat` gains `ACTION_FAST_FORWARD`/`ACTION_REWIND` so the system draws seek affordances rather than skip arrows that lie about what they do | Playback | UR-040, UR-006 | Done |
+| DR-202 | Video keeps the display awake. Android counts its display timeout from the last *user input*, and watching something is exactly the case where there is none, so the screen dimmed and slept mid-film unless the user kept tapping it. Nothing held it: `FLAG_KEEP_SCREEN_ON` appeared nowhere in the app, and neither renderer supplies a hold for free — ExoPlayer's `setWakeMode` is a CPU/wifi wake lock that says nothing about the display, and it draws into the `TextureView` this app owns (DR-192) rather than media3's `PlayerView`, which is the widget that would otherwise set `keepScreenOn` itself; the WebView `<video>` path is no better, because the display wake lock Chrome takes for video lives in the browser layer and not in an embedded WebView. `ScreenWakeManager` toggles `FLAG_KEEP_SCREEN_ON` on the Activity window — window-scoped, so it stops applying the moment the app is not visible and cannot outlive a crash the way an explicitly acquired `PowerManager.WakeLock` can, and it needs no permission (the manifest's `WAKE_LOCK` is the media service's). The two rendering paths are independent holders OR-ed in the pure `ScreenWakeState`: the native path follows `onIsPlayingChanged` plus surface teardown, so the hold tracks what ExoPlayer *reports* rather than what the UI intends, and the webview path reuses the `setHtml5VideoState` report the frontend already sends for PiP (DR-160) rather than adding a bridge. Audio is deliberately not a holder — playing music with the screen off is the point of that path — so the hold is gated on the media type being video, and it is dropped on pause, on stop, on surface teardown, and on a new WebView, since a page that goes away never sends its own final `active = false`. Also the repo's first Kotlin JVM unit tests: `ScreenWakeState` is framework-free so the decision is testable off-device with `./gradlew :app:testUniversalDebugUnitTest`. Verified on device (FP5, native path): `IS PLAYING CHANGED: true` → `keepScreenOn = true` 17 ms later and `fl=KEEP_SCREEN_ON` on the window in `dumpsys`, a pause releasing it and the resume re-taking it. The webview path is unverified | Android | UR-003, UR-004 | Done |
+| DR-203 | The background-audio handoff stops silently rewinding to the point it started. A player retry is only a *retry* if it can resume where the load failed, and ExoPlayer decides that in `ProgressiveMediaPeriod.configureRetry`: it keeps the load position when the content length is known or the extractor produced a seek map with a duration, and otherwise assumes the source is live — the data at the URL is taken to have changed, so every sample queue is reset and the URL is re-requested from offset 0. The handoff transcode (`/Audio/{id}/universal?Container=mp3&TranscodingProtocol=http`, DR-129) satisfies neither condition: chunked, so no `Content-Length`, and a live mp3 encode carries no `Xing` header, so the duration is unset — on device every position tick reads `<position> / 0.0`. Its URL carries `StartTimeTicks` = the handoff point, so "from offset 0" is the handoff point, and after any transient load error playback resumed there and ran on normally. Nothing was reported: a successful retry raises no error and no `STATE_ENDED`, so neither arm of DR-129 was ever consulted, no `onPositionDiscontinuity` handler existed, and the app's only trace of it was a position that went backwards — which is why it read as random, since it needs a network blip to land while a load is in flight rather than while the ~50s buffer covers it, and why it survived the two earlier fixes for the same *symptom* (DR-129's phantom end, DR-159's relative-timeline leak). The decision is Rust's: `player_retry_restarts_stream` marks a `Remote` audio-only video item, and `loadWithMetadata` carries the answer to Kotlin, where the pure `StreamRetryDecision` holds it for a `DefaultLoadErrorHandlingPolicy` subclass that returns `C.TIME_UNSET` — which makes `onLoadError` answer `DONT_RETRY_FATAL` *before* reaching `configureRetry`. The rewind therefore becomes a recoverable error, and `recoverable_error_resume` already knows what to do with one: re-open at the position playback actually reached, `StartTimeTicks` rewritten, with backoff and the shared attempt budget. Every other source keeps the player's retry, because a static file and an HLS playlist both declare their timeline and are resumed in place. A `onPositionDiscontinuity` handler is added for the log line alone, so a recurrence is visible rather than invisible — loud for `DISCONTINUITY_REASON_INTERNAL`, which is the rewind's own signature, and quiet for the backwards jump a resume's re-prepare legitimately makes. Reproduced and verified on device (FP5), same procedure both times: background-audio handoff, 60s to fill the buffer, a 45s radio outage, then watch. **Before** — the outage passed unnoticed and 3.5 minutes later, with nothing logged in between, `BUFFERING` → `READY` → position `1165.4s` → `840.3s`, exactly the handoff base, no error and no `STATE_ENDED`; the same log line reports `Media ready! Duration: -9.223372036854776E15`, which is `C.TIME_UNSET` and the precondition itself. **After** — `Load error on a stream that cannot be resumed in place — declining the player's retry` at the outage, playback continuing undisturbed off the buffer for 69s (a fatal load error is only raised when the renderer next needs data), then `ERROR_CODE_IO_NETWORK_CONNECTION_FAILED` → `re-opening at 785.6s in 2s` → `READY`, playing on from 785.6s with no rewind in the following 7 minutes | Playback | UR-040, UR-004 | Done |
+| DR-199 | The webview stops undoing the network security config. `MainActivity.configureWebViewSettings` set `mixedContentMode = MIXED_CONTENT_ALWAYS_ALLOW` together with `allowFileAccess = true` and `allowContentAccess = true`, which is a blanket cleartext opt-in reached by hand — exactly the thing `network_security_config.xml` exists to prevent and its own comment warns against (DR-138). Nothing needed any of the three. `file://` is never loaded: cached thumbnails go through `convertFileSrc`, which on Android resolves to `http://asset.localhost/…` and is answered by wry's request interceptor rather than the filesystem, and downloaded media goes over the loopback HTTP server (DR-137), which exists precisely because the asset/file route cannot stream a large file. `content://` is never loaded either — the manifest's `FileProvider` is for outbound share intents, not webview navigation. And mixed content never arises: Tauri serves the UI from `http://tauri.localhost` (`use_https_scheme` defaults false and is not set in `tauri.conf.json`), while both `127.0.0.1` and `asset.localhost` are loopback/`.localhost` origins that Chromium treats as potentially trustworthy, so they are not mixed content to begin with. A plain-HTTP *remote* Jellyfin server would be, but the network security config already rejects it before any mixed-content check runs — so `ALWAYS_ALLOW` bought nothing and only widened the hole. `COMPATIBILITY_MODE` rather than `NEVER_ALLOW` is a deliberate hedge and not the default — the platform default at targetSdk 21+ *is* `NEVER_ALLOW` — because none of this can be verified anywhere but a device, and compatibility mode keeps passive content (images) working if the analysis missed a path. `allowFileAccess = false` restores the targetSdk-30+ default; `allowContentAccess = false` is a genuine tightening (its default is true) and is the first thing to look at if something that used to render stops. The two files now cross-reference each other so the pair cannot drift apart again | Security | UR-071 | Done (pending device verification) |
+| DR-194 | Stale pixels in the letterbox bars — the rotation "flash of the previous frame", a ghost control bar stranded in the top bar, each new clock digit drawn over the last (`35:42` with the `1` still showing through the `2`), and menus (sleep timer, quality) leaving their imprint behind. One cause for all of it: **nothing painted the bars.** The window surface is opaque (the theme is not translucent), and for an opaque surface HWUI deliberately does not clear the damaged region before replaying a frame — it assumes the view hierarchy covers every pixel. That hierarchy is window background → video `TextureView` → transparent WebView, and `fitSurfaceToScreen` sizes the TextureView to the *letterboxed* video rect, so the bars were the window background's alone to paint. `setTransparent(true)` cleared that background to `TRANSPARENT`, leaving the bars painted by nobody and whatever was last in the framebuffer surviving in them. Fixed by keeping the window background opaque black while compositing; the WebView's own background is what lets the video through, and the TextureView is drawn on top of the window background, so an opaque one cannot hide it. Three earlier fixes aimed at the window's rotation animation and at TextureView frame-retention (two `postOnAnimation` hops, an `onSurfaceTextureUpdated` reveal, then `ROTATION_ANIMATION_JUMPCUT` + `FLAG_FULLSCREEN`) all missed, because the pixels were never the animation's; the alpha-hiding among them made it worse by blanking the one view that reliably paints its own rect. Those are removed, `FLAG_FULLSCREEN` included — it fought edge-to-edge insets for no gain. Verified on device: ghosting reproduced with native video on, then absent after the fix, across playback, the control bar and a rotation round-trip | Android | UR-003, UR-066 | Done |
+| 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 | Superseded by DR-196 |
+| 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 |
+| DR-204 | A leveled logging facade for the frontend, replacing raw `console.*` calls. One module owns the log sinks, so a level (error/warn/info/debug) decides at run time what is emitted rather than every call site deciding permanently at authoring time: a release build stays quiet, a developer chasing a playback bug turns the player's debug output on without editing and rebuilding, and nothing that reaches the console is written by a `console.log` nobody can find again. Scoped loggers carry the subsystem in the message, so a filtered console is usable while a player, a download worker and a store are all talking | Tooling | - | Done |
+| DR-205 | ESLint + Prettier run as a gate over the frontend, so lint and formatting are decided once by configuration rather than per reviewer. Formatting is not a matter of opinion at review time, and the classes of bug a linter sees (unused bindings, floating promises, accidental globals) should never reach a human reviewer at all. Wired as an npm script so the same command runs locally and in CI, matching how `check:boundary` and the traceability gate already work | Tooling | - | Done |
+| DR-206 | The Rust toolchain is pinned in-repo (`rust-toolchain.toml`) and the pin is what both a developer's machine and CI use. Without it, `cargo fmt --check` and `cargo clippy` are run by whatever version each host happens to have, so a formatting or lint result differs between a laptop and the builder image and CI fails on a diff that was clean locally — the failure mode is a red build nobody can reproduce. The builder image carries the pinned toolchain, so pinning is a *declaration*, not a CI-time install (see the no-toolchain-installs rule) | Tooling | - | Done |
+| DR-207 | A pre-commit hook runs the "Before Committing" gates — frontend checks and tests, `cargo fmt`, clippy, the boundary tripwire and the traceability checks — so the gates are enforced at the commit rather than discovered in CI. The gates already exist and are already documented; what is missing is that nothing runs them, which makes compliance a matter of memory. The hook is the mechanism that makes the documented list actually binding | Tooling | - | Done |
+| DR-208 | Documentation link integrity is checked mechanically (`scripts/check-doc-links.sh`): every relative markdown link in every tracked `.md` must resolve to a file that exists on disk. This is a real defect class, not hygiene — the generated traceability matrix shipped ~2,800 dead file links because it was written to `docs/` while its hrefs were repo-root-relative, and nothing noticed for months because no check existed and nobody clicks 2,800 links. The check validates *paths*, deliberately not anchors or external URLs: anchor resolution needs a markdown renderer's slug rules and network checks make the gate flaky, so both are out of scope and stated as such in the script | Tooling | - | Done |
+| DR-209 | Library folders are excluded from music browsing **server-side, by folder id**, replacing a hardcoded frontend filter that dropped anything whose name contained "Podcasts". The name filter was wrong in three separate ways: it encoded a domain classification in the presentation layer, it matched on a title rather than on what an item *is* (so an album legitimately called "Podcasts" vanished while a podcast folder named anything else did not), and it applied only where someone had remembered to call it, so the same library was in scope on one screen and out of scope on the next. Excluded folder ids are stored as user configuration and applied by the repository layer to every music query — libraries, artists, albums, genres, search and the home rows — so scope is decided in one place and is the same everywhere | Repository | UR-076 | Done |
+| DR-210 | Thumbnail cache writes are confined to the cache directory. The filename was built from `item_id`, `image_type` and `tag`, but only `tag` was sanitised — and `Path::join` neither folds `..` nor keeps the base when handed an absolute path, so a value arriving verbatim from server JSON decided where a file landed. The tag's existing rule (non-alphanumerics become `_`) now applies to all three parts, and the resolved path is checked with `starts_with(cache_dir)` at the point of use. The database keeps the raw key and the resolved path, so lookups still match and pre-existing rows still resolve. Not exploitable as shipped — server URLs must be HTTPS and Android blocks cleartext, so the id comes from a server the user chose to trust — the value is making the write path consistent with how caller-supplied paths are handled elsewhere | Storage | UR-012 | Done |
+| DR-211 | Download paths are confined to the download root. `file_path` and `target_dir` reached `PathBuf::join` unchecked from the frontend, and `mark_download_completed` persisted a caller-supplied path later passed to `remove_file`. A correct sanitiser already existed and `download_item_and_start` used it, but `download_item` is itself a command accepting `file_path` raw, so the guard was bypassable rather than absent — the fix moves it inside instead of adding a second one. Sanitising is **per path component**: whole-string sanitising would rewrite `downloads/x.mp3` to `downloads_x.mp3` and relocate every existing download. Confinement happens after the join, since a join with an absolute second half discards the root | Downloads | UR-011 | Done |
+| DR-212 | Query and URL construction bind or encode their inputs. Three sites interpolated caller-supplied values directly: the offline `get_items` item-type filter built `IN ('a','b')` by string formatting, `build_get_items_endpoint` wrote `ParentId`/`IncludeItemTypes`/`SortBy`/`SortOrder` into a URL unencoded, and `player_set_volume` accepted NaN and out-of-range floats. Each is a *consistency* defect rather than a novel one — the same file already did it correctly a few lines away (parameter placeholders in `search`, `urlencoding::encode` for genres, `clamp` in every player backend). List separators stay unencoded and encoding is per element, because Jellyfin splits these parameters on the comma | Repository | UR-007, UR-065 | Done |
+| DR-213 | Containerised builds hand their artifacts back to the host user. The compose services bind-mount the repo and run as root — their caches live at `/root/.cargo` and `/root/.bun`, so a non-root container user cannot write them — which leaves root-owned files accumulating in the developer's working tree: 11,124 of them when this was found, enough that `cargo clean` and `scripts/clean.sh` failed with EACCES and a plain `cargo build` died part-way, since build scripts compile for the host and land in `target/debug` even during a cross-build. Ownership is restored at the end of each containerised build, reading the intended owner from the checkout so no uid needs plumbing through. Running the containers as the host uid is the tidier fix and remains open; it needs the cache volumes relocated off `/root` first | Tooling | - | Done |
+| DR-214 | The app identifies itself correctly everywhere a user or a package manager reads its name. `productName` was the scaffold's lowercase `jellytau`, which is what the Android release build showed under its icon and what the deb/rpm/NSIS bundles carried as their display name — invisible in development because `build.gradle.kts` overrides the label to "JellyTau Debug" for the debug build type, so the install a developer looks at daily was the only correctly-cased one. `mainBinaryName` pins the executable filename so nothing that resolves a path by name has to change. `strings.xml` moves into the canonical android tree, where `sync-android-sources.sh` already copies `res/values/*.xml`, so the fix survives regenerating `gen/`. Bundle metadata (publisher, copyright, category, descriptions, licence) was entirely absent, which is why the packages shipped with no maintainer or description — the hand-written Arch PKGBUILD and `.desktop` had all of it, so only the *generated* packaging was wrong | Packaging | - | Done |
+| DR-215 | Frontend test coverage is a ratcheted CI gate rather than a number nobody looks at. `test:coverage` had been configured since the suite was created and was silently broken: `@vitest/coverage-v8` resolved to 4.1.10, whose peer range pins `vitest` exactly, while `package.json` asked for `>=1.0.0 <5.0.0` and got 4.0.16 — so every invocation died on a missing `BaseCoverageProvider` export and no coverage figure had been produced in months. Fixing the range is half the requirement; the other half is that a measured figure that gates nothing decays the same way an unrun script does. Thresholds sit a few points under the measured result (statements 54.6, branches 48.7, functions 49.6, lines 55.1 when this landed) and only ever move up, matching `MIN_THRESHOLD` in the traceability gate and the eslint `--max-warnings` ratchet. The absolute numbers are held down by `.svelte` components, which this project deliberately does not test directly — the pattern is to extract the logic to a plain module and test that | Tooling | - | Done |
+| DR-216 | Dependencies are gated on known vulnerabilities and on licence compatibility, and the build graph is pinned to what is actually shipped. The project had no scanning of any kind: nothing checked the ~500-crate Rust graph or the JS packages against an advisory feed, and nothing checked that everything redistributed inside an MIT-licensed bundle permits it. The first run found eight vulnerabilities and one unsoundness — `bytes`, four in `rustls-webpki`, `time`, two in `quick-xml`, `rand` — every one closed by a `cargo update` nobody had reason to run. `cargo deny` (src-tauri/deny.toml) now runs in CI over advisories, licences, bans and sources. Two structural fixes matter as much as the gate: the graph is scoped to the targets actually shipped, so an advisory against an Apple-only path is correctly absent rather than ignored by ID; and the one git dependency (`libmpv`) is pinned by revision instead of by branch, since a branch means any `cargo update` silently substitutes new upstream code in the one dependency that is unsigned and links a C library into the player. Licence findings are recorded rather than waved through — `libmpv`/`libmpv-sys` are LGPL-2.1, which the app satisfies by dynamic linking, and that carries obligations (keep the linkage dynamic; ship libmpv's licence text with any bundle carrying the .so) | Tooling | - | Done |
+| DR-217 | In-app update, desktop only, over a manifest we control. `tauri-plugin-updater` and `tauri-plugin-process` are compiled for everything except Android/iOS — spelled as a target-triple cfg rather than `cfg(desktop)`, which Cargo does not evaluate in a `[target.'cfg(…)']` table and which therefore drops the dependency silently, surfacing much later as "Permission updater:default not found". The release workflow signs updater artifacts with a minisign key held in Gitea secrets and publishes `latest.json` to a dedicated `updater` branch, read over Gitea's raw-file URL: this instance serves `/releases/download/<tag>/<asset>` but returns 404 for `/releases/latest/download/<asset>`, so there is no stable latest-release URL to point at, and the docs branch is force-pushed by publish-docs.yml so it cannot host the manifest either. Bundle targets gain `appimage`, which the release notes had been advertising for months while `tauri.conf.json` never built it — the artifact step globbed for `*.AppImage`, found nothing, and said nothing | Tooling | UR-077 | Done |
+| DR-218 | Persistent, redacted logging and a diagnostics export. `tauri-plugin-log` replaces the `env_logger` stdout-only init, giving a rotating 5 MB file, a webview target in dev, and — the single largest gain — logcat on Android, where `env_logger`'s stdout went nowhere. **Redaction runs in the log formatter, not at export**: a credential in a file on the device is already a disclosure, so stripping it on the way out would be too late; the exporter redacts a second time to cover files written by older builds. `api_key`/`X-Emby-Token`/`Authorization`/`"AccessToken"`/`Token="…"` all reduce to `[REDACTED]` while host, item ids and filenames are deliberately kept — a bundle scrubbed of those is one nobody can debug from. The server URL is reduced to scheme and host, dropping any embedded `user:pass@`. The panic hook chains to the previous hook rather than replacing it, because `utils/lock.rs` installs a silencing hook around tests that provoke poisoned locks on purpose. The chosen level persists to disk and is re-applied at startup, since reproducing a bug usually means restarting into it. The frontend facade keeps its untouched `console.*` pass-through (DR-204) and additionally forwards a stringified copy at info and above, so one file holds both halves of the app in order — which is what makes a race between them legible after the fact | Tooling | UR-078 | Done |
+| DR-219 | Release notes are the reviewed CHANGELOG entry, not a generated draft. Every release from v0.0.1 to v0.9.1 published the same ~1,050 bytes of generic install instructions whose "What's New" section said "See CHANGELOG.md" — a link that does not resolve from a release page. Thirty-five releases, byte-identical, telling a reader nothing about what changed. The workflow now publishes the `## <version>` section of CHANGELOG.md and fails the release if that section is absent, since notes that say nothing are worse than a build that waits for two sentences. `release:notes` is printed into the job log as a drafting aid but is deliberately *not* published: CLAUDE.md calls its output "a reviewed draft, not a final changelog", and publishing it unreviewed proved why — a range containing a repo-wide formatting sweep resolved to nearly the entire requirement matrix and produced notes claiming one release had added the whole application. The script now skips cosmetic commits (`chore(format)`, `chore(deps)`, `style`) when deriving a range's files, and says how many it skipped rather than silently reporting a smaller set | Tooling | - | Done |
+| DR-220 | A release ships only its own artifacts. `src-tauri/target/*/release/bundle/` is not versioned, cargo never cleans it, and the CI runner reuses the target directory — so the copy step's `bundle/**/*-setup.exe` glob collected every installer ever built there. Every release from v0.1.0 to v0.8.2 shipped its predecessors': sixteen Windows installers on v0.8.2, thirteen of them stale, and a download list on v0.5.0 reaching back to 0.1.0. It went unnoticed for eight months because there was nothing to notice — the upload loop reported success, the files were real, and the page looked busy rather than wrong. It stopped only when an unrelated cache change wiped the runner's target dir, leaving the defect dormant rather than fixed. Both desktop builds now clear the bundle directory first, so a stale file cannot exist to be copied — filtering the copy by version would have hidden it instead. `scripts/check-release-artifacts.sh` is the backstop for the next route nobody predicts: it runs before the SBOM, the checksums and the upload, and refuses to publish when any artifact's embedded version disagrees with the tag | Tooling | - | Done |
+| DR-221 | The release path is exercised before a tag exists. Nothing in `build-and-test.yml` runs `tauri build` — only a tag does — so a whole class of breakage was invisible until release day, and two instances of it were sitting on master at once. Tauri refuses to build when a plugin's Rust crate and npm package differ by minor version, which the updater and logging work had introduced (`tauri-plugin-log 2.8.0` against `@tauri-apps/plugin-log 2.9.0`) while `cargo check`, clippy, the tests and `svelte-check` all passed; both sides are now pinned exactly rather than by caret, since a caret is what let them separate, and CI runs `tauri info` to compare them without building. The AppImage target had never once been built: linuxdeploy carries a `strip` too old to parse the `.relr.dyn` section modern toolchains emit, so bundling failed on every library — and Ubuntu 23.10+ links with `-z pack-relative-relocs` by default, so the builder image fails the same way a modern Arch host does. `NO_STRIP=true` is linuxdeploy's documented escape hatch; the cost is a larger, unstripped bundle. Both were found by building the target locally before tagging rather than by publishing a release that could not build | Tooling | - | Done |
+| DR-222 | Build tooling matches the package manager the project declares. `scripts/build-android.sh` ran `npm install` on its clean-build path — in a bun project, where `packageManager` says bun and `bun.lock` is the committed lockfile. npm ignores that lockfile, re-resolves the whole tree from package.json, and writes a `package-lock.json` that `.gitignore` then hides. That is not a style preference: the JS halves of the Tauri plugins are pinned exactly against Cargo.lock because the CLI refuses to build when a plugin's crate and package differ by minor version, and a silent re-resolve is precisely how they drift apart. It survived because clean builds are rare — the shape shared by nearly every defect found preparing v0.10.0, where the code running on every commit was healthy and the code running on a release, a tag or a clean build had no guard at all. `scripts/check-tooling.sh` fails on any npm/yarn/pnpm invocation or foreign lockfile | Tooling | - | Done |
+| DR-223 | The Android JavaVM and Application are published into `ndk_context` by this crate, not by a transitive dependency. Seven call sites (five in credentials.rs, two in lib.rs) read that process-global to reach JNI, and nothing here ever set it — `tao` did, three levels below anything this project names in Cargo.toml. tao 0.35.3 moved those pointers into a private struct and stopped publishing them, so the Tauri 2.11 upgrade made the first credential read abort the process on every launch: `PANIC ... android context was not initialized`. Our code had not changed; an undocumented side effect of the windowing layer had gone. The invariant is now owned here rather than assumed: `JNI_OnLoad` captures the JavaVM as the shared library loads, and the Application is resolved lazily via `ActivityThread.currentApplication()` and pinned as a global reference for the process lifetime — the Application rather than the Activity, since that is what `SecureStorage.initialize()` immediately reduces its argument to. Failure degrades to the encrypted-file credential path and is logged, rather than aborting. Found only by installing on a device: nothing in CI runs the app | Security | UR-012 | Done |
+| DR-224 | Backgrounding the app obeys the background-audio toggle on every renderer. The toggle (UR-040) was built for the WebView `<video>` path, where losing visibility kills the decode: it chose between handing off to a native audio stream and letting playback stop. Native video then became the default renderer (DR-188), and on that path playback runs through ExoPlayer inside a `MediaSessionService` — a foreground media service whose purpose is to keep playing while the app is hidden. Nothing paused it and nothing in the codebase paused on background, so locking the screen kept the audio going whether or not the toggle was on: the toggle governed a handoff that no longer had a gap to bridge, and users got background playback they never asked for. The decision now lives in Rust (`player/background_policy.rs`) and both renderers obey it: a video with the toggle off pauses, with the toggle on hands off to audio, music is never paused by backgrounding, and picture-in-picture keeps playing because the window is still on screen (UR-041). It takes no renderer parameter on purpose — the split between the two paths is what produced the defect | Player | UR-040 | Done |
+| DR-225 | `StreamSelection` replaces the bare URL returned for playback: URL, `Transport` (hls / progressive / localFile), `PlaybackKind` (directPlay / directStream / transcode), the negotiated `Rendition`, the ladder this source can offer, and a `needs_transcoding` flag derived in Rust so "which kinds count as transcoding" is answered once. Both enums are serde-tagged (`{"type":"hls"}`) so the frontend matches a discriminant rather than comparing text. The field that mattered most is `transport`: `VideoPlayer.svelte` chose its loader with `url.includes(".m3u8")` in two places, a domain fact reconstructed in the presentation layer — the same class of error as leaking item-type taxonomy, and one that fails silently in both directions (a progressive file served from a path containing the substring gets an HLS loader; a playlist served from one without it does not). The paths that never negotiate — a downloaded file, a live channel — get the same shape from Rust (`media_local_selection`, `LiveStreamInfo.transport`) rather than having the page assemble one, so there is no second place where a transport is decided | Playback | UR-079 | Done |
+| DR-226 | The bandwidth ceiling is two-level: a durable device default (Settings, persisted, restored at startup) and a per-playback override the in-player picker sets. The picker's own documentation had called it a "this film, this connection" control since it was written, but it was implemented by writing the process-wide default — so dropping one awkward film to 2 Mbps silently capped every video played afterwards for the rest of the process, while the Settings screen still displayed the old value and nothing in the UI admitted the change. The override is cleared whenever playback moves to a new item, which is what keeps it from surviving into an autoplayed next episode where nobody would reopen the picker. `effective_streaming_quality()` is the single resolution point; every URL builder and the `PlaybackInfo` negotiation go through it, because a negotiation that authorises a direct play the URL builder then constrains (or the reverse) leaks the cap | Playback | UR-074, UR-079 | Done |
+| DR-227 | The quality picker is filled from what *this* media source can offer, not from the fixed eight-rung enum. Rust marks each rung `exceeds_source` when its ceiling is at or above the source's own bitrate — such a rung produces the same bytes as `Original`, so offering it is another way to spell one choice — and the frontend simply does not draw those. `Original` is never marked (it *is* the source) and a source whose bitrate the server does not report (the sampled library has `avi` files with none) marks nothing redundant, keeping every rung offered, which is the safe direction. The picker also shows what the server is actually doing with the stream, which only became knowable once `PlaybackKind` existed. Labels and detail lines come from Rust beside the numbers they describe, so a relabelled rung cannot drift out of step with what it does | UI | UR-070, UR-079 | Done |
+| DR-228 | Direct play and direct stream are negotiated rather than assumed away. `get_video_stream_url` always built an HLS transcode URL, so every video play burned server CPU even when the file would have played untouched. The decision now comes from `PlaybackInfo` under the device profile and the ceiling in force, with two client-side overrides applied on top because the server's answer is right about the *file* and wrong about what this app will do with it: undecodable audio (Jellyfin 10.11.5 honours a DirectPlayProfile's container and video codec but ignores its audio codec, so it offers direct play for an E-AC-3 track the webview renders in silence) and a viewer-pinned audio track the source file does not default to. Measured against the development server over a 400-item sample: **85% direct play on the Android profile, 7% on the Linux one** — the library is ~80% hevc and WebKitGTK can only claim h264, so the Linux figure is a property of the renderer, not of this code, and is what `linux-native-video-spike.md` exists to change. A direct *stream* is a remux and is deliberately not counted as transcoding | Playback | UR-079 | Done |
+| DR-229 | Mid-playback re-negotiation on throughput was scoped and **dropped on measurement**. The premise — that hls.js gives this app real adaptive bitrate and mpv would lose it — does not hold: a master playlist from the development server carries exactly one `EXT-X-STREAM-INF`, because Jellyfin builds it from the single rendition the request asked for rather than publishing a ladder. There is no adaptation to preserve, so "adapt mid-stream" collapses into "pick well at open", which is what DR-225 and DR-226 already are. Recorded rather than deleted because the conclusion is a measurement, not an opinion, and a server that does publish a ladder would change it — the DR-224 re-negotiation path is the hook that work would build on | Playback | UR-079 | Won't Do |
+| DR-230 | Every player backend consumes the same selection, proving the contract is player-agnostic rather than HTML5-shaped. The queue item carries the negotiated `transport`, so `player_seek_video` picks its seek strategy from the backend's own decision instead of the last `stream_url.contains(".m3u8")` in the codebase; items queued by a path that never negotiated (audio tracks, direct URLs) carry `None` and fall back to `needs_transcoding`, which is exact rather than a guess because every transcode this app requests is HLS (DR-140). The webview adapter's bridge carries the whole selection rather than a URL, so the component's HLS effect reads a tag instead of searching a string, and the background-audio handoff states the transport it is moving to (progressive mp3 out, HLS back) rather than leaving it to be inferred | Playback | UR-003, UR-004, UR-079 | Done |
+| DR-231 | An mpv video backend that composites beneath the transparent webview, the desktop counterpart of the Android TextureView arrangement. mpv renders through its **render API** into an FBO the toolkit binds (`vo=libmpv` + `mpv_render_context_create` with `MPV_RENDER_PARAM_OPENGL_FBO`), rather than by embedding a foreign window — which is what the 2024 "not possible on Wayland at all" conclusion was about and why it does not apply. On Linux that is a `GtkOverlay` with a `GtkGLArea` as main child and Tauri's own webview reparented as the overlay child; the mpv half is shared and only the surface differs per platform. Webview transparency alone suffices — no window-level transparency is used or needed | Playback | UR-080 | Proposed |
+| DR-232 | The mpv render context's lifetime is bound to the GL context it draws into: created on `realize`, freed on `unrealize`, on the same thread, with the update callback unregistered *before* the free so a callback cannot land on a freed context. This is DR-184 on Android restated — a surface outliving its player — and it is a requirement in its own right rather than a fix for a specific crash. The spike observed one SIGSEGV in a decoder thread that three targeted soaks failed to reproduce; what is not in doubt is that the spike never called `mpv_render_context_free` and never tore down on `unrealize`, so nothing defended against the GL context being recreated underneath. Removing the likeliest cause is worth doing whether or not it was the cause | Playback | UR-080 | Proposed |
+| DR-233 | Frame pacing goes through mpv's update callback, with `mpv_render_context_report_swap` after each render. Recorded as a requirement because the failure mode misleads: driving the widget's frame clock every tick without reporting the swap leaves mpv with nothing to time against, which looks fine in a window and **judders at fullscreen** — reading as a compositing or GPU limit and being neither | Playback | UR-080 | Proposed |
+| DR-234 | The device profile is derived from the **renderer that will decode the stream**, not from a compile-time platform constant. `video_codecs` was `#[cfg(target_os)]`, which is correct only while a build has one video renderer; once mpv and the webview element coexist it must be runtime state. This is the change that converts the measured 7% desktop direct-play rate toward the 85% the Android profile achieves on the same library, because the two differ by nothing except which component decodes. It looks like configuration and is not — it is the input that decides whether the server re-encodes, and getting it wrong fails silently, a claimed codec the renderer cannot decode being a black picture or silence (DR-148, and DR-227's audio override). The webview's narrower *audio* set stops applying to the video path once mpv decodes it, while the multichannel bound still does, since a 5.1 track direct-played into a two-channel sink is silence or inaudible dialogue | Repository | UR-080, UR-070 | In Progress |
+| DR-235 | The webview video path is deleted, not merely bypassed. Staged, because a path cannot be removed while a shipped platform still needs it: Linux moves to mpv first, Windows follows, and only then do `hls.js`, `html5Adapter.ts`, `videoLoaderFor` and the `<video>` element go. The staging is the point — a Linux-only version would leave the fork alive permanently, taking video from three renderers to four and giving every seek strategy, track switch and lifecycle bug one more place to be got right. Android keeps ExoPlayer and keeps the webview as its documented opt-out; the background-audio `<audio>` path is untouched. With no HTML5 fallback left, a failed mpv init emits `backend-init-failed` and surfaces a real error rather than silently degrading to the transcode this work exists to stop paying for | Playback | UR-080 | Proposed |
+| DR-236 | Hardware-decode policy is decided from what mpv reports it **selected** (`hwdec-current`), never from what it was asked for. The spike established that hardware decode works through the render API at all — the load-bearing result, since it means direct play is not bought with software decoding — but also that `auto` reached for the discrete GPU in copy-back mode on a hybrid Intel+NVIDIA laptop, the least efficient hardware path, and that `vaapi` fell back to software silently because the libva driver was absent. So zero-copy VA-API on the integrated GPU is preferred where the driver is present, `auto` is a fallback rather than the default, and a missing driver is detected and logged rather than mistaken for a compositing limit | Playback | UR-080 | Proposed |
+| DR-237 | Windows reaches the same mpv path, reusing everything except the surface. The surface is genuinely different code — a native child window beneath a transparent WebView2, not GTK — but the render context, lifetime discipline, frame pacing, device profile and hwdec policy are shared, which is why none of them may be guarded on `cfg!(target_os = "linux")`. The cost is mostly build, not video: `libmpv` is currently a Linux-only dependency while Windows is cross-compiled from Linux via `x86_64-pc-windows-msvc` + `cargo-xwin`, so a Windows libmpv must reach that cross-build and its DLL must ship in the NSIS bundle, carrying the LGPL obligations DR-216 already records — dynamic linkage, licence text shipped alongside. Windows gains a native audio decoder as a side effect, which is what the long-blocked Windows audio work wants and cannot otherwise have | Playback | UR-080 | Proposed |
+| DR-198 | The webview runs under a real Content-Security-Policy, and the asset protocol is scoped to the one directory it still serves. `csp` was `null`, which disables CSP entirely: any script that reached the web layer — through a future `{@html}`, a dependency, or a devtools paste — would have inherited the whole IPC surface, and with it the user's session. `script-src 'self'` (Tauri injects a nonce for SvelteKit's inline bootstrap script at build time, so no `'unsafe-inline'` is needed) plus `object-src`/`frame-src 'none'` and `base-uri 'self'` is the part that is genuinely restrictive. `img-src`/`media-src`/`connect-src` cannot be: the Jellyfin origin is typed in by the user at run time and is commonly plain `http` on a LAN, so they allow `http:`/`https:` — a wide grant for *data*, but one that still bars `file:`, `filesystem:` and scripting schemes, and leaves `script-src` untouched. `style-src` keeps `'unsafe-inline'` because Svelte compiles `style="…"` attributes (including `app.html`'s `display: contents` wrapper) into markup; this is safe only while no `<style>` element survives into `index.html`, since a nonce there would make Tauri's injection outrank — and therefore void — `'unsafe-inline'`. `worker-src blob:` and `media-src blob:` are hls.js: it demuxes in a worker built from a blob and attaches MSE through `URL.createObjectURL`. `asset:` and `http://asset.localhost` are the same protocol under the two naming schemes `convertFileSrc` emits (custom scheme on Linux/macOS, `http` host on Windows/Android); `ipc:`/`http://ipc.localhost` is the invoke transport, which would otherwise be blocked by `connect-src`. A run-time CSP naming the server origin exactly was rejected: Tauri computes the header from immutable config when it serves the HTML, so it would mean rebuilding config and reloading the webview on every server change, for a policy the user can already point anywhere. The asset-protocol scope narrows from `$APPDATA/**` to `$APPDATA/thumbnails/**` — since DR-137 moved downloaded media to the loopback server, `imageCache` is the only `convertFileSrc` caller left, so the database and the encrypted-token fallback file no longer sit inside the grant | Security | UR-012, UR-071 | Done |
 
 ---
 
@@ -253,29 +444,29 @@ 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-006 | IR-005, IR-006, IR-007, IR-008 | - |
+| 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, DR-196 |
+| 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, DR-203 |
+| 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 | DR-200, DR-201 |
 | UR-007 | IR-010 | DR-007, DR-008, DR-016 |
 | UR-008 | IR-010 | DR-007, DR-011 |
 | UR-009 | IR-009, IR-010, IR-011 | - |
 | UR-010 | IR-012, IR-021 | DR-037, DR-059 |
 | UR-011 | IR-013 | DR-003, DR-015, DR-018 |
-| UR-012 | IR-009, IR-014 | - |
+| UR-012 | IR-009, IR-014 | DR-198 |
 | UR-013 | IR-013 | DR-017 |
 | UR-014 | IR-010 | DR-014, DR-019 |
 | 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-021 | IR-016, IR-019 | DR-024 |
+| UR-020 | IR-016, IR-018 | DR-023, DR-176 |  <!-- IR-018 delivered by ExoPlayer + HTML5 `<track>`, not libmpv -->
+| UR-021 | IR-016, IR-019 | DR-024 |  <!-- IR-019 delivered by ExoPlayer + HLS stream re-open, not libmpv -->
 | 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 |
@@ -290,25 +481,46 @@ 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, DR-196, DR-201, DR-203 |
+| 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 |
+| 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, DR-198, DR-199 |
+| UR-072 | - | DR-156 |
+| UR-073 | - | DR-158 |
+| UR-074 | - | DR-162, DR-177, DR-181 |
+| UR-075 | - | DR-174, DR-175 |
+| UR-076 | - | DR-209 |
+| UR-077 | - | DR-217 |
+| UR-078 | - | DR-218 |
+| UR-079 | - | DR-225, DR-226, DR-227, DR-228, DR-229, DR-230 |
+| UR-080 | IR-033 | DR-231, DR-232, DR-233, DR-234, DR-235, DR-236, DR-237 |
 
 ---
 
@@ -379,6 +591,7 @@ 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 | 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 |
@@ -398,6 +611,134 @@ Internal architecture, components, and application logic.
 | 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-188 | The control-bar auto-hide rule permits hiding only during uninterrupted playback: it declines while paused, while a seek is in flight, and while a track/subtitle/quality menu is open — asserted against the pure `shouldHideControls` rule rather than a clock or a DOM | DR-189 | 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 |
+| UT-190 | `build_next_up_endpoint` sends `EnableResumable=false` with the user and limit, and no `SeriesId` filter when none was requested | DR-197, JA-036 | Done |
+| UT-191 | A per-series next-up query keeps `SeriesId` and the resumable exclusion, and defaults the limit | DR-197 | Done |
+| UT-192 | `filterInProgressNextUpItems` drops an episode present in the resume list, keeps the genuinely unstarted next episode, leaves the rest of the row intact, and is a no-op when nothing is in progress | DR-197 | Done |
+| UT-193 | The shipped Tauri security config stays restrictive: `csp` is set, `script-src` carries no `'unsafe-inline'`/`'unsafe-eval'`/wildcard, `object-src`/`frame-src` are `'none'`, the directives playback needs (asset scheme, loopback, `blob:`, `ipc:`) are present, and the asset-protocol scope covers only the thumbnail cache — never the storage root that holds the database | DR-198 | Done |
+| UT-194 | Normal audio (no background-audio handoff) keeps queue advance on both skip buttons | DR-201 | Done |
+| UT-195 | In background-audio mode a skip scrubs +30s/-10s instead of advancing the queue — the reported defect | DR-201 | Done |
+| UT-196 | Skipping back near the start clamps to zero rather than seeking negative | DR-201 | Done |
+| UT-197 | Skipping forward near the end clamps to the duration rather than running past it into an EOF-driven advance | DR-201 | Done |
+| UT-198 | An unknown duration still scrubs and still refuses to go negative | DR-201 | Done |
+| UT-199 | The screen-wake decision: video playing holds the display, pausing releases it, audio playing never holds it, a webview element going inactive releases even without a pause report, either renderer alone is enough to hold, and teardown drops both | DR-202 | Done |
+| UT-201 | The logging facade gates by level: a message below the active level is not emitted at all, one at or above it reaches the sink, changing the level at run time changes what passes without touching the call sites, and a scoped logger tags its output with the subsystem | DR-204 | Proposed |
+| UT-202 | Generated traceability-matrix file links resolve from `docs/`: an emitted href, resolved against the directory `traceability.md` is written to, points at a file that exists on disk; the visible link text stays repo-root-relative; the `#Lnn` anchor survives; and a bare repo-root href — the regression that made every link 404 as `docs/<path>` — is rejected | DR-093 | Done |
+| UT-203 | Library folder exclusion filters by id, not by name: an excluded folder's items are absent from a music query, an item whose *title* merely contains an excluded folder's name is kept, and clearing the exclusion restores the items | DR-209 | Proposed |
+| UT-204 | Thumbnail cache writes stay inside the cache directory: a traversal-style and an absolute `item_id` both fail to produce a file outside it, a filename made only of already-safe characters is byte-identical to the one the previous code produced, and an odd id still round-trips through `get_cached_path` | DR-210 | Done |
+| UT-205 | Queued download paths cannot escape the download root — traversal, absolute and `..` forms are refused — while the four real path shapes the app builds, including the absolute one `download_series` produces, come back unchanged; and a completed download cannot register a file outside the root | DR-211 | Done |
+| UT-206 | The offline item-type filter is bound rather than interpolated (a value containing a quote and `OR 1=1` matches nothing instead of disabling the `WHERE`), `build_get_items_endpoint` percent-encodes its values while preserving the commas Jellyfin splits on, and volume normalisation clamps out-of-range input and maps NaN to a finite value | DR-212 | Done |
+| UT-200 | The stream a player could only restart is refused its retry: the handoff transcode answers yes to `player_retry_restarts_stream` while music, video and a downloaded episode answer no, and the Kotlin decision starts permissive, flips on a non-resumable load, and is restored by the next ordinary one | DR-203 | Done |
+| UT-207 | The hero banner's rotation timer restarts from the moment of a manual change: a swipe 5.5s into a 6s interval waits a further 6s instead of firing the leftover 500ms, repeated restarts never stack timers, and `stop()` ends rotation | DR-038 | Done |
+| UT-208 | The update decision: each numeric version field is compared in order, the installed version is not offered to itself, a leading `v` is tolerated because that is how the tags are written, a pre-release sorts below the release of the same number so 0.9.2-rc1 is not offered to somebody on 0.9.2, a missing patch field reads as zero rather than NaN, mobile reports link-only while desktop reports install, and absent release notes normalise to null rather than undefined | DR-217 | Done |
+| UT-209 | Redaction and forwarding. Rust: every credential shape reduces to `[REDACTED]` while the host, username and neighbouring parameters survive; redaction is idempotent, leaves ordinary lines alone, does not fire on the word "token" in prose, and does not panic on multi-byte input; a server URL keeps only scheme and host and drops an embedded `user:pass@`; an unparseable level falls back to info rather than failing at startup. Frontend: info and above forward while debug does not, a message the level filter suppressed is not forwarded, a throwing forwarder neither propagates nor prevents the console write, and an `Error` renders as name and message rather than the `{}` that `JSON.stringify` produces | DR-218 | Done |
+| UT-210 | Cosmetic-commit detection for release notes: a `chore(format)`, `chore(deps)` or `style` subject is skipped when deriving a range's changed files, while `fix`, `feat`, `ci`, `docs`, a bare `chore:` and `chore(release):` are kept; and the word "format" appearing later in a subject ("fix(duration): format times over 24 hours") does not make a real fix look cosmetic | DR-219 | Done |
+| UT-211 | The background decision: a video with the toggle off pauses (the reported defect, where the media service kept playing regardless), a video with it on hands off to audio, music keeps playing whatever the toggle says because it has no picture to lose, picture-in-picture keeps playing in every combination since the window is still visible, and the answer does not vary by renderer | DR-224 | Done |
+| UT-212 | The stream-selection contract. `Transport` and `PlaybackKind` each serialise to exactly the tag the frontend matches (`{"type":"hls"}`, `{"type":"directPlay"}`, …) and round-trip; nested `StreamSelection` fields are camelCase on the wire including `playbackKind`, `mediaSourceId` and `maxBitrate`; only `Transcode` counts as transcoding, so a direct stream does not; a local file is a direct play over a local transport with no ladder. The ladder: every rung at or above a 1.12 Mbps source is marked redundant while the three that constrain it are not, `Original` is never marked for any bitrate including zero and unknown, an unreported source bitrate keeps all eight rungs offered, a 40 Mbps source marks none, and each option carries the ladder's own label and detail | DR-224, DR-226 | Done |
+| UT-213 | The direct-play negotiation, one test per branch, against `PlaybackInfo` fixtures whose shapes were all observed on a live server: a supported source direct-plays; a remuxable one direct-streams and reports itself as *not* transcoding; an unsupported codec transcodes; undecodable audio overrides the server's direct-play offer (silent picture is worse than a transcode); a pinned audio track forces a transcode; a ceiling below the source bitrate transcodes even though the codec is fine, and the ladder agrees that rung constrains it; direct play wins over direct stream when both are offered. Plus the ceiling: a per-playback override governs the stream being opened without disturbing the durable default the Settings screen shows, and dropping it returns to that default | DR-225, DR-227 | Done |
+| UT-214 | The loader comes from the transport, never the URL. hls.js is attached for `hls` when available and the element's own loader when not; progressive and local files load directly; the element's `src` is emptied only when hls.js drives it. The two cases that fail against a substring check, and the reason the field exists: a `progressive` stream whose URL contains `.m3u8` is *not* given an HLS loader, and an `hls` stream whose URL contains no `.m3u8` *is*. Both failed against the pre-DR-225 implementation before the fix landed | DR-224 | Done |
+| UT-215 | Waiting for the repository rather than racing it: it resolves immediately when the session is already restored, resolves when the session arrives later (the race the player page lost on mount), still rejects when there genuinely is no session, unsubscribes once settled so a later store change cannot re-settle it, and leaves no armed timer to reject an already-resolved promise | DR-013 | Done |
 
 ### Integration Tests
 
@@ -410,8 +751,8 @@ Internal architecture, components, and application logic.
 | IT-005 | MPRIS lockscreen controls on Linux | IR-005, UR-006 | Pending |
 | IT-006 | Offline mode with local database | IR-013, UR-002 | Pending |
 | IT-007 | Media download and local playback | DR-015, UR-011 | Pending |
-| IT-008 | Subtitle track selection via libmpv | IR-018, UR-020 | Pending |
-| IT-009 | Audio track selection via libmpv | IR-019, UR-021 | Pending |
+| IT-008 | Subtitle track selection on the video backends (ExoPlayer sideloaded tracks; HTML5 `<track>` children) — *not* via libmpv, which does not implement it | IR-018, UR-020 | Pending |
+| IT-009 | Audio track selection on the video backends (ExoPlayer track switch; HTML5 stream re-open at the chosen `AudioStreamIndex`) — *not* via libmpv, which does not implement it | IR-019, UR-021 | Pending |
 | IT-010 | Playback progress sync to Jellyfin | IR-015, UR-025 | Pending |
 | IT-011 | Resume playback from server position | IR-015, UR-019 | Pending |
 | IT-012 | Equalizer bands via libmpv | IR-020, UR-027 | Pending |
@@ -423,6 +764,50 @@ Internal architecture, components, and application logic.
 
 ## 5. Technical Debt
 
+### Open items carried over from the v0.6.0 codebase audit
+
+The 2026-08-16 audit (v0.6.0, commit `be907b49`) was a point-in-time snapshot
+with no status markers, and by v0.8.2 most of it had been either fixed or
+overtaken. It was **retired** rather than left to rot into a document that
+half-describes the code: what survived it is the table below, which is now the
+record. Each row is self-contained — the audit is not needed to act on it.
+
+What was dropped as demonstrably closed, so it is not re-raised: the CSP and
+asset-protocol scope findings (now DR-198), cloud backup and credential restore,
+the WebView mixed-content override (DR-199), `POST_NOTIFICATIONS` and the
+media-session exemption (DR-200), the `jvmTarget` 1.8 pin (now 17), the
+half-declared Android TV leanback category (removed), the untraced-but-Done
+requirements and the contradictory UR/IR statuses (re-scoped in §2.1), the 50%
+traceability gate (ratcheted, and gated on a live denominator by DR-093), the
+flaky `offlineCatalog` test, the clippy warning backlog (cleared, and `cargo
+fmt --check` plus clippy now run in CI), and the "820 production `unwrap()`s"
+figure — a measurement error that counted test modules, corrected in the audit
+itself to ~19 and standing at 27 today, none of them in a command handler. The
+three `Runtime::new().unwrap()` sites that genuinely matter survive as row 5.
+
+Ordered by what would hurt most if left.
+
+> **Closed 2026-08-17:** the R8-minified release APK was validated on device.
+> That was the last item gating confidence in the v0.8.0 release itself; R8
+> stripping JNI-loaded classes has broken release builds here before, and
+> v0.8.0 added a new Kotlin path (`onFastForward`/`onRewind`) that the
+> unminified debug pass did not cover.
+
+| # | Item | Why it matters | Size |
+|---|------|----------------|------|
+| 1 | **Android 16 Local Network Protections** | The rare platform change that could stop the app working at all: JellyTau's core function is reaching a Jellyfin server that, for most users, is on the LAN. Opt-in for testing in Android 16, enforcement signalled for a later release — so nothing is broken today and no device test will surface it. Far cheaper to handle before it is mandatory. An Android 16 device is already to hand to test the opt-in flag against | M |
+| 2 | **The traceability matrix cannot see Kotlin** | `scripts/extract-traces.ts` walks only `src`, `src-tauri/src` and `scripts`, so every `TRACES:` comment in `src-tauri/android/**` is invisible — pre-existing ones included. A whole platform is unmeasured, which is plausibly why the Android IRs sat untagged for so long, and it means the 90% coverage figure is computed over a codebase that excludes the Android tree | S |
+| 3 | **Delete the asset protocol outright** | It is not narrowly used, it is **unused**. `getCachedImageUrl` has no production callers (only its own test file), so `convertFileSrc` never executes; images arrive as base64 `data:` URIs from `image_get_url`. Confirmed on device: zero `asset.localhost` requests across a full browsing session. Dropping `protocol-asset` and the `assetProtocol` block retires the surface instead of shrinking it, and `imageCache.ts` goes with it | S |
+| 4 | **Tighten `img-src`** | The v0.8.0 CSP grants `img-src … http: https:` on the premise that thumbnails are fetched direct-from-server by the webview. They are not (see #3). With no webview-side server image loads anywhere in `src/`, `'self' data: blob:` should suffice. Needs its own device pass — a wrong `img-src` blanks every image, silently | S |
+| 5 | **Three `Runtime::new().unwrap()` in playback-critical threads** | `session_poller/mod.rs:102`, `player/mpv_backend.rs:424`, `player/android/mod.rs:761`. A panic strands the app offline with nothing surfaced, freezes the scrubber mid-playback, or kills progress reporting across a JNI boundary. One shared helper returning `Option<Runtime>` and logging on failure retires all three. (The wider "820 unwraps" figure was a measurement error — the real count is 19, and none are in command handlers) | S |
+| 6 | **Confirm the playback service rejects unknown callers** | `JellyTauPlaybackService` is `exported="true"` with a `MediaSessionService` intent filter — conventional for Media3, but it means any app on the device can attempt to bind and drive playback. The session's `onConnect` should reject unknown packages. (Predictive back, raised alongside this, was verified working on device and needs nothing) | S |
+| 7 | **Media3 is several minor versions behind** | Pinned at 1.5.0 across exoplayer/hls/session/common. Much of this app's hard-won behaviour lives in ExoPlayer edge cases — truncated progressive streams, background-audio handoff, HLS resume — so its bug-fix releases have unusually high value here. Schedule with a device pass over the playback regression list | M |
+| 8 | **Shipped desktop bundles have no update path** | deb/rpm/nsis are built but `tauri-plugin-updater` is absent, so every desktop user upgrades by manually fetching a package — in practice a long tail of installs pinned to whatever they first downloaded. Add the updater with a signed manifest, or document the manual path so the omission is deliberate | M |
+| 9 | **`DR-042` overstates what ships** | It promises "poster cards, year, **and rating badges**", but `MediaCard.svelte` renders only `productionYear`; `CommunityRating`/`OfficialRating` appear solely as sort keys, never as a badge. Either build the badge or correct the requirement text — a requirement that describes unbuilt behaviour is worse than an untraced one | S |
+| 10 | **Stray duplicate `JellyTauPlayer.kt`** | A copy exists at `src-tauri/android/app/src/main/java/.../player/JellyTauPlayer.kt`, outside the canonical `src-tauri/android/src` tree that `sync-android-sources.sh` reads. Two files with one name in a tree with a strict canonical-source rule is a trap for the next edit | S |
+| 11 | **Six modules carry a disproportionate share of the complexity** | `src-tauri/src/player/mod.rs` (4,732 lines), `src-tauri/src/repository/offline.rs` (4,705), `src-tauri/src/repository/online.rs` (3,760), `src-tauri/src/commands/player/mod.rs` (3,327), `src-tauri/src/commands/download/mod.rs` (3,238) and `src/lib/components/player/VideoPlayer.svelte` (2,786) — all still growing. The cost is not the line count itself, it is that **these are the same modules `CLAUDE.md`'s Gotchas section keeps having to warn about**: the deadlock rule about locking in event callbacks, the `AutoplayDecision` scrutinee, the "no lifecycle calls after an `await` in `onMount`" rule, the HLS `master.m3u8` rule, the download concurrency cap. A file that needs a standing warning in the project's onboarding document is a file whose invariants are no longer local to it, and every such warning is a rule a newcomer has to be *told* rather than one the structure enforces. **Recorded, not scheduled** — a speculative refactor of six files this size buys nothing on its own. The trigger is the next time one of them needs substantial work: splitting it then is likely cheaper than growing it, and each rule that moves from Gotchas into a module boundary is one fewer thing to remember | L |
+
+
 ### Linux Keyring Integration Workaround
 
 **Issue**: The `keyring-rs` crate (v3.x) has issues with retrieving credentials from the Linux Secret Service API, despite successfully saving them.
@@ -467,18 +852,18 @@ Linux-specific `secret-tool` save/get/delete paths.
 
 **Issue**: The Linux (MPV) and Android (ExoPlayer) playback backends have diverged in feature implementation and architecture patterns.
 
-**Symptoms**:
-- Audio settings (crossfade, gapless playback, volume normalization) work on Linux but not on Android
+**Symptoms** (as first recorded; the audio half is now closed — see Status):
+- Audio settings (gapless playback, volume normalization, equalizer) worked on Linux but not on Android
 - Position update frequency differs between platforms (Linux: 250ms polling, Android: on-demand callbacks)
 - Thread safety models differ (Linux: `Arc<Mutex<>>`, Android: global `OnceLock` statics)
 
 **Root Cause**:
-The `PlayerBackend` trait defines optional audio settings methods with default empty implementations. The Linux `MpvBackend` overrides these with full MPV property commands, but `ExoPlayerBackend` uses the defaults.
+The `PlayerBackend` trait defines optional audio settings methods with default empty implementations. `MpvBackend` overrode these with MPV property commands; `ExoPlayerBackend` took the silent defaults, so the Settings › Audio panel rendered controls that did nothing on Android. `ExoPlayerBackend` now overrides them too, but the trait default is still a silent `Ok(())` — a backend that omits the method still reports success rather than failing loudly.
 
 **Affected Files**:
-- [src-tauri/src/player/backend.rs](../src-tauri/src/player/backend.rs) - Trait with default empty implementations
+- [src-tauri/src/player/backend.rs](../src-tauri/src/player/backend.rs) - Trait; defaults still return `Ok(())` silently
 - [src-tauri/src/player/mpv_backend.rs](../src-tauri/src/player/mpv_backend.rs) - Full audio settings support
-- [src-tauri/src/player/android/mod.rs](../src-tauri/src/player/android/mod.rs) - Missing audio settings implementation
+- [src-tauri/src/player/android/mod.rs](../src-tauri/src/player/android/mod.rs) - Audio settings carried to Kotlin as JSON over JNI
 
 **Feature Parity Matrix**:
 
@@ -487,22 +872,36 @@ 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/architecture/05-platform-backends.md, "Audio settings on ExoPlayer"):
+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
 
@@ -520,7 +919,7 @@ The `PlayerBackend` trait defines optional audio settings methods with default e
 **Affected Files**:
 - [src/lib/components/player/AudioPlayer.svelte](../src/lib/components/player/AudioPlayer.svelte) - Duplicate handlers
 - [src/lib/components/player/MiniPlayer.svelte](../src/lib/components/player/MiniPlayer.svelte) - Duplicate handlers
-- [src/lib/services/playbackControl.ts](../src/lib/services/playbackControl.ts) - Position conversion
+- [src/lib/utils/playbackUnits.ts](../src/lib/utils/playbackUnits.ts) - Position conversion (the shared helper the "Future Fix" below called for; `playbackControl.ts`, previously listed here, has since been removed)
 - [src/lib/stores/playbackMode.ts](../src/lib/stores/playbackMode.ts) - Position conversion
 - [src/lib/services/playbackReporting.ts](../src/lib/services/playbackReporting.ts) - Position conversion
 

docs/specs/README.md

@@ -0,0 +1,81 @@
+# Specs index
+
+Feature specs for JellyTau. Start a new one from
+[SPEC-TEMPLATE.md](SPEC-TEMPLATE.md) and run it past
+[SPEC-REVIEW-CHECKLIST.md](SPEC-REVIEW-CHECKLIST.md) before accepting it.
+
+## What lives here
+
+**Only work that has not shipped.** Once a spec is fully implemented its design
+is folded into the architecture docs — which are the maintained description of
+the build — and the spec file is deleted. Git history keeps the original,
+including its rejected alternatives and acceptance criteria; the architecture
+docs keep the reasoning that a future change still needs.
+
+So: a file in this directory is a **promise, not a description**. If you want to
+know how something *works*, read
+[docs/architecture/](../architecture/README.md). If you want to know what is
+*planned*, read here.
+
+**Status vocabulary**
+
+| Status | Meaning |
+|---|---|
+| Proposed | Written, not accepted. Nothing built. |
+| Accepted | Agreed as the design; implementation not started or not finished. |
+| Partially implemented | Some parts shipped; the spec names what is left. |
+| Design authority | No code of its own — it records a decision later specs act on. |
+
+**Next free requirement ids** (always re-check
+[requirements.md](../requirements.md) before allocating): **UR-079**,
+**IR-033**, **DR-232**. Three specs below suggested ids that have since been
+taken by other work; each carries a ⚠️ note at the top.
+
+## Partially implemented
+
+| Spec | What landed | What is left |
+|---|---|---|
+| [frontend-domain-model.md](frontend-domain-model.md) | Catalog surface: `MediaKind`, `from_jellyfin` isolated, ticks → ms | `primaryImageTag` → `imageId` (~30 sites); player/session/reporting tick math; `stream.type` |
+| [libmpv2-migration.md](libmpv2-migration.md) | `LICENSE` | The `libmpv` → `libmpv2` crate swap |
+| [read-through-media-cache.md](read-through-media-cache.md) | DR-126…128, DR-133…138 — cache entries *are* download rows; local playback of downloads | DR-122/124/125 — the read-through capture. DR-121 shipped as backend-owned stream selection and left this spec |
+| [scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md) | Stage 1: `SearchScope` owned by Rust (DR-063…067) | Stage 2: result-side grouping (`GROUP_ITEM_TYPES` still in `searchScope.ts`) |
+
+## Not started
+
+| Spec | Blocked on / note |
+|---|---|
+| [desktop-native-video.md](desktop-native-video.md) | mpv draws video on every desktop platform, then the webview `<video>` path and hls.js are deleted. Converts a measured 7% direct-play rate toward Android's 85%. Stacked on backend-owned stream selection. |
+| [build-provenance.md](build-provenance.md) | `build.rs` is still bare. ⚠️ suggested id DR-093 is taken. |
+| [player-facade-enforcement.md](player-facade-enforcement.md) | ~60 `commands.player*` sites still outside the facade; no lint rule. ⚠️ suggested id DR-095 is taken. |
+| [windows-native-audio-backend.md](windows-native-audio-backend.md) | Blocked on the libmpv2 swap. ⚠️ suggested id IR-030 is taken. |
+| [linux-native-video-spike.md](linux-native-video-spike.md) | **Spike run 2026-08-21: compositing works on Linux, X11 and Wayland.** G1-G6 green bar the Tauri `default_vbox()` half of G1. The adaptive-bitrate question it was waiting on is **answered**: the server publishes one `EXT-X-STREAM-INF`, so there is no ladder for mpv to lose (DR-229). `StreamSelection` (DR-225) is the contract to consume. |
+
+## Design authority
+
+| Spec | Role |
+|---|---|
+| [playback-backend-unification.md](playback-backend-unification.md) | Why video cannot unify onto one native engine and audio can. The audio half has since shipped on Android; Windows has not. |
+| [scoped-search-boundary.md](scoped-search-boundary.md) | The boundary design the `check:boundary` rule came from. Stage 1 built. |
+| [scoped-search.md](scoped-search.md) | Superseded in part — its "frontend only, no Rust changes" decision is the leak the boundary spec reversed. UX still current. |
+
+## Where the shipped specs went
+
+Sixteen specs were folded into the architecture docs and deleted (2026-08-21).
+Where to look for each:
+
+| Shipped work | Now documented in |
+|---|---|
+| Account menu & global chrome | [02-svelte-frontend.md](../architecture/02-svelte-frontend.md) — App Shell and Chrome |
+| Library mosaic | [02-svelte-frontend.md](../architecture/02-svelte-frontend.md) — Library Mosaic |
+| Series current-episode navigation | [02-svelte-frontend.md](../architecture/02-svelte-frontend.md) — Series and Episode Navigation |
+| Downloads as an offline library | [02-svelte-frontend.md](../architecture/02-svelte-frontend.md) — Downloaded Browse |
+| Favourites browsing | [01-rust-backend.md](../architecture/01-rust-backend.md) — Favorites System |
+| Streaming bitrate cap | [01-rust-backend.md](../architecture/01-rust-backend.md) — Streaming quality ladder |
+| Locally-indexed search | [03-data-flow.md](../architecture/03-data-flow.md) — Search Flow; [01-rust-backend.md](../architecture/01-rust-backend.md) — Background workers |
+| Offline downloaded-only filter | [06-downloads-and-offline.md](../architecture/06-downloads-and-offline.md) — Offline Catalog Visibility |
+| Audio equalizer · Android audio settings parity | [05-platform-backends.md](../architecture/05-platform-backends.md) — Audio settings on ExoPlayer |
+| Android native video spike | [05-platform-backends.md](../architecture/05-platform-backends.md) — Native Video Compositing |
+| Video background audio | [05-platform-backends.md](../architecture/05-platform-backends.md) — Background Audio Handoff |
+| Traceability gate repair | [traceability-ci.md](../traceability-ci.md) |
+| Boundary tripwire hardening | `scripts/check-frontend-boundary.sh` (its header is the spec) |
+| Playback docs corrections · req-coverage script removal | Nothing to document — both were corrections that have been applied |

docs/specs/SPEC-REVIEW-CHECKLIST.md

@@ -50,7 +50,20 @@ Copy the boxes into the review comment (or the PR) and tick them.
 - [ ] Linked to existing URs, or new URs/DRs are allocated in
       [requirements.md](../requirements.md).
 - [ ] Requirement-implementing code will carry `// TRACES:` comments (CLAUDE.md).
-- [ ] Traceability coverage stays ≥ 50% (the CI gate).
+- [ ] Traceability coverage stays ≥ 88% (the CI gate — a ratchet, so check
+      `bun run traces:coverage` rather than trusting this number).
+
+## Lifecycle
+
+- [ ] **"Destination on completion" names a real architecture doc and section.**
+      This spec file is deleted when it ships; something has to absorb the
+      design. If nothing fits, the layer assignment is probably unclear — go back
+      to that table.
+- [ ] The spec separates the **durable half** (invariants, rejected alternatives,
+      the defect a decision exists to prevent) from the **disposable half**
+      (phases, migration steps, acceptance criteria). Only the first is folded in.
+- [ ] Anything listed as out of scope but still worth doing is written where it
+      will be found after this file is gone — beside the code it concerns.
 
 ## Conflicts & hygiene
 

docs/specs/SPEC-TEMPLATE.md

@@ -6,12 +6,27 @@
   "Layer assignment" — read its comment before writing it.
 
   Before merging a spec, run it past docs/specs/SPEC-REVIEW-CHECKLIST.md.
+
+  LIFECYCLE: this file is temporary. docs/specs/ holds only unshipped work — when
+  the last acceptance criterion is met, the design is folded into
+  docs/architecture/ and this file is deleted in the same commit. Write it
+  knowing that: the durable half is the reasoning (invariants, rejected
+  alternatives, the defect a decision prevents), and the disposable half is the
+  plan (phases, migration steps, acceptance criteria).
 -->
 
-**Status:** Proposed <!-- Proposed | Accepted | Implemented | Superseded -->
+**Status:** Proposed <!-- Proposed | Accepted | Partially implemented | Superseded.
+     NOT "Implemented" — a fully shipped spec is folded into docs/architecture/
+     and deleted. See "Destination on completion" below. -->
 **Requirements:** <!-- UR-xxx → DR-yyy; allocate new DRs in requirements.md. -->
 **UX spec:** <!-- link to the relevant ux-flows.md section, or "n/a". -->
 **Supersedes / revises:** <!-- link any spec this changes, or delete this line. -->
+**Destination on completion:** <!--
+     Which architecture doc absorbs this design when it ships, and roughly which
+     section. e.g. "05-platform-backends.md — a new section beside
+     ExoPlayerBackend". Name it NOW: a feature that fits no existing doc usually
+     has an unclear layer assignment, which is worth finding out at spec time.
+     This spec file is deleted in the same commit that folds it in. -->
 
 ## Summary
 

docs/specs/account-menu.md

@@ -1,164 +0,0 @@
-# Spec: Account menu and global chrome availability
-
-**Status:** Implemented
-**Scope:** Frontend only. No Rust changes required.
-**Requirements:** UR-054 → DR-075, DR-076, DR-077 (see
-[requirements.md](../requirements.md)).
-**UX spec:** [ux-flows.md §1.2–1.4](../ux-flows.md).
-
-## Summary
-
-Account actions — Settings, Downloads, Display preferences, Sign out — are
-currently reachable **only from `/library/*`**. Move them into a single shared
-account menu anchored to the user's name, and make that menu available on every
-authenticated non-immersive screen.
-
-## Motivation
-
-A user sitting on the home screen cannot open Settings or sign out. The bottom
-nav offers Home / Search / Library only, and the header that hosts those actions
-belongs to the library layout. The user has to guess that account actions live
-*inside* Library — an unrelated section — and navigate there first.
-
-Desktop and mobile also disagree today: desktop shows an unlabeled logout icon
-with no grouped menu, mobile shows a three-dot overflow with labelled items. The
-same two actions are found two different ways.
-
-## Background: verified current state
-
-1. **The header is not global.** It is defined in
-   [library/+layout.svelte](../../src/routes/library/+layout.svelte). The root
-   layout [+layout.svelte](../../src/routes/+layout.svelte) renders no header at
-   all.
-
-2. **`routeOwnsLayout`** in
-   [layoutShell.ts](../../src/lib/utils/layoutShell.ts) returns true for
-   `/library`, `/player/`, `/login` — those routes own their own full-height
-   flex column. Everything else renders into the root scroller with the root's
-   `BottomUi` below it.
-
-3. **Bottom nav is Home / Search / Library only**
-   ([BottomNav.svelte](../../src/lib/components/BottomNav.svelte)) — no Settings
-   or account entry.
-
-4. **Net effect:** on `/`, `/search`, and `/downloads` there is no route to
-   Settings or Sign out.
-
-5. **Desktop username is inert text** — a `<span>` next to the icons, not a
-   trigger.
-
-6. **The mobile overflow menu already has the right contents** (Downloads,
-   Settings, divider, Sign out) and the right dismissal behaviour (backdrop
-   click, keyboard handler). **Extract and reuse it rather than rewriting it.**
-
-7. **`viewMode` is already a persisted store** in
-   [library.ts](../../src/lib/stores/library.ts) (`jellytau-view-mode`,
-   `localStorage`). The Display setting is a second view onto it — **no new
-   state, no migration.**
-
-## Design
-
-### `AccountMenu` component (DR-075)
-
-One component used by both breakpoints. Contents in fixed order:
-
-```
-Signed in as <name>        ← identity block, not interactive
-<server host>
-────────────────────────
-Downloads
-Settings
-Display                    ← grid/list preference
-────────────────────────
-Sign out                   ← destructive, last, after a divider
-```
-
-- **Trigger is the username/avatar**, not a bare three-dot icon. On mobile where
-  horizontal space is tight, the avatar (or initial) alone is acceptable; the
-  name shows inside the open menu regardless.
-- **Same items, same order, both platforms.**
-- Preserve the existing dismissal behaviour: click-outside backdrop, `Escape`,
-  and focus return to the trigger on close.
-- Menu items are real links/buttons — keyboard reachable, correct roles,
-  `aria-expanded` on the trigger.
-
-"Display" may either navigate to the Settings Display section or expose the
-grid/list choice inline. Prefer navigating — it keeps one source of truth for
-preferences and avoids a nested control inside a dropdown.
-
-### Global chrome (DR-076)
-
-Make the header — and therefore the account menu — available on `/`, `/search`,
-and `/downloads`.
-
-The cleanest route is to lift the header out of the library layout into a shared
-component rendered by the root layout, with the library layout consuming the
-same component rather than defining its own. **Do not duplicate the markup into
-each route.**
-
-Constraints that must survive the change:
-
-- `/player/*` and `/login` stay chrome-free.
-- `/settings` already owns its layout; it needs no account menu (the user is
-  already there), but must not double up on chrome.
-- The root layout's flex/scroller structure is deliberate — the comments in
-  [layoutShell.ts](../../src/lib/utils/layoutShell.ts) and
-  [+layout.svelte](../../src/routes/+layout.svelte) explain why routes own their
-  own column. Preserve the scroll containment; a regression here reintroduces
-  the "last row hidden behind the nav" bug called out in those comments.
-- Mini-player and bottom-nav visibility rules (`showGlobalMiniPlayer`,
-  `showBottomNav`) must be unchanged.
-
-### Display section in Settings (DR-077)
-
-Add a Display section to [settings/+page.svelte](../../src/routes/settings/+page.svelte)
-with the grid/list control bound to the existing `viewMode` store via
-`library.setViewMode(...)`. The page-header toggle in `LibraryGrid` stays — both
-controls drive the same store, so they stay in sync for free.
-
-## Out of scope
-
-- Redesigning the Settings page or reorganising its existing sections.
-- Multi-server / account switching (UR-047) — the identity block displays the
-  active server but offers no switcher.
-- Changing the bottom nav's three destinations.
-
-## Acceptance criteria
-
-- [ ] Settings and Sign out are reachable from `/`, `/search`, and `/downloads`
-      without first navigating into Library.
-- [ ] Desktop and mobile show the same account menu items in the same order.
-- [ ] The username/avatar opens the menu; it is a real button with
-      `aria-expanded`.
-- [ ] Sign out is last, after a divider, and still logs out + resets library
-      state + redirects as it does today.
-- [ ] `/player/*` and `/login` remain chrome-free.
-- [ ] Settings has a Display section that changes grid/list, and the change is
-      immediately reflected by the library page-header toggle (same store).
-- [ ] No regression in scroll containment, mini-player visibility, or bottom-nav
-      visibility on any route.
-- [ ] `bun run check` and `bun run test` pass.
-
-## Testing
-
-- Extend the existing `layoutShell` tests: chrome-visibility for `/`, `/search`,
-  `/downloads` (now true) and `/player/*`, `/login` (still false).
-- `AccountMenu`: renders the documented items in order; trigger toggles
-  `aria-expanded`; `Escape` and backdrop click close it; Sign out invokes the
-  logout handler.
-- Display setting: writes through to the `viewMode` store and persists.
-
-New requirement-implementing code needs `TRACES:` comments — see
-[CLAUDE.md](../../CLAUDE.md). Suggested: `AccountMenu` → `UR-054 | DR-075`,
-shell/header changes → `UR-054 | DR-076`, Settings Display section →
-`UR-054, UR-029 | DR-077`.
-
-## Notes for the implementer
-
-- Read [ux-flows.md §1.2–1.4](../ux-flows.md) first — behavioural spec; this is
-  the implementation plan.
-- The layout shell is subtle and the existing comments record real bugs that
-  were fixed there. Read them before restructuring.
-- Another session may be active in this repo, including in
-  `src/routes/settings/+page.svelte`. Check `git diff` before "repairing"
-  unexpected changes, and expect to coordinate on that file.

docs/specs/audio-equalizer.md

@@ -1,171 +0,0 @@
-# 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:** —
-
-## 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).
-- 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/build-provenance.md

@@ -0,0 +1,256 @@
+# Spec: Build provenance (git describe + build profile)
+
+**Status:** Proposed — not started. `src-tauri/build.rs` still contains only
+`tauri_build::build()`, and nothing reports a version over IPC. Note that
+`scripts/set-version.sh` has since landed, which changes the "three hand-bumped
+files" premise below: versions are now stamped from one place.
+**Requirements:** ⚠️ the suggested id **DR-093 has since been allocated** to the
+traceability coverage gate — allocate a fresh id (DR-215 or later) on
+implementation. 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/desktop-native-video.md

@@ -0,0 +1,423 @@
+# Spec: Desktop native video — mpv renders the picture, everywhere
+
+**Status:** Proposed
+**Requirements:** UR-080 (new) → DR-231 … DR-237 (new); IR-033 (new)
+**UX spec:** n/a — nothing about the player's appearance changes. What changes is
+what is behind the controls.
+**Supersedes / revises:** consumes and closes
+[linux-native-video-spike.md](linux-native-video-spike.md), whose gates
+authorised exactly this spec and nothing more. Settles finding 2 of
+[playback-backend-unification.md](playback-backend-unification.md) on the
+desktop; finding 3 was already settled by DR-229. Absorbs the video half of what
+[windows-native-audio-backend.md](windows-native-audio-backend.md) leaves open.
+**Depends on:** backend-owned stream selection (DR-225 … DR-230), the branch
+below this one. mpv is a *consumer* of `StreamSelection`, never a second place to
+decide what to play.
+
+**Destination on completion:**
+[05-platform-backends.md](../architecture/05-platform-backends.md) — a "Native
+Video Compositing (Desktop)" section beside the existing Android one, which this
+mirrors; and [01-rust-backend.md](../architecture/01-rust-backend.md) — the
+device profile becomes renderer-dependent, beside the stream-selection section.
+**The spike is deleted in the same commit**, its three traps and its
+hardware-decode table folded in; they are the durable half.
+
+## Summary
+
+mpv decodes and draws video on **every desktop platform**, composited beneath the
+transparent webview, exactly as Android already does with ExoPlayer. The HTML5
+`<video>` path and hls.js are then **deleted**, not merely bypassed.
+
+The user-visible change is that most video stops being re-encoded by the server
+before it can be watched. The change for whoever maintains this is that video
+goes from three renderers to two.
+
+## Motivation
+
+### The transcode is a decoder constraint, not a rendering one
+
+Desktop video goes through an h264 HLS transcode because the picture is drawn by
+a WebKitGTK `<video>` element, and that element decodes little else. The device
+profile therefore claims `h264` alone. That is not a statement about the machine
+— the same machine runs mpv, which decodes essentially everything in the library
+— it is a statement about which widget is holding the frame.
+
+DR-228 made the cost measurable. Over 40 items negotiated against the development
+server:
+
+| Profile | Direct play |
+|---|---|
+| Desktop / WebKitGTK — `h264` only, 2ch | **7%** |
+| Android / ExoPlayer — `h264,hevc,vp8,vp9,av1,mpeg4` + `ac3,eac3`, 6ch | **85%** |
+
+The sampled library is ~80% hevc. **Those rows differ only by which component
+decodes.**
+
+Moving the picture to mpv is what lets the desktop row claim what the machine
+can actually do, and that — not the compositing — is the product.
+
+> **The 85% is a ceiling, not a shipped result.** It was measured with a profile
+> containing `ac3,eac3`. The Android device later used for verification reports
+> neither in its `MediaCodecList` — no Dolby licence, normal for a tablet — so
+> eac3 content, about a third of the sampled library, correctly transcodes there.
+> Realising any of this depends on DR-234, deriving the profile from the renderer
+> rather than from the platform, which is why that requirement is load-bearing
+> and not tidy-up.
+
+### One desktop video path, not two
+
+This is why the spec covers Windows rather than stopping at Linux.
+
+Today video has **three** renderers: ExoPlayer, the WebKitGTK `<video>` element,
+and (on Android, via the opt-out) that same element again. A Linux-only version
+of this work would make it four, permanently: mpv on Linux, HTML5 on Windows,
+ExoPlayer on Android, plus hls.js underneath the HTML5 one. Every seek strategy,
+every track switch, every quality change, every lifecycle bug would then have one
+more place to be got right — and the HTML5 path would survive indefinitely
+because *something* would still need it.
+
+Finishing the job removes that: **mpv on desktop, ExoPlayer on Android**, and
+`hls.js`, `html5Adapter.ts`, `videoLoaderFor` and the webview video element all
+go. The maintenance win is the reason Windows is in this spec and not in a
+follow-up that never gets written.
+
+### Three blockers are gone
+
+1. **Compositing works, including Wayland.** The spike ran all six gates; the
+   2024 "not possible on Wayland at all" claim is out of date when the render API
+   is used instead of foreign-window embedding.
+2. **There is no ABR to lose.** DR-229: the server's master playlist carries one
+   `EXT-X-STREAM-INF`. hls.js was demuxing, not adapting.
+3. **A direct-play path exists.** It did not when the spike was written. DR-228
+   built it; DR-230 proved the contract is player-agnostic.
+
+And on Windows specifically, `tauri-plugin-libmpv` lists Windows as its **fully
+tested** platform — the inverse of the Linux situation the spike had to
+disprove. The embedding difficulty was always WebKitGTK-specific.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|---|---|---|
+| **Which codecs this device can decode** | **Rust** | Domain: it is the input to Jellyfin's `PlaybackInfo` negotiation. It stops being a property of the *platform* and becomes a property of *the renderer in use* — see "The structural change". |
+| Which backend renders video | **Rust** | Rust already owns this (`use_html5_element` / `VideoBackend`). It stops being a `cfg!` constant and becomes a runtime fact. |
+| What stream to play (direct / remux / transcode, transport, ceiling) | **Rust — already decided** | DR-225. mpv consumes `StreamSelection`. Re-deriving any of it in a new backend would be the defect DR-225 exists to remove, restated. |
+| Creating the GL surface, reparenting the webview, owning the render context | **Rust (platform layer)** | Native window and GL-context lifetime. Not presentation, and not expressible above the IPC boundary at all. |
+| Render-context ↔ GL-context lifetime binding | **Rust** | A correctness invariant over native resources. DR-232. |
+| Frame pacing (update callback, `report_swap`) | **Rust** | Timing against the compositor; mpv's own contract. |
+| Hardware-decode selection | **Rust** | A capability question about the machine, answered from what mpv reports it actually selected. |
+| Z-order of controls over video, overlay chrome, letterbox colour | **Frontend / mpv** | Presentation. Controls already draw over a transparent webview on Android; mpv paints its own letterbox bars (better than the Android equivalent, which shipped DR-194 as a defect). |
+| Whether the surface is visible right now | **Frontend** | `nativeVideoActive` already exists and toggles `data-native-video`. Unchanged. |
+
+### The structural change
+
+Everything above is routine except one row, and it carries the whole benefit.
+
+`video_codecs` in `build_device_profile` is a **compile-time constant per
+platform**:
+
+```rust
+#[cfg(all(not(target_os = "android"), target_os = "linux"))]
+let (video_codecs, audio_codecs) = ("h264".to_string(), "aac,mp3,opus,…");
+```
+
+That is correct only while a build has exactly one video renderer. It must be
+derived from **which renderer will decode this stream**, which is runtime state.
+
+It looks like configuration and is not: it is the input that decides whether the
+server re-encodes, it changes when Jellyfin's API or our renderer changes, and
+getting it wrong fails *silently* — a claimed codec the renderer cannot decode is
+a black picture or silence, which is DR-148 and DR-228's audio override already.
+
+**Write this against "the active video renderer", never `cfg!(target_os)`.** It
+is the single piece that must not be Linux-shaped, because phase 2 reuses it
+unchanged.
+
+## Design
+
+### Backend and compositing (DR-231, IR-033)
+
+An `MpvVideoBackend` beside the existing `MpvBackend` (audio). The mpv side —
+render context, FBO, update callback, hwdec — is **shared**; only the surface
+differs per platform:
+
+| Platform | Surface | Status |
+|---|---|---|
+| Linux (X11 + Wayland) | `gdk_cairo_draw_from_gl()` in the default vbox's `draw` handler, over a `GdkGLContext` on its `GdkWindow`. No reparenting — see below | Render path proven by the spike; the *overlay* approach it used is rejected |
+| Windows | Native HWND child beneath a transparent WebView2 | Phase 2 |
+
+`vo=libmpv` plus `mpv_render_context_create` with `MPV_RENDER_PARAM_OPENGL_FBO`.
+Webview transparency via `with_transparent(true)` — no window-level transparency;
+the spike showed it is neither used nor needed.
+
+**G1's untested half failed, and the design changed because of it.**
+
+Reparenting Tauri's webview into a `GtkOverlay` attaches cleanly and then aborts
+the process on the first click. `tauri-runtime-wry` connects a
+button-press handler to the webview that walks a hard-coded path:
+
+```rust
+webview.parent()            // "This one should be GtkBox"
+       .parent()            // ...and this one the GtkWindow
+       .downcast::<gtk::Window>().unwrap()
+```
+
+An overlay makes that chain `webview → GtkOverlay → GtkBox`, the downcast fails,
+and the panic is non-unwinding so it kills the app. Nothing in configuration
+avoids it: on Linux `attach_resize_handler` is called **unconditionally** (the
+Windows equivalent is guarded by `is_decorated()`), and the decoration check that
+would make the handler inert runs *after* the unwrap.
+
+**So the webview is not moved at all.** mpv draws into the *default vbox's own
+`draw` handler* instead, via `gdk_cairo_draw_from_gl()` over a `GdkGLContext`
+created on that widget's `GdkWindow`. GTK3 draws a container before its children,
+so the webview composites on top for free — the same z-order the overlay was for,
+without touching the widget tree Tauri walks.
+
+That is strictly better than the overlay it replaces: no reparent, no extra
+widget, and the arrangement cannot be broken by a Tauri upgrade that assumes its
+own layout. It is also why "the surface attached successfully" is not the gate —
+a click is.
+
+Three traps from the spike, each of which cost a debugging cycle and each of
+which looks like a platform limitation and is not:
+
+1. **`LC_NUMERIC` must be reset *after* `gtk::init()`.** mpv refuses to start
+   under a non-C numeric locale. `mpv_backend.rs` already handles this but has no
+   GTK init in front of it; here `gtk::init()` applies the user's locale
+   afterwards and `mpv_create` returns null.
+2. **libepoxy exports GL entry points as *data* symbols.** There is no `glFoo`
+   function — there is `epoxy_glFoo`, a variable holding a lazily-resolving
+   pointer. `get_proc_address` must return the pointer **stored at** that symbol;
+   returning the symbol's own address makes mpv jump into non-executable data and
+   take SIGSEGV on the first GL call. The `epoxy` crate does this correctly but is
+   unusable — its `gl_generator` dependency pulls a yanked `xml-rs`.
+3. **Frame pacing is not optional and its symptom misleads.** See DR-233.
+
+### Render-context lifetime (DR-232) — the crash defence
+
+The spike's one unexplained SIGSEGV landed in a *decoder* thread with no Tauri,
+GTK or GL frame in the stack, and three plausible causes failed to reproduce it
+across ~13 minutes of targeted stress.
+
+What is **not** unexplained is that the spike had no defence: it never calls
+`mpv_render_context_free` and never tears down on `unrealize`, so nothing stopped
+the GL context being recreated beneath the render context. That is DR-184 on
+Android restated — a surface outliving its player.
+
+Built as a requirement in its own right, not as a fix for a crash we cannot yet
+reproduce:
+
+- Render context created on `realize`, freed on `unrealize`, same thread, before
+  the GL context goes away.
+- The update callback is unregistered **before** the context is freed, so a
+  callback cannot land on a freed context.
+- Playback teardown and surface teardown are ordered, not racing.
+
+If the crash recurs after this, it is a different bug and the likeliest cause is
+out of the search space. If it does not, we needed this anyway.
+
+### Frame pacing (DR-233)
+
+Register `mpv_render_context_set_update_callback`; redraw only when it reports a
+frame ready; call `mpv_render_context_report_swap` after each render.
+
+Recorded because the failure mode is a trap: driving `queue_render()` off the
+frame clock every tick without reporting the swap leaves mpv nothing to time
+against. It looks fine in a window and **judders at fullscreen**, which reads as
+a compositing or GPU limit and is neither.
+
+### Renderer-dependent device profile (DR-234)
+
+`build_device_profile` takes the active video renderer and derives the codec
+lists from it:
+
+| Renderer | Video codecs | Audio (video direct play) | Channels |
+|---|---|---|---|
+| mpv (desktop native) | `h264,hevc,vp8,vp9,av1,mpeg4` | platform list incl. `ac3,eac3` where the sink can voice it | from the audio route |
+| WebKitGTK `<video>` | `h264` | webview-decodable set only | 2 |
+| ExoPlayer (Android) | unchanged | unchanged | unchanged |
+
+The existing `video_audio_codecs()` narrowing exists because *the webview decodes
+a narrower audio set than the platform*. With mpv decoding, that no longer
+applies to the video path — but the multichannel bound still does, since a 5.1
+track direct-played into a 2-channel sink is silence or inaudible dialogue. Both
+constraints stay, sourced from the renderer rather than assumed.
+
+**This is what converts the 7% figure upward** (toward, not necessarily to, the 85% ceiling — see the caveat above), and it is also the change most able to break
+playback silently — so it lands after compositing is proven, covered by the
+DR-228 override tests.
+
+### Deleting the webview video path (DR-235)
+
+`get_player_status` stops reporting `use_html5_element: true` on desktop;
+`supports_native_video` becomes true there.
+
+Deletion is staged, because a path cannot be removed while a shipped platform
+still needs it:
+
+| Phase | Linux | Windows | HTML5 video path |
+|---|---|---|---|
+| 1 | mpv | HTML5 | alive — Windows needs it |
+| 2 | mpv | mpv | alive but unreached |
+| 3 | mpv | mpv | **deleted**, with hls.js |
+
+Phase 3 is a real phase with its own acceptance criterion, not a "later". The
+whole maintenance argument for including Windows collapses if the fork survives.
+
+Android keeps ExoPlayer and keeps the webview as its documented opt-out; the
+`<audio>` element and the background-audio handoff are untouched throughout.
+
+**What happens when mpv fails to initialise.** With no HTML5 path there is no
+silent fallback, and inventing one resurrects what we deleted. The
+graceful-backend-init principle applies as written: fall back to the no-op
+backend, emit `backend-init-failed`, and surface a real error rather than a black
+rectangle. An honest failure beats a hidden downgrade to the transcode we are
+trying to stop paying for.
+
+### Hardware decode (DR-236)
+
+The spike established the load-bearing fact: **hardware decode works through the
+render API** (`hwdec-current` reported `nvdec-copy` on the discrete GPU), so the
+direct-play prize is not traded for software decoding.
+
+Policy is decided from what mpv reports it *selected*, never from what it was
+asked for:
+
+- Prefer zero-copy VA-API on the integrated GPU where the driver is present.
+- `auto` reached for the discrete GPU in **copy-back** mode on a hybrid
+  Intel+NVIDIA laptop — the least efficient hardware path — so `auto` is a
+  fallback, not the default.
+- `vaapi` silently fell back to software on the spike box because `vainfo` was
+  absent. A missing driver must be detected and logged, not mistaken for a
+  compositing limit.
+- Log `hwdec-current` at start-up; knowing what was actually chosen is the whole
+  diagnostic value.
+
+### Windows: what phase 2 actually costs (DR-237)
+
+Not hidden, because it is the part most likely to be underestimated:
+
+- **The surface is different code.** WebView2 in an HWND, not GTK. A transparent
+  WebView2 over a native child window is a solved arrangement, but DR-231's
+  Linux surface does not transfer. Everything else does.
+- **libmpv is currently a Linux-only dependency**, and Windows is
+  **cross-compiled from Linux** via `x86_64-pc-windows-msvc` + `cargo-xwin`. Phase
+  2 must source a Windows libmpv (DLL + import library) into that cross-build and
+  ship the DLL in the NSIS bundle.
+- **LGPL obligations follow the DLL.** DR-216 already records them for Linux:
+  keep the linkage dynamic, ship libmpv's licence text with any bundle carrying
+  it. The Windows bundle inherits both.
+- **`bun run test:rust` and CI must still build.** Per the CI rule, any tool this
+  needs goes into the builder image and is pushed — never installed at job time.
+
+Windows also gains a native *audio* decoder as a side effect, which is what
+[windows-native-audio-backend.md](windows-native-audio-backend.md) wants and
+cannot currently have. If that spec lands first, phase 2 inherits its build work
+and shrinks to the surface.
+
+## Out of scope
+
+- **Android.** Unchanged in every respect.
+- **macOS.** Not a shipped target. If it becomes one it joins phase 2's shape.
+- **Audio backends.** mpv already plays audio on Linux; this adds a video
+  renderer beside it. Windows audio is its own spec.
+- **HDR, tone mapping, multi-window.** Not exercised by the spike at all.
+- **Re-deciding what stream to play.** DR-225 owns that. If this spec finds
+  itself choosing a URL, something has gone wrong.
+
+## Acceptance criteria
+
+**Phase 1 — Linux**
+
+- [ ] Tauri's own webview reparents into the overlay (the untested half of G1),
+      on X11 **and** Wayland.
+- [ ] Video plays, seeks and switches audio track in mpv, with the Svelte
+      controls composited over it and alpha blending intact.
+- [ ] The render context is freed on `unrealize` and the update callback
+      unregistered before the free; a test demonstrates the ordering.
+- [ ] A direct-play negotiation returns `DirectPlay` for an hevc source that
+      today returns `Transcode`, and it plays.
+- [ ] Direct-play rate over the same 40-item sample rises from 7% toward the
+      Android figure. **Record the number.**
+- [ ] mpv init failure emits `backend-init-failed` and surfaces an error rather
+      than falling back to a transcode.
+- [ ] `hwdec-current` is logged and is not copy-back where zero-copy is available.
+- [ ] A soak covering seek, track switch and fullscreen runs clean for an agreed
+      duration. **The spike's SIGSEGV is why this is a criterion.**
+
+**Phase 2 — Windows**
+
+- [ ] libmpv links in the `cargo-xwin` cross-build; the DLL and its licence ship
+      in the NSIS bundle; any new tool lives in the builder image, not in a CI step.
+- [ ] Video plays composited under a transparent WebView2.
+- [ ] The device profile, lifetime and hwdec code are **reused, not
+      reimplemented** — a reviewer confirms no `cfg!(target_os = "linux")` guards
+      them.
+
+**Phase 3 — deletion**
+
+- [ ] `use_html5_element` is false on every desktop platform.
+- [ ] `hls.js` is gone from `package.json`; `html5Adapter.ts`, `videoLoaderFor`
+      and the `<video>` element are deleted; Android's opt-out and the
+      background-audio `<audio>` path still work.
+
+**Throughout**
+
+- [ ] `bun run check`, `bun run test`, `bun run format:check`, `bun run lint` pass.
+- [ ] `cargo fmt` clean, `cargo clippy -D warnings` clean, `bun run test:rust` passes.
+- [ ] `bun run check:boundary` passes, and a reviewer confirms no stream decision
+      was reconstructed in the new backend.
+- [ ] `bindings.ts` regenerated from Rust.
+- [ ] `bun run traces:validate` passes; coverage stays ≥ the CI ratchet.
+- [ ] The spike and this spec are folded into
+      [05-platform-backends.md](../architecture/05-platform-backends.md) and both
+      deleted in the same commit.
+
+## Testing
+
+- **Rust, pure:** the device profile per renderer — mpv claims hevc, the webview
+  does not, the multichannel bound survives both. The DR-234 table as a
+  table-driven test.
+- **Rust, pure:** `PlaybackInfo` fixtures that transcode under the webview
+  profile and direct-play under the mpv profile — the direct-play conversion as a unit
+  test, not only as a measurement.
+- **Rust:** teardown ordering — callback unregistered before context freed, freed
+  before GL context destroyed. Structure it so the ordering is assertable without
+  a live GL context.
+- **Frontend:** no desktop path selects an HTML5 video adapter. After phase 3,
+  the adapter does not exist and the test goes with it.
+- **Manual / soak:** the criterion above. The spike's automated fullscreen and
+  resize soaks are reusable and already written.
+
+## TRACES
+
+| Piece | Tag |
+|---|---|
+| mpv video backend + compositing | `UR-080 \| DR-231, IR-033` |
+| Render-context lifetime binding | `UR-080 \| DR-232` |
+| Frame pacing | `UR-080 \| DR-233` |
+| Renderer-dependent device profile | `UR-080, UR-070 \| DR-234` |
+| Webview video path removed | `UR-080 \| DR-235` |
+| Hardware-decode policy | `UR-080 \| DR-236` |
+| Windows surface + cross-build | `UR-080 \| DR-237` |
+
+## Notes for the implementer
+
+- **Read the spike before writing a line.** Its three traps and its
+  hardware-decode table are the most valuable things in this directory, and each
+  cost a debugging cycle to find.
+- **mpv consumes `StreamSelection`; it does not decide.** The transport is on the
+  queue item (DR-230). If you are parsing a URL, stop.
+- **Guard nothing on `cfg!(target_os = "linux")` that phase 2 will need.** That is
+  the one avoidable mistake here.
+- The Android backend is the reference for the *shape* of this — transparent
+  webview over a native surface at index 0. Read `05-platform-backends.md`'s
+  Android section for what shipped and what its defects were (DR-184 surface
+  lifetime, DR-194 letterbox).
+- Do not call sync/blocking APIs from mpv event callbacks that can re-enter the
+  player or hold a lock. The existing deadlock gotchas apply.
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes.
+- This branch is stacked on backend-owned stream selection. Rebase when that
+  merges rather than merging master into it.

docs/specs/diagnostics-and-logging.md

@@ -0,0 +1,192 @@
+# Spec: Diagnostics and persistent logging
+
+**Status:** Proposed
+**Requirements:** UR-078 → DR-218; tests UT-209
+**UX spec:** n/a (one Settings section; no new flow)
+**Destination on completion:** [09-security.md](../architecture/09-security.md)
+for the redaction rules, and a new "Logging and diagnostics" section in
+[01-rust-backend.md](../architecture/01-rust-backend.md) for the capture path.
+
+## Summary
+
+JellyTau records what it does, keeps it in a size-capped file on disk, survives a
+crash, and can hand the whole thing to the user as one file to attach to a bug
+report. Credentials never reach that file.
+
+## Motivation
+
+Today the app forgets everything the moment it exits.
+
+The Rust half logs through `env_logger` to **stdout only**. A user who launched
+from a desktop icon has no stdout. On Android it is worse than useless:
+`env_logger` writes to stdout, which is not logcat, so **the Rust backend's
+output is invisible on the platform where most of the hard bugs have been** — the
+autoplay deadlock, the truncated-stream restart, the background-audio stall. The
+frontend has a proper leveled facade (`logger.ts`, DR-204) but it only reaches
+the webview console, which nobody can read on a phone.
+
+The practical consequence is visible in this project's history: several bugs took
+multiple rounds of "can you reproduce it under `adb logcat`" before anyone could
+even see what happened. A user reporting "the episode randomly restarted" is
+reporting the symptom of a race whose evidence was discarded microseconds later.
+
+There is also no crash record at all. If the app panics, the user sees it vanish
+and we learn nothing.
+
+## Layer assignment
+
+| Logic / responsibility | Layer | Why it belongs there |
+|---|---|---|
+| What is captured, at what level, and where it is written | Rust | Retention and capture policy is backend behaviour; it must work identically whether the UI is open, backgrounded, or gone |
+| Log rotation and the size cap | Rust | Storage management, same class as the download and image caches |
+| **Redaction of credentials** | Rust | Security-critical, and the values (tokens, `api_key`, keyring payloads) are domain vocabulary owned by the auth layer. A frontend that redacted its own messages would still not cover anything Rust wrote |
+| Panic capture and persistence | Rust | Only Rust can install a panic hook |
+| Assembling the export (archive + environment summary) | Rust | Touches the filesystem and the app's own paths; also the last point at which redaction can be enforced over everything |
+| Which log level is active | Rust owns the *stored* setting and applies it; the frontend renders the picker | Same split as every other setting: the value is state the backend acts on, the control is presentation |
+| Showing the export path / opening the folder | Frontend | Pure presentation |
+| Formatting a log line for the webview console | Frontend | `logger.ts` already owns this; unchanged |
+
+Borderline: **the frontend forwarding its own messages into the Rust sink.**
+Arguably presentation "sending data down". Placed as: the frontend calls a
+plugin, and the *decision of what to persist and how to redact it* stays in Rust
+— which is the tie-breaker, because a bug report containing a token would be a
+security defect regardless of which half wrote the line.
+
+## Design
+
+### Capture
+
+Replace the `env_logger` init in `lib.rs` with `tauri-plugin-log`, which is the
+official plugin and already does the three things we would otherwise hand-roll
+(CLAUDE.md: prefer official plugins before writing native code):
+
+| Target | Purpose |
+|---|---|
+| `Stdout` | unchanged behaviour for `bun run tauri dev` |
+| `LogDir { file_name: "jellytau" }` | the persistent, rotating file |
+| `Webview` (dev only) | Rust lines visible in the webview console while developing |
+
+On Android the plugin routes to **logcat**, which is the single largest
+improvement here and needs no code of ours.
+
+Rotation: `RotationStrategy::KeepAll` is wrong for a phone. Use a size cap
+(5 MB) with one retained previous file, so a long session cannot fill a device
+and yesterday's evidence still exists.
+
+Level: default `Info`, `RUST_LOG` still honoured, and a stored user preference
+that survives restart (a user reproducing a bug needs debug logging *across* the
+restart that reproduces it).
+
+### Redaction
+
+A pure function in a new `src-tauri/src/utils/diagnostics.rs`:
+
+```rust
+pub fn redact(line: &str) -> String
+```
+
+It replaces the value in each of these with `[REDACTED]`, case-insensitively:
+
+- `api_key=…` and `ApiKey=…` in URLs and query strings
+- `X-Emby-Token: …`, `X-MediaBrowser-Token: …`, `Authorization: …` headers
+- `"AccessToken":"…"` in JSON bodies
+- `MediaBrowser Token="…"` in the Emby auth header form
+
+What it deliberately does **not** remove: the server host, item ids, and
+filenames. Those are what make a log useful, they are not secrets, and stripping
+them would produce a diagnostic bundle nobody can diagnose anything from.
+
+Applied at two points: on every line the export copies, and — because the export
+is not the only way a file leaves a device — inside the log formatter itself, so
+the token never reaches disk in the first place. The export-time pass exists to
+cover files written before an upgrade.
+
+### Export
+
+```rust
+#[tauri::command]
+pub async fn diagnostics_export(app: AppHandle) -> Result<DiagnosticsBundle, String>
+```
+
+Writes a single `.zip` and returns where it went:
+
+```rust
+#[derive(Serialize, Type)]
+#[serde(rename_all = "camelCase")]
+pub struct DiagnosticsBundle {
+    pub path: String,
+    pub size_bytes: u64,
+    pub file_count: usize,
+}
+```
+
+Contents: the current and previous log files (redacted), plus `environment.txt`
+— app version, OS and arch, whether the build is debug, the active log level, and
+the *scheme and host* of the configured server. No token, no username, no path
+inside the user's home beyond the app's own directories.
+
+### Frontend
+
+`logger.ts` keeps its `console.*` pass-through untouched — live object references
+in devtools are a stated design goal of DR-204 — and *additionally* forwards a
+stringified copy at `info` and above to the plugin, so one timeline contains both
+halves of the app. Forwarding is fire-and-forget and never throws into a caller:
+a logging failure must not become an application failure.
+
+A `Diagnostics` section in Settings shows the log location, a level picker, and
+an **Export diagnostics** button that reports the resulting path and, on desktop,
+offers to reveal it.
+
+## Out of scope
+
+- **An Android share sheet.** Export writes to the app's files directory and
+  reports the path; wiring a native `ACTION_SEND` intent is a Kotlin change that
+  belongs with the other native work, not here.
+- **Uploading anywhere.** Nothing is transmitted. The user attaches the file
+  themselves, which is also what keeps this from becoming telemetry.
+- **Frontend `debug` forwarding.** Only `info`+ crosses the IPC boundary; per-tick
+  player debug would be thousands of calls a minute.
+
+## Acceptance criteria
+
+- [ ] Rust logs reach a rotating file on Linux and **logcat** on Android.
+- [ ] A panic is recorded and is present in the next export.
+- [ ] Frontend `info`/`warn`/`error` appear in the same file as Rust's lines.
+- [ ] An export containing a request URL with `api_key=` shows `[REDACTED]`, and
+      a test greps the produced bundle for the token to prove it.
+- [ ] The log file cannot exceed the cap.
+- [ ] `bun run check`, `bun run test`, `cargo fmt`, `cargo clippy -D warnings`,
+      `bun run test:rust`, `bun run check:boundary` all pass.
+- [ ] `bindings.ts` regenerated (new command and struct).
+- [ ] New code carries `TRACES:` comments.
+
+## Testing
+
+**Rust** (`cargo test`): `redact` over each credential shape, including one
+already-redacted line (idempotent) and a line containing no secret (unchanged);
+that the environment summary contains a host but no token; that rotation respects
+the cap.
+
+**Frontend** (`vitest`): that the forwarder is called for `info`+ and not for
+`debug`; that a rejected forward does not propagate to the caller.
+
+## TRACES
+
+| Piece | Tag |
+|---|---|
+| `utils/diagnostics.rs` | `UR-078 \| DR-218` |
+| `commands/diagnostics.rs` | `UR-078 \| DR-218` |
+| logging init in `lib.rs` | `UR-078 \| DR-218` |
+| `logger.ts` forwarding | `UR-078 \| DR-204, DR-218` |
+| Settings section | `UR-078 \| DR-218` |
+| tests | `\| DR-218 \| UT-209` |
+
+## Notes for the implementer
+
+- A parallel Claude session may be active — `git diff` before "repairing"
+  anything unexpected.
+- `utils/lock.rs` already sets and restores a panic hook in its tests. The
+  diagnostics hook must chain to the previous hook rather than replace it, or
+  those tests start reporting panics they deliberately suppress.
+- Do not call the exporter from an event callback that can re-enter the player:
+  it does blocking file I/O (see the deadlock note in CLAUDE.md).

docs/specs/downloads-as-offline-library.md

@@ -1,166 +0,0 @@
-# Spec: Downloads as a browsable offline library
-
-**Status:** Draft — ready to implement
-**Scope:** Frontend-heavy; one new repository-client browse path. Minimal Rust.
-**Requirements:** UR-055 → DR-081, DR-082, DR-083, DR-084; UR-056 → DR-085
-(see [requirements.md](../requirements.md)).
-**UX spec:** [ux-flows.md §7.2–7.7](../ux-flows.md).
-
-## Summary
-
-Replace the flat Active/Completed download list with two views under
-`/downloads`:
-
-1. **Downloaded** (default) — the library, filtered to what's on the device,
-   using the *same* browse screens as online (grids, cards, detail pages).
-2. **Transfers** — the existing progress-row list, demoted to a secondary tab,
-   showing only in-flight transfers.
-
-Plus per-item disk usage (UR-056) shown in familiar units on cards, detail
-pages, a device total, and the remove confirmation.
-
-## Motivation
-
-A user who downloaded three seasons and two albums sees ~70 individual transfer
-rows today, with no grouping and no reuse of the library UI. "What do I have
-offline" and "what is downloading" are different questions crammed into one flat
-list. Browsing offline should feel exactly like browsing online.
-
-## Background: verified current state
-
-1. **The offline repository is already a browsable tree.**
-   [offline.rs](../../src-tauri/src/repository/offline.rs) — `get_items` returns
-   downloaded items **plus** containers (MusicAlbum, Series, Season) that have at
-   least one downloaded child. `get_libraries`, `get_item`, and `search` all
-   filter to downloaded content via CTEs. This is the data source for
-   Downloaded; **do not build a new query layer.**
-
-2. **The client cannot reach it independently.**
-   [repository-client.ts](../../src/lib/api/repository-client.ts) `getItems` →
-   `repositoryGetItems` always goes through the **hybrid** repository
-   ([hybrid.rs](../../src-tauri/src/repository/hybrid.rs)), which merges cache and
-   server. There is no "offline only" browse path exposed. This is the one real
-   backend gap (DR-082).
-
-3. **Downloads page is a flat two-tab list.**
-   [downloads/+page.svelte](../../src/routes/downloads/+page.svelte) — Active /
-   Completed tabs, one `DownloadItem` row per transfer, no browsing.
-
-4. **Library browse components are reusable as-is.** `LibraryGrid`, `MediaCard`,
-   the `/library/[id]` detail page (§5A/§5B) render whatever items they are
-   given. Downloaded browse is those components with an offline-scoped source.
-
-5. **A related fallthrough bug is already tracked** (DR-080, another session):
-   `HybridRepository::get_items` treats an empty offline result as a cache miss
-   and falls through to the server. The offline-only browse path (DR-082) must
-   **not** share that behaviour — an empty result there is authoritative "nothing
-   downloaded here."
-
-6. **Concurrency, the 3-download cap, and the auto-pump are backend concerns.**
-   Do not surface them as manual controls; do not loop `startDownload` from the
-   frontend (see [CLAUDE.md](../../CLAUDE.md) gotchas).
-
-## Design
-
-### View split (DR-081)
-
-`/downloads` renders a **Downloaded** / **Transfers** switch. Downloaded is the
-default. Transfers shows a count/badge only while transfers are active.
-Initiating downloads stays on item/album/series detail pages (§7.1) — this page
-does not start downloads.
-
-### Offline-scoped browse source (DR-082, DR-083)
-
-Add an explicit offline-only browse path so Downloaded never merges server
-results and never depends on reachability. Two viable shapes — pick per the
-codebase, do not do both:
-
-- **(a)** A dedicated command (e.g. `repository_get_downloaded_items` /
-  `_libraries`) that calls the offline repository directly, with a matching
-  client method; or
-- **(b)** An explicit `offlineOnly`/scope flag on the existing get-items path
-  that bypasses the hybrid merge and the empty→fallthrough behaviour.
-
-Either way: an empty result is authoritative (do **not** reuse the DR-080
-fallthrough), and the path is available while the server is reachable (a user
-online still wants to browse their downloads).
-
-Downloaded then reuses `LibraryGrid` / `MediaCard` / the detail page against this
-source. Omit libraries and containers with no downloaded content. Badge
-partially- vs fully-downloaded containers. Play uses the local file; remove is
-available at item / album / season / series level and removes a container from
-the browse when its last downloaded child goes.
-
-### Transfers view (DR-084)
-
-The existing list, filtered to in-flight rows only: downloading (with progress),
-queued, paused, failed, waiting-for-WiFi (the DR-074 state from the other
-session). Controls: Pause / Resume / Cancel / Retry. Completed transfers leave
-this view — they appear in Downloaded. Empty state points at the library.
-
-### Disk usage (DR-085, UR-056)
-
-- **Source the bytes from the download manager** — it writes the files and can
-  stat them. Aggregate to album/season/series subtotals and a device total.
-  This is display + aggregation, **not** new tracking.
-- **Format once, consistently.** One shared formatter, human units, 2–3
-  significant figures (`1.2 GB`, `340 MB`). Binary vs decimal — pick one and use
-  it everywhere.
-- **Surface it in familiar places:** a secondary size label on the card and
-  detail page; a device total at the top of Downloaded (`3.4 GB · 12 items`)
-  that reconciles with the listed sum; a reclaim figure in the remove
-  confirmation ("frees 1.2 GB"). No separate "storage report" screen.
-- Sort/filter by size is a nice-to-have, not required for v1.
-
-## Out of scope
-
-- Changing download initiation, the 3-concurrent cap, or the auto-pump.
-- The catalog-browse / show-server-catalog toggle (UR-052, another session) —
-  that governs the *online offline-fallback* library; this is the dedicated
-  Downloads surface. They should be consistent but are separate work.
-- Fixing the DR-080 hybrid fallthrough bug (owned elsewhere) — just don't depend
-  on that behaviour here.
-
-## Acceptance criteria
-
-- [ ] `/downloads` opens on Downloaded and can switch to Transfers.
-- [ ] Downloaded lists only libraries/containers with downloaded content, using
-      the same grids/cards/detail pages as online browsing.
-- [ ] Browsing Downloaded never shows non-downloaded server items, online or off.
-- [ ] An empty Downloaded result reads as "nothing downloaded," never falls
-      through to the server.
-- [ ] Play from Downloaded plays the local file.
-- [ ] Remove works at item/album/season/series level and updates the browse.
-- [ ] Transfers shows only in-flight rows with working controls; finished
-      transfers move to Downloaded.
-- [ ] Each downloaded item/container shows its on-disk size; a device total is
-      shown and reconciles with the sum; remove states the reclaim amount.
-- [ ] `bun run check`, `bun run test`, and (if Rust touched) `cargo test` +
-      `cargo clippy` pass.
-
-## Testing
-
-- Repository client: the offline-only browse path returns downloaded content and
-  its containers, and an empty result does **not** trigger server fallthrough.
-- Downloaded view: libraries/containers with no downloads are omitted;
-  partial/full container badging.
-- Transfers: only in-flight statuses render; a completed transfer disappears.
-- Size formatter: rounding and unit thresholds; subtotal aggregation; device
-  total reconciles with listed items.
-- If a Rust command is added, add the tauri IPC param-naming coverage per
-  [CLAUDE.md](../../CLAUDE.md) (camelCase rule).
-
-New requirement-implementing code needs `TRACES:` comments. Suggested tags:
-view split `UR-055 | DR-081`; offline browse path `UR-055 | DR-082, DR-083`;
-Transfers `UR-055 | DR-084`; size display `UR-056 | DR-085`.
-
-## Notes for the implementer
-
-- Read [ux-flows.md §7.2–7.7](../ux-flows.md) first — behavioural spec; this is
-  the implementation plan.
-- The offline repository already does the hard part. The main work is a clean
-  offline-only client path and reusing the library components — resist
-  rebuilding browse UI.
-- Another session is active in downloads/offline/connectivity code (DR-074,
-  DR-078–080). Coordinate on [downloads/+page.svelte](../../src/routes/downloads/+page.svelte)
-  and the repository layer; check `git diff` before repairing unexpected changes.

docs/specs/libmpv2-migration.md

@@ -0,0 +1,171 @@
+# Spec: Migrate to libmpv2 and declare the project licence
+
+**Status:** Partially implemented — the `LICENSE` file has landed (part 2). The
+`libmpv` → `libmpv2` swap (part 1) is **not** done: `src-tauri/Cargo.toml` still
+pins the abandoned crate to a git branch.
+**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/linux-native-video-spike.md

@@ -0,0 +1,503 @@
+# Spec: Linux native video — bounded compositing spike
+
+**Status:** **Run 2026-08-21 — compositing works; G5 carries an open crash.**
+The compositing claim it set out to test is falsified on Linux. See "Result".
+This file stays open until the implementation spec exists. **ABR is resolved** —
+the playlist carries one `EXT-X-STREAM-INF`, so finding 3 is false and there is
+no adaptation for mpv to lose. The remaining blocker is the unexplained SIGSEGV
+under G5, which is a lifetime problem, not a compositing one.
+**Requirements:** none allocated. This spike produces a decision record, not
+product code — same shape as
+[playback-backend-unification.md](playback-backend-unification.md), which is
+Accepted with no requirement ids of its own. Ids are allocated by the
+*implementation* spec that follows a green result.
+**UX spec:** n/a
+**Supersedes / revises:** re-opens finding 2 of
+[playback-backend-unification.md](playback-backend-unification.md) on Linux only.
+Its findings 3, 4, 5 and 6 stand unchallenged and are **not** in scope here.
+
+**Destination on completion:**
+[05-platform-backends.md](../architecture/05-platform-backends.md) — a "Native
+Video Compositing (Linux)" section alongside the existing Android one. The
+durable half is the mechanism and the two traps below; the gates and phases are
+disposable.
+
+## Summary
+
+Test one falsifiable claim: *a native video surface cannot be composited with a
+Tauri webview on Linux.* The claim is load-bearing — it is why Linux video goes
+through an h264 HLS transcode into a WebKitGTK `<video>` element instead of
+decoding directly in the mpv instance we already run. The spike renders one mpv
+frame beneath the webview, on both X11 and Wayland, and stops. It ships no
+product code and flips no defaults.
+
+A green result does **not** authorise native video on Linux; it authorises
+writing the spec that would.
+
+## Motivation
+
+[playback-backend-unification.md](playback-backend-unification.md) finding 2
+concluded that native video cannot be composited with a Tauri webview, on
+evidence from `tauri-plugin-libmpv`'s platform table, wry#284, tauri#6343, and a
+Tauri maintainer's 2024 statement that a GTK widget as a child X11 window is
+"a bit hacky and it is not possible on Wayland at all."
+
+Two things have changed since that was written, and one thing was never tested.
+
+**1. The general claim has already been falsified on one platform — by us.**
+Android now renders ExoPlayer video on a TextureView at index 0 *behind a
+transparent Tauri WebView*, with the Svelte controls drawn over it, on by
+default. See
+[05-platform-backends.md](../architecture/05-platform-backends.md#native-video-compositing-android).
+That is exactly the composition finding 2 said was impossible, shipped. What
+survives of the finding is a narrower, WebKitGTK-specific claim — which is worth
+testing on its own terms rather than inheriting.
+
+**2. A Tauri app now ships Linux native mpv as an active platform.**
+[MaxVideoPlayer](https://github.com/MaxMB15/MaxVideoPlayer) (354 commits) embeds
+libmpv via **EGL + X11 child window / Wayland subsurface**, with Linux and macOS
+active and Windows only planned — the inverse of the plugin matrix finding 2
+sampled. Its existence does not prove our case works, but it does mean the
+Wayland half of the maintainer quote is out of date.
+
+**3. The render API was never tested.** Every source in finding 2 describes
+*foreign-window embedding*: `--wid`, child windows, a second toplevel
+position-synced to a `getBoundingClientRect()` div. That is a different mechanism
+from mpv's render API, where **we** own the GL context and mpv draws into an FBO
+we hand it (`mpv_render_context_create` / `mpv_render_context_render`, with an
+upstream [GTK example](https://github.com/mpv-player/mpv-examples/pull/44/files)).
+Tauri v2 exposes `WebviewWindow::gtk_window()` and `default_vbox()`, so the
+target is a widget inside Tauri's own GTK tree — not a foreign window, not a
+second toplevel, and therefore not the thing that was found broken.
+
+The prize is direct play: no h264 transcode, hardware decode, libass subtitles,
+and no server CPU burned on every Linux play.
+
+## The blocker a green spike does not clear
+
+🔴 **Read this before treating a green result as a green light.**
+
+Finding 3 of the unification spec stands: **mpv has no adaptive bitrate.** It
+delegates HLS to FFmpeg's demuxer, which picks one variant at open and never
+adapts. The webview path has real ABR via hls.js. Compositing is necessary for
+native video on Linux; it is not sufficient.
+
+There is a plausible answer, and this spike exists partly to make it testable:
+**ABR only matters on the transcode path.** A direct-played file has no variant
+ladder to adapt between — the adaptation the server offers *is* the transcode.
+So "mpv when the stream is direct-play, HTML5 + hls.js when the server
+transcodes" would sidestep finding 3 rather than fight it, and it maps onto a
+decision Rust already makes when it builds the stream URL.
+
+That is a **hypothesis, not a conclusion.** It is out of scope here. Record it in
+the spike's decision note so the follow-up spec starts from it.
+
+## Layer assignment
+
+The spike introduces no product logic. The table below is the assignment the
+*follow-up* would inherit, written now so a green result cannot drift into
+frontend decisions during implementation.
+
+| Logic / responsibility | Layer | Why it belongs there |
+|------------------------|-------|----------------------|
+| Which backend renders video on this platform (`use_html5_element`, `supports_native_video`) | Rust | Already there — `get_player_status` in `commands/player/mod.rs` computes it from a `cfg!`. The spike would widen that `cfg!`, not relocate the decision. The frontend already consumes it via `createAdapter`. |
+| Whether *this stream* is direct-play or transcoded, and therefore whether mpv or hls.js renders it | Rust | Domain. It depends on Jellyfin's `PlaybackInfo` response, container/codec support, and the bitrate cap — all of which change when Jellyfin's API or our quality ladder changes. The frontend must never re-derive it from a URL shape. |
+| Creating, sizing, and destroying the GL surface; the mpv render context | Rust | Owns the backend and the GTK window handle. There is no presentation decision in it. |
+| Where controls, subtitles, and the mini-player sit above the video, and the letterbox/poster treatment | Frontend | Pure presentation; changes only if the UI is redesigned. Precisely the split the Android path already uses. |
+| Reserving the video rectangle in layout and marking the shell transparent | Frontend | Presentation. `nativeVideo.ts` + the `[data-native-video="active"]` rule in `app.css` already do this for Android and are platform-agnostic. |
+
+Borderline row, stated with its tie-breaker: *"is the surface currently
+attached?"* reads like view state, but the Android work found that a surface left
+in the hierarchy outlives its player (DR-184). Attachment is backend lifecycle →
+**Rust**, with the frontend told about it, not asked.
+
+## Design
+
+A throwaway branch. No merge to `master` except the decision note.
+
+### What gets built
+
+One `#[cfg(target_os = "linux")]` experiment behind a feature flag, in a scratch
+binary or an ignored test — **not** in `MpvBackend`'s constructor path:
+
+1. From `app.get_webview_window(...)`, take `gtk_window()` and `default_vbox()`.
+2. Reparent the webview into a `gtk::Overlay`: `GLArea` as the main child, the
+   webview as the overlay child.
+3. Set the webview background to fully transparent (wry does this when
+   `"transparent": true`; verify it reaches `webkit_web_view_set_background_color`).
+4. In the `GLArea`'s `render` signal, drive
+   `mpv_render_context_render` with `MPV_RENDER_PARAM_OPENGL_FBO` pointing at the
+   FBO GTK bound for us.
+5. Play one local file. Draw an opaque HTML element over the video area.
+
+`video = no` and `audio-display = no` are set in
+[mpv_backend.rs:135-141](../../src-tauri/src/player/mpv_backend.rs#L135-L141);
+the spike overrides them on its own `Mpv` handle rather than editing that path.
+
+### Bindings
+
+The current pin is `libmpv = { git = "…/libmpv-rs", branch = "master" }` — the
+dead pin [libmpv2-migration.md](libmpv2-migration.md) exists to replace. The
+render API lives in `libmpv2-sys` (`mpv_render_context_render`); the safe wrapper
+was only ever a PR against the old crate. **Use `libmpv2-sys` raw FFI directly in
+the spike.** Do not block the spike on the migration, and do not let the spike
+half-perform it — if the spike goes green the migration becomes a hard
+prerequisite of the implementation, which is the ordering
+[windows-native-audio-backend.md](windows-native-audio-backend.md) already sits
+in.
+
+### IPC
+
+None. The spike crosses no boundary. If it goes green, the follow-up changes only
+the *value* of the existing `useHtml5Element` / `supportsNativeVideo` fields — no
+new wire shapes, no `bindings.ts` regeneration.
+
+## Gates
+
+Each is pass/fail with a named failure. Stop at the first red and write it up —
+a red result is a successful spike.
+
+| # | Question | Fails if |
+|---|---|---|
+| G1 | Can a custom GTK widget join Tauri's widget tree and survive the window's lifetime? | `default_vbox()` is absent/unusable, or reparenting the webview breaks input or crashes. |
+| G2 | Does the webview still paint, with a transparent backdrop, over that widget? | The backdrop renders opaque black ([wry#1540](https://github.com/tauri-apps/wry/issues/1540)) or the webview stops repainting ([tauri#12800](https://github.com/tauri-apps/tauri/issues/12800)). **This is the highest-risk gate.** |
+| G3 | Does mpv render a frame into our FBO? | The render context refuses GTK's context, or frames land in the wrong buffer. |
+| G4 | Does HTML drawn over the video area actually appear over it? | Video covers the controls — the exact failure wry#284 and tauri#6343 report. Without this, the whole thing is worthless: our controls, subtitles and mini-player all sit over the video. |
+| G5 | Does it survive resize, fullscreen, and SPA navigation away and back? | Flicker on resize, or a surface that outlives its route. |
+| G6 | Does it hold on **both** X11 and Wayland? | Either session backend fails. Wayland is the one the 2024 maintainer quote says is impossible — test it first, not last. |
+
+G6 is not a nice-to-have. A result that only holds on X11 is red for a project
+shipping to current desktops.
+
+### Time box
+
+If G1–G4 are not all green, stop and write the result up. The value of this spike
+is a dated, method-specific answer — including "still no, and here is the
+mechanism" — not a working player.
+
+## Result (2026-08-21)
+
+Run on GNOME, kernel 7.1.8, libmpv 2.5.0 (mpv 0.41.0), GTK 3.24.52, WebKitGTK
+2.52.6, wry 0.53.5 — the versions `src-tauri/Cargo.lock` resolves. Spike source:
+a ~250-line standalone crate using wry + gtk + `libmpv2-sys` raw FFI, driving
+mpv's render API with an update callback, frame-gated repaints and
+`report_swap`.
+
+| Gate | Result | Observed mechanism |
+|---|---|---|
+| G1 widget in GTK tree | 🟡 **partial** | `GtkOverlay` with `GtkGLArea` as main child and the wry webview as overlay child works, built directly. **Tauri's own `default_vbox()` was not exercised** — see below. |
+| G2 webview paints transparently over it | ✅ green | `with_transparent(true)` alone. No window-level transparency was used or needed. |
+| G3 mpv renders into our FBO | ✅ green | `vo=libmpv` + `mpv_render_context_create` with `MPV_RENDER_PARAM_OPENGL_FBO` into the FBO GTK binds. |
+| G4 HTML over video | ✅ green | Opaque panel and a translucent control bar both drew over moving video. |
+| G5 resize / drag / fullscreen | 🟡 **green on appearance, suspect underneath** | No flicker, gap or misalignment, and smooth once frame pacing was correct (trap 3). But the only crash observed came from the only session where fullscreen was exercised — see "What is still open". |
+| G6 X11 **and** Wayland | ✅ green | Identical on both; `GDK_BACKEND` flipped between runs. |
+
+**Finding 2 of [playback-backend-unification.md](playback-backend-unification.md)
+is false on Linux** when tested by the render API rather than by foreign-window
+embedding. Wayland — the half the 2024 maintainer quote called impossible — is
+green.
+
+Better than the gate asked for: the translucent bar composited *alpha* against
+the video, not merely opaque-over. Scrims, gradient fades and subtitle backdrops
+therefore work, which is most of how a player UI actually looks. mpv also painted
+the letterbox bars black on its own — the Android equivalent was a shipped defect
+(DR-194).
+
+### Three traps, each of which cost a debugging cycle
+
+Carry these into the implementation; each produced a failure that looked like a
+platform limitation and was not.
+
+1. **`LC_NUMERIC` must be reset *after* `gtk::init()`, not before.** mpv refuses
+   to start under a non-C numeric locale. `mpv_backend.rs` already handles this,
+   but it has no GTK init in front of it; on this path `gtk::init()` applies the
+   user's locale afterwards and `mpv_create` returns null.
+2. **libepoxy exports GL entry points as *data* symbols.** There is no `glFoo`
+   function to resolve — there is `epoxy_glFoo`, a variable holding a lazily
+   resolving function pointer. `get_proc_address` must return the pointer *stored
+   at* that symbol; returning the symbol's own address makes mpv jump into
+   non-executable data and take SIGSEGV/SEGV_ACCERR on the first GL call. The
+   `epoxy` crate does this correctly but is unusable — its `gl_generator`
+   dependency pulls a yanked `xml-rs`.
+3. **Frame pacing is not optional, and its symptom is misleading.** Driving
+   `queue_render()` off the widget's frame clock on every tick, without calling
+   `mpv_render_context_report_swap` after each render, leaves mpv with nothing to
+   time against. Playback looks fine in a window and **judders at fullscreen** —
+   which reads as a compositing or GPU limit and is neither. The fix is to
+   register `mpv_render_context_set_update_callback`, redraw only when it says a
+   frame is ready, and report the swap afterwards. Fullscreen was smooth
+   immediately once both were in place.
+
+### Hardware decode through the render API
+
+Tested by asking mpv what it actually selected (`hwdec-current`), not what it was
+asked for. All three ran 20s clean at a steady 30 fps.
+
+| `hwdec` | `hwdec-current` | Note |
+|---|---|---|
+| `vaapi` | `no` | **Did not engage** on this box — silently fell back to software. `vainfo` is not installed, so the libva driver for the Iris Xe iGPU is likely absent. No render-API error; this looks like a missing driver package, not a compositing limit. |
+| `auto` | `nvdec-copy` | Hardware decode **does** work through the render API, on the discrete RTX 3050. Copy-back rather than zero-copy interop. |
+| `no` | `no` | Software. Clean baseline. |
+
+The load-bearing result is the middle row: **hardware decode is compatible with
+mpv's render API**, so the direct-play prize is real and not traded away for
+software decoding. Which decoder to prefer is an implementation question — on a
+hybrid Intel+NVIDIA laptop `auto` reached for the discrete GPU in copy-back mode,
+which is the least efficient hardware path. An implementation should evaluate
+zero-copy VA-API on the iGPU (after confirming the driver is installed) before
+accepting `auto`.
+
+`hwdec=auto-safe` probes Vulkan video decode, which this GPU does not support.
+It logs two `Failed setup for format vulkan` / `no frame!` pairs at start-up and
+then settles on `nvdec-copy` — the same place `auto` lands. A first reading of
+these logs mistook the start-up pair for a per-frame flood; **it is not**. Every
+run, clean or crashed, contains exactly two. `auto-safe` is not implicated in
+anything.
+
+### What is still open
+
+- **The Tauri half of G1.** The spike built its own `GtkOverlay`. The app must
+  instead reach `WebviewWindow::gtk_window()` / `default_vbox()` and reparent
+  Tauri's existing webview into an overlay. Low risk — the same widgets, one
+  extra reparent — but unproven, and it is the only place Tauri-specific
+  behaviour could still bite.
+- ✅ **ABR — resolved. Finding 3's premise is false.** Finding 3 said mpv would
+  regress streaming quality because "the webview path already has real ABR via
+  hls.js". Three pieces of evidence in this repo suggested that is **not true of
+  the URLs we actually build**:
+
+  1. `get_video_stream_url` (`repository/online.rs`) requests a *single*
+     rendition — one `VideoBitrate`, one `MaxStreamingBitrate`, one `MaxHeight`.
+     Jellyfin transcodes to what it is asked for; it does not build a ladder.
+  2. The frontend contains **no level-handling code at all** — no `hls.levels`,
+     no `LEVEL_SWITCH`, no `currentLevel`. The `abrEwma*` options in
+     `VideoPlayer.svelte` are default tuning with nothing to act on. hls.js is
+     serving as an HLS *demuxer* (WebKitGTK cannot play HLS natively), not as an
+     adaptation engine.
+  3. That function's own comment describes a quality switch as **rebuilding the
+     URL** — "every path that re-opens a stream (quality switch, transcoded seek,
+     audio-track switch)". Manual selection by stream re-open is what you build
+     when there is no adaptation, and mpv can do the same thing.
+
+  **The decisive test has now been run** (2026-08-21, against the development
+  server, Jellyfin 10.11.5):
+
+  ```
+  curl -s ".../Videos/<itemId>/master.m3u8?…&TranscodingProtocol=hls&…" \
+    | grep -c EXT-X-STREAM-INF
+  1
+  ```
+
+  **One line.** The playlist carries a single `EXT-X-STREAM-INF` plus an
+  `EXT-X-IMAGE-STREAM-INF` trickplay entry, which is not a rendition. Jellyfin
+  builds the master playlist from the rendition the request asked for; it does
+  not publish a ladder. So **there is no ABR to lose, and this blocker is
+  closed** — hls.js is serving as an HLS demuxer, exactly as (2) above supposed,
+  and mpv gives up nothing by replacing it.
+
+  Recorded as DR-229 (Won't Do) rather than deleted, because it is a
+  measurement: a server that *does* publish a ladder would change the answer, and
+  the re-negotiation path is the hook that work would build on.
+
+  **The direct-play path now exists.** It did not when this spike was written —
+  every video play went through the HLS transcode endpoint. Backend-owned stream
+  selection (DR-225 … DR-230) built it: Rust negotiates direct play / direct
+  stream / transcode and hands every backend one `StreamSelection` carrying the
+  URL, the transport and the chosen rendition. **That is the contract this
+  implementation consumes** — mpv is a consumer of a decision already made, not a
+  place to re-derive it.
+
+  It also sizes the prize precisely. Measured over the same server, 40 items
+  through a real negotiation per profile:
+
+  | Profile | Direct play |
+  |---|---|
+  | Linux / WebKitGTK — `h264` only, 2ch | **7%** |
+  | Android / ExoPlayer — `h264,hevc,vp8,vp9,av1,mpeg4` + `ac3,eac3`, 6ch | **85%** |
+
+  **The 85% is a ceiling, not a shipped result** — it was measured with a
+  profile containing `ac3,eac3`, which the Android device later used for
+  verification does not support.
+
+  The library sampled is ~80% hevc. Linux sits at 7% **solely because the
+  WebKitGTK profile can only claim h264** — not because of anything about the
+  server or the negotiation. mpv decodes hevc, so widening the Linux device
+  profile once mpv renders the picture is what converts that 7% toward the
+  Android figure. That conversion is the actual product of this work; the
+  compositing proven above is the mechanism that permits it.
+- 🔴 **One unexplained SIGSEGV.** A ~180s
+  run died in a *decoder* thread (libavcodec -> `av_log` -> libmpv's log handler
+  -> libc). No Tauri, wry, WebKitGTK, GTK or GL frame appears anywhere in the
+  stack, so the fault is on the mpv/ffmpeg side of the process rather than in the
+  compositing seam.
+
+  Three hypotheses were tested and **none reproduced it**:
+
+  | Hypothesis | Test | Result |
+  |---|---|---|
+  | `hwdec=auto-safe`'s Vulkan failures | 300s soak on `auto-safe` | Survived. Also based on a misreading — the failures are 2 per run at start-up, not per-frame. Dead. |
+  | Fullscreen transitions recreating the GL context under mpv's render context | 240s soak, ~120 automated transitions | Survived, no core dumped. |
+  | Continuous resize thrashing the GL framebuffer | 240s soak, ~2000 resizes | Survived, no core dumped. |
+
+  **The crash is therefore unexplained.** It was observed exactly once, in the
+  only session a human interacted with, and did not recur in ~13 minutes of
+  targeted stress across the three most plausible causes. It is recorded here
+  rather than dismissed precisely because nothing explains it: an intermittent
+  fault that nobody can reproduce is worse to inherit than a deterministic one,
+  not better.
+
+  The underlying concern stands regardless of which test eventually reproduces
+  it. A SIGSEGV in an unrelated thread is characteristic of memory corruption,
+  and this spike never calls `mpv_render_context_free` and never tears down on
+  `unrealize` — it has no defence against the GL context being recreated beneath
+  the render context. That is DR-184 on Android restated: a surface outliving its
+  player. An implementation must bind the two lifetimes together whether or not
+  this particular crash is ever explained.
+
+  **Therefore G5 is recorded green on appearance only**, and this crash is the
+  single largest piece of unfinished business in the spike. Do not read the green
+  gates above as "safe to build on" until it is explained or a long soak clears
+  it.
+- Long-run stability, seeking, track switching, HDR, and multi-window were not
+  exercised at all.
+
+## Out of scope
+
+- Any change to the shipping Linux video path. `experimentalNativeVideo` in
+  `adapters/index.ts` is a **suppressor, never a promoter**; the spike must not
+  change that.
+- Adaptive bitrate. See "The blocker a green spike does not clear".
+- Windows and macOS — different mechanisms, and **Windows is the easier case, not
+  the endangered one**. See below.
+- Android. Already shipped; it is the precedent, not the target.
+- Crossfade, the libmpv2 migration, and the audio-parity work.
+
+### Why Windows is unaffected, and cheaper
+
+Nothing here can regress Windows. `use_html5_element` is already a per-platform
+`cfg!` in `get_player_status` — Android native, everything else HTML5 — so
+divergent video paths are the existing design rather than something this
+introduces. Windows keeps `<video>` + hls.js whatever this spike returns.
+
+The mechanism does not port: `default_vbox()`, `GtkOverlay` and `GtkGLArea` are
+GTK3/WebKitGTK concepts. But the *question* is already answered more favourably
+there. Both mpv plugins list Windows as **fully tested** and Linux as broken,
+because WebView2 honours a transparent background — the "native surface beneath a
+transparent webview" approach that fails on WebKitGTK is the one that works on
+Windows. That asymmetry is why
+[windows-native-audio-backend.md](windows-native-audio-backend.md) can call
+Windows "the cleanest available win".
+
+Windows' cost is packaging, not compositing: the build cross-compiles with MSVC +
+`cargo-xwin`, so libmpv arrives as a bundled prebuilt DLL (the ⚠️ in finding 5's
+comparison table). That cost is already committed for *audio*. Once the DLL ships
+to replace `WebviewAudioBackend`, Windows video is largely a follow-on.
+
+Sequencing, if native video is ever pursued on both:
+
+1. [libmpv2-migration.md](libmpv2-migration.md) — prerequisite for either.
+2. [windows-native-audio-backend.md](windows-native-audio-backend.md) — already
+   specced; lands the DLL and a real Windows backend.
+3. Windows native video — cheap once 2 exists, and does not need this spike.
+4. Linux native video — needs this spike, and runs independently of 1–3.
+
+### Does this add a backend?
+
+No — and the trajectory is convergence, not proliferation.
+
+`create_player_backend` in `lib.rs` already selects between four
+`PlayerBackend` impls by `cfg!`: `MpvBackend` (Linux), `ExoPlayerBackend`
+(Android), `WebviewAudioBackend` (Windows and anything else), and `NullBackend`
+as the graceful-init fallback. The HTML5 video path is not among them — it is a
+frontend adapter reporting through `player_report_*`, not a `PlayerBackend`.
+
+This spike adds none of these. `MpvBackend` already exists and already runs on
+Linux; it merely sets `video = no` at construction. Giving it video widens an
+existing backend rather than introducing an engine.
+
+Following the sequence above, the count goes **down**: replacing
+`WebviewAudioBackend` with mpv on Windows leaves two native engines — mpv
+(Linux + Windows) and ExoPlayer (Android) — with native video riding on both.
+
+Two is the floor, for a reason worth stating so nobody re-litigates it: Android
+cannot drop ExoPlayer even if libmpv runs there, because the foreground service,
+`MediaSessionCompat` and lockscreen control are built on it (finding 7 puts the
+cost at that rewrite, not at the bindings). The HTML5 path does not go away
+either — it is the transcode/ABR route and the fallback.
+
+The trait surface converges too: `ExoPlayerBackend` already implements the
+video-surface lifecycle for Android native compositing, so teaching `MpvBackend`
+video follows a path already walked rather than opening a second one.
+- Adopting `tauri-plugin-libmpv` or `tauri-plugin-mpv` as dependencies. Both
+  report Linux window embedding as not working and are small projects
+  (20 and ~70 commits); read them, do not depend on them.
+
+## Acceptance criteria
+
+The deliverable is a decision, not a feature.
+
+- [ ] Each of G1–G6 recorded green/red **with the observed mechanism**, not just
+      the verdict.
+- [ ] X11 and Wayland results reported separately, each naming the compositor
+      and WebKitGTK version tested.
+- [ ] The direct-play/transcode ABR hypothesis recorded as open, with whatever
+      the spike learned about it.
+- [ ] `docs/specs/README.md` updated — this spec listed, and its row moved or
+      deleted per the result.
+- [ ] The Linux claim in the `createAdapter` doc comment
+      ([adapters/index.ts:12-13](../../src/lib/player/adapters/index.ts#L12-L13))
+      corrected either way: if red, cite this spike instead of asserting it; if
+      green, it is wrong and must be rewritten.
+- [ ] On **red**: finding 2 of
+      [playback-backend-unification.md](playback-backend-unification.md) gains a
+      dated note naming the render-API method as also tested, and this file is
+      deleted. The verdict lives in the design-authority spec, not in a second
+      file that contradicts nothing.
+- [ ] On **green**: an implementation spec exists, allocating ids from
+      **UR-077 / IR-033 / DR-216** (re-check `requirements.md` — the README's
+      "next free DR-215" is stale, DR-215 landed), and it must answer ABR before
+      being accepted.
+- [ ] No spike code on `master`. If any lands, the standard gates apply:
+      `bun run check`, `bun run test`, `bun run check:boundary`, `cargo fmt`,
+      `cargo clippy`, `bun run test:rust`.
+
+## Testing
+
+No automated tests. A compositing result is a visual, per-session-backend
+observation and cannot be asserted in `cargo test` or vitest — pretending
+otherwise would produce a test that passes on a headless runner and tells us
+nothing.
+
+Capture a screenshot per gate. G4 specifically: an opaque HTML element over the
+video area, photographed showing the video *behind* it.
+
+If it goes green, the implementation spec inherits the testable surface the
+Android work already established — `nativeVideoLayers.test.ts` asserts the
+`app.css` selector list and the `data-native-video` contract, and both are
+platform-agnostic.
+
+## TRACES
+
+None. No requirement-implementing code is produced. The implementation spec that
+follows a green result allocates from DR-216 and tags there.
+
+## Notes for the implementer
+
+- **Read [playback-backend-unification.md](playback-backend-unification.md)
+  first, in full.** This spike disputes exactly one of its six findings, on one
+  platform, by one method it did not try. Everything else in it is still binding
+  — particularly finding 3.
+- Test **Wayland first**. It is the gate most likely to be red and the one that
+  makes the rest moot.
+- The frontend plumbing already exists from the Android work: `createAdapter`,
+  `NativePlayerAdapter`, `nativeVideo.ts`, `videoSurface.ts`, and the
+  `[data-native-video="active"]` rule. A green spike is far cheaper to implement
+  than it would have been a year ago — which is itself part of why the question
+  is worth re-asking.
+- The Android record in
+  [05-platform-backends.md](../architecture/05-platform-backends.md#native-video-compositing-android)
+  lists six shipped defects from getting this right on one platform. Expect the
+  Linux equivalents (the surface outliving its player, the shell painting over
+  it, unpainted letterbox bars) rather than rediscovering them.
+- A parallel Claude session may be active in this repo — `git diff` before
+  "repairing" unexpected changes.

docs/specs/offline-downloaded-only-filter.md

@@ -1,235 +0,0 @@
-# Spec: Offline "downloaded only" filtering (issue #10)
-
-**Status:** Implemented
-**Scope:** Frontend (connectivity store) + Rust (hybrid repository). No new
-commands, no schema changes, no UI additions.
-**Requirements:** UR-052 → DR-078, DR-079, DR-080
-(see [requirements.md](../requirements.md)).
-**Tracking:** issue #10 — *"when offline the filter to show only downloaded
-media does not work."*
-
-## Summary
-
-Offline, a library page is supposed to show **only media on the device**, with a
-"Show all server media" toggle that additionally reveals the cached server
-catalog greyed out (queueable for download on reconnect). In practice the toggle
-does not gate the listing — every server item still appears. This spec fixes
-that with two independent changes; either one alone leaves the bug visible.
-
-## Background: what already exists
-
-Verified in code. **The feature is built and mostly correct — this is a
-two-point repair, not new infrastructure.** Do not rebuild the toggle, the
-command, or the SQL gate.
-
-1. **The SQL gate works and is unit-tested.**
-   [offline.rs](../../src-tauri/src/repository/offline.rs) — `get_items` appends
-   the synced-catalog `UNION` branch only when `include_catalog_browse()` is
-   true; with it false, only downloaded/local rows return. Guarded by
-   `test_get_items_toggle_gates_synced_catalog` (UT-067). **Do not touch the
-   query.**
-
-2. **The toggle → backend path is wired.** The `showServerCatalog` store and the
-   `set_show_server_catalog` command
-   ([catalog.rs](../../src-tauri/src/commands/catalog.rs)) drive the process-wide
-   `INCLUDE_CATALOG_BROWSE` flag. `pushCatalogVisibility` in
-   [offlineCatalog.ts](../../src/lib/services/offlineCatalog.ts) computes
-   `include = connected || showCatalog` and pushes it on every change.
-
-3. **Home-screen queries are already downloads-only.** `get_latest_items`,
-   `get_resume_items`, `get_recently_played_audio`, `get_resume_movies` all
-   `INNER JOIN downloads ... status = 'completed'`. They are unaffected — leave
-   them.
-
-4. **`MediaCard` already greys and queues.**
-   [MediaCard.svelte](../../src/lib/components/library/MediaCard.svelte) —
-   `isServerOnly` renders the greyed, inert card with a queue button; the queued
-   row heals its `stream_url` on reconnect via the offlineCatalog service. Leave
-   it.
-
-## The two defects
-
-### Defect A — offline is never actually entered (DR-079)
-
-`pushCatalogVisibility` keys off `isConnected`, but
-[connectivity.ts](../../src/lib/stores/connectivity.ts) derives:
-
-```ts
-isConnected = isOnline && isServerReachable   // isOnline = navigator.onLine
-```
-
-`navigator.onLine` is documented in that same file as **advisory only** — the
-Rust `ConnectivityMonitor` is the source of truth (principle: *reachability from
-real traffic*, DR-055). When the server is unreachable but the device link is
-up (server down, wrong LAN, VPN dropped), `isOnline` stays true, so `isConnected`
-stays true, so `include` stays true, so the backend keeps returning the full
-catalog. The user is "offline" in every meaningful sense but the toggle never
-gets a chance to gate anything.
-
-This is the primary cause: it explains why the filter looks dead rather than
-merely inverted — the gate never closes.
-
-### Defect B — an intentionally empty result falls through to the server (DR-080)
-
-With the gate off and nothing downloaded in a library, offline `get_items`
-correctly returns few or zero rows. But
-[hybrid.rs](../../src-tauri/src/repository/hybrid.rs) treats a cache result as a
-hit only `if data.has_content()`. An empty offline result is indistinguishable
-from a cache miss, so `HybridRepository::get_items` (and `parallel_race`, used by
-~10 other reads) falls through to the server and returns the full server list —
-re-defeating the filter even after Defect A is fixed.
-
-## Design
-
-### Fix A: `isConnected` follows backend reachability alone (DR-079)
-
-In [connectivity.ts](../../src/lib/stores/connectivity.ts), redefine the derived
-store:
-
-```ts
-export const isConnected = derived(
-  connectivity,
-  ($c) => $c.isServerReachable
-);
-```
-
-`navigator.onLine` stays wired to what it is good for — a *trigger* for an
-immediate recheck (`online`/`offline` listeners already call
-`checkServerReachable()`); it must no longer be a *term* in the offline decision.
-Leave `isOnline` on the state object and the listeners intact.
-
-Consider whether the optimistic `isServerReachable: true` startup default
-([connectivity.ts](../../src/lib/stores/connectivity.ts)) should hold until the
-first real check resolves. Keep it — flipping the app to "offline" on launch is a
-worse regression than a brief full-catalog flash before the first probe. Note the
-choice in a comment.
-
-**Blast radius — this is the reason this is a spec, not a patch.** `isConnected`
-is consumed beyond this feature (banners, `MediaCard`, mini-player gating,
-anything importing it). Enumerate consumers first:
-
-```
-grep -rn "isConnected" src/ | grep -v node_modules
-```
-
-For each, confirm "server unreachable" (not "device link down") is the correct
-trigger. It almost always is — that is the whole point of the reachability model
-— but verify rather than assume, and call out anything that genuinely wanted the
-device link in the PR description.
-
-### Fix B: an empty offline result is authoritative when the gate is off (DR-080)
-
-The backend must distinguish "cache is cold, go ask the server" from "user asked
-for downloads only and there are none here." The gate flag already encodes intent
-— reuse it.
-
-Add a getter beside the existing setter in
-[offline.rs](../../src-tauri/src/repository/offline.rs):
-
-```rust
-pub fn include_catalog_browse() -> bool { /* pub, already exists privately */ }
-```
-
-In [hybrid.rs](../../src-tauri/src/repository/hybrid.rs) `get_items`: when
-`!include_catalog_browse()`, treat the offline result as authoritative and return
-it **as-is even when empty** — do not spawn/await the server fallback for this
-call. When the flag is on (online fast-path, or offline with the toggle on),
-behaviour is unchanged: empty cache still falls through to the server.
-
-Keep it surgical:
-
-- Scope the change to `get_items`. The gate is a `get_items` concept; do not
-  thread it into `parallel_race` or the other readers, which have no catalog
-  gate and legitimately want the server on an empty cache.
-- Preserve the online path exactly: with the flag on (its default, and always so
-  while reachable) the method behaves as it does today, including the background
-  cache refresh on a hit.
-- The flag is process-global `Relaxed`; it is set from the frontend before the
-  query. That ordering already holds for the SQL gate — no new synchronization.
-
-### Why both
-
-Fix A closes the gate; Fix B stops the hybrid from re-opening it. A alone: with
-downloads present the list still gets padded by the server fallback whenever a
-library's cache is thin. B alone: the gate never closes because `isConnected`
-never goes false on a live link. Ship them together.
-
-## Out of scope
-
-- The SQL gate, the toggle, the command, `INCLUDE_CATALOG_BROWSE` — all correct.
-- `MediaCard` greying / queue-on-reconnect — correct.
-- Home-screen and resume queries — already downloads-only.
-- The Rust `ConnectivityMonitor` reachability logic itself — unchanged; this
-  spec only stops the *frontend* from diluting its verdict with `navigator.onLine`.
-- Any new IPC command, DB column, or settings entry.
-- Making the "Show all server media" toggle reachable from Settings (that is a
-  UX-placement question, tracked separately under UR-051's toggle note).
-
-## Acceptance criteria
-
-- [~] With the server unreachable on a live device link, a library page lists
-      only downloaded media when the toggle is off (IT-016 — pending e2e; unit
-      coverage via UT-069 + gate tests).
-- [x] Turning the toggle on reveals the greyed-out cached catalog; turning it off
-      hides it again — without leaving/re-entering the page (SQL gate + toggle
-      wiring unchanged; UT-068 confirms the flag is pushed on toggle change).
-- [x] A library with downloads and a thin cache does not get padded with
-      non-downloaded server items when offline with the toggle off (Defect B —
-      UT-070: gate off + empty offline result returned as-is, server not queried).
-- [x] `isConnected` is false whenever the server is unreachable, regardless of
-      `navigator.onLine`; true for a reachable server even if the browser reports
-      offline (UT-069).
-- [x] Every existing `isConnected` consumer still behaves correctly (banner in
-      `+layout.svelte`, `MediaCard`, `favorites.ts` server-write skip — all want
-      "server unreachable", which is the new semantics; `CastButton`'s local
-      `isConnected` is unrelated). Full frontend suite (616 tests) green.
-- [x] Online behaviour is unchanged: with the flag on (its default, always so
-      while reachable) `get_items` keeps the offline fast-path and background
-      refresh (UT-067 + gate-on fall-through test).
-- [~] A download queued from a greyed offline card resolves and starts on
-      reconnect (IT-017 — regression check, no code change; offlineCatalog
-      resume path untouched).
-- [x] `bun run check`, `bun run test`, and `bun run test:rust` pass;
-      `cd src-tauri && cargo fmt && cargo clippy` clean (no new warnings in the
-      touched files).
-
-## Testing
-
-Rust ([offline.rs](../../src-tauri/src/repository/offline.rs) /
-[hybrid.rs](../../src-tauri/src/repository/hybrid.rs) test modules):
-
-- **UT-070** — hybrid `get_items` with the gate off returns an empty offline
-  result as-is and does **not** query the server. Assert via a mock online repo
-  whose `get_items` bumps a call counter that must stay at zero.
-- Gate on + empty cache still falls through to the server (guard the online path).
-- UT-067 (`test_get_items_toggle_gates_synced_catalog`) must still pass untouched.
-
-Frontend (vitest, `src/lib/**/*.test.ts`):
-
-- **UT-069** — `isConnected` follows `isServerReachable` alone: false when
-  unreachable with `navigator.onLine === true`; true when reachable with
-  `navigator.onLine === false`.
-- **UT-068** — `pushCatalogVisibility` resolves `serverReachable || showCatalog`
-  and pushes to the backend on a change of either input (extend the existing
-  offlineCatalog tests).
-
-Integration (IT-016, IT-017) are documented as pending in
-[requirements.md](../requirements.md); wire them if the e2e harness can simulate
-an unreachable-server-on-live-link state, otherwise leave them pending with a note.
-
-New/changed requirement code keeps its `TRACES:` comments — see
-[CLAUDE.md](../../CLAUDE.md). The affected files already carry tags:
-`connectivity.ts` (`… | DR-079`), `hybrid.rs` (`… | DR-080`), `offline.rs`
-(`… | DR-078`). Update the getter's tag when you expose it.
-
-## Notes for the implementer
-
-- Read [docs/architecture/07-connectivity.md](../architecture/07-connectivity.md)
-  before Fix A — it is the canonical statement of the reachability model this fix
-  restores fidelity to.
-- Fix B relies on the frontend having pushed the flag before the query runs; that
-  ordering already holds for the SQL gate today. No new locking.
-- Another session is active in this repo (WiFi-only downloads, account menu
-  landed alongside this work). Check `git diff` before "repairing" unexpected
-  changes, and expect requirement IDs around UR-052 / DR-078 to be adjacent to
-  other new rows.

docs/specs/playback-backend-unification.md

@@ -0,0 +1,251 @@
+# 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:** informed the Android native-video and audio-parity
+work (both since shipped — see
+[05-platform-backends.md](../architecture/05-platform-backends.md)) and
+[windows-native-audio-backend.md](windows-native-audio-backend.md), still open
+
+## 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.**
+
+> **Re-opened on Linux (2026-08-21).** This finding's general form has since been
+> falsified on Android — native video now composites behind a transparent Tauri
+> WebView and ships on by default (see
+> [05-platform-backends.md](../architecture/05-platform-backends.md#native-video-compositing-android)).
+> The evidence above is also entirely about *foreign-window embedding*; mpv's
+> render API, drawing into a GL context we own inside Tauri's own GTK tree, was
+> never tested. [linux-native-video-spike.md](linux-native-video-spike.md) tests
+> that one claim on Linux. **Findings 3-6 below are untouched by it** - in
+> particular finding 3, which is an independent disqualifier a green spike would
+> not clear.
+
+### 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.
+
+> **Premise in doubt (2026-08-21).** "The webview path already has real ABR"
+> was not verified against the URLs this app actually builds.
+> `get_video_stream_url` requests a *single* rendition (one `VideoBitrate`, one
+> `MaxHeight`), the frontend has **no** level-handling code (`hls.levels`,
+> `LEVEL_SWITCH`, `currentLevel` appear nowhere), and this repo implements a
+> quality switch by *re-opening the stream* — all of which point to a
+> single-variant playlist, i.e. no ABR to lose. The decisive test is counting
+> `#EXT-X-STREAM-INF` lines in a real `master.m3u8`; it needs a live server and
+> has not been run. See
+> [linux-native-video-spike.md](linux-native-video-spike.md).
+
+### 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
+   [05-platform-backends.md](../architecture/05-platform-backends.md#native-video-compositing-android).
+3. **Audio parity is the real gap** and is achievable without touching any of the
+   above. See [05-platform-backends.md](../architecture/05-platform-backends.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/player-facade-enforcement.md

@@ -0,0 +1,263 @@
+# Spec: Enforce the unified player boundary
+
+**Status:** Proposed — not started. The count below has not improved: ~60
+`commands.player*` call sites still live outside `src/lib/player/`, and no lint
+rule enforces the boundary. This remains the one stated design principle with no
+automated check.
+**Requirements:** ⚠️ the suggested id **DR-095 has since been allocated** to seek
+clamping — allocate a fresh id (DR-215 or later) on implementation. 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 +1,227 @@
+# Spec: Two-path media — selectable playback bitrate, independent whole-file download
+
+**Status:** Partially implemented. Landed: the cache/download unification
+(DR-126, DR-127 — a cache entry *is* a `downloads` row with a shorter life, and
+eviction only reclaims the temporary tier), local playback of downloaded media
+(DR-128), and the one-path/one-row invariants that followed (DR-133 … DR-138).
+DR-123 is in progress. Still open: the read-through capture itself — DR-122,
+DR-124, DR-125.
+
+**DR-121 has shipped and left this spec.** The player quality selector, the
+per-playback bitrate ceiling, and the backend-owned stream decision it needed
+were built as *backend-owned stream selection* (DR-225 … DR-228) and are
+described in
+[01-rust-backend.md](../architecture/01-rust-backend.md#stream-selection) and
+[03-data-flow.md](../architecture/03-data-flow.md#video-stream-selection-flow).
+The settings-level ceiling (DR-162) is the same section. What remains here is the
+*capture* half only — this spec no longer specifies anything about choosing a
+bitrate.
+
+**Requirements:** UR-070, UR-071 → DR-122, DR-123, DR-124, DR-125; IR-032
+**Related:** the locally-indexed search and downloaded-browse work, both
+shipped — see
+[03-data-flow.md](../architecture/03-data-flow.md) and
+[06-downloads-and-offline.md](../architecture/06-downloads-and-offline.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 — moved out (shipped)
+
+Bitrate selection in the player shipped as DR-225 … DR-228; see
+[01-rust-backend.md](../architecture/01-rust-backend.md#stream-selection).
+
+The one constraint here that the capture work still has to respect: a quality
+change re-negotiates **within HLS**. Returning a progressive `stream.mp4` for a
+transcode means playback never starts, because the server encodes the whole file
+before serving a byte (DR-140). That is why DR-122 below abandons a capture on a
+quality change rather than trying to splice one.
+
+### 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 |
+|---|---|
+| 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/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 by the tripwire hardening (DR-094, shipped);
+   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
+the hardened tripwire (DR-094) can pass.
+
+**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
+  the hardened tripwire (DR-094, see `scripts/check-frontend-boundary.sh`).
+- 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
+  the hardened tripwire (DR-094)**. 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-boundary.md

@@ -1,6 +1,13 @@
 # Spec: Move search scope taxonomy behind the Rust boundary
 
-**Status:** Proposed
+**Status:** Design authority — **Stage 1 implemented**, Stage 2 outstanding.
+The scope→item-type mapping now lives in Rust (`SearchScope::item_types()` in
+`repository/types.rs`, DR-063 … DR-067). The *result-side* grouping table
+(`GROUP_ITEM_TYPES` in `src/lib/utils/searchScope.ts`) is still in the
+frontend, and `check:boundary` does not match its shape. Delivery status and
+the remaining work live in
+[scoped-search-boundary-implementation.md](scoped-search-boundary-implementation.md);
+this spec remains the design authority.
 **Scope:** Rust + Frontend. **Revises a decision in
 [scoped-search.md](scoped-search.md).**
 **Requirements:** UR-049, UR-050 (existing) → new DRs for the boundary move

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/video-background-audio.md

@@ -1,233 +0,0 @@
-# Spec: Background audio for video playback (Android)
-
-**Status:** Draft
-**Scope:** Android only (v1). Linux noted as future work.
-**Branch base:** `android-picture-in-picture`
-**Requirements:** UR-040 → IR-025, JA-032, DR-051, DR-052 (see
-[requirements.md](../requirements.md)). Tests: UT-059, UT-060, UT-061, IT-013.
-
-## Summary
-
-Add a per-player toggle that lets the **audio** of a video keep playing when the
-app is backgrounded or the screen is locked, while **video decoding stops**.
-When the app returns to the foreground, video decoding resumes from the current
-audio position.
-
-This is the audio-first counterpart to the existing Picture-in-Picture feature
-(which keeps the *whole video* decoding in a floating window). The two are
-mutually exclusive: enabling background audio suppresses auto-PiP.
-
-## Motivation
-
-Users watching talk-heavy content (podcasts-as-video, lectures, music videos,
-concert films) want to lock the phone or switch apps and keep listening without
-draining battery on video decode or needing a visible floating window.
-
-## Background: how playback actually works here
-
-Two facts drive the entire design (verified in code, not assumed):
-
-1. **Video renders through the HTML5 `<video>` element in the WebView on both
-   platforms.** The native ExoPlayer *video* surface path is disabled — see the
-   INTERIM override in
-   [VideoPlayer.svelte](../../src/lib/components/player/VideoPlayer.svelte)
-   around the `playerPlayItem` response handling (`useHtml5Element` is forced
-   `true`, native backend is stopped). So "video decoding" == the WebView
-   `<video>` element, and the WebView is what Android suspends on background.
-
-2. **An Android WebView `<video>` element does not keep playing audio when the
-   app is backgrounded / locked.** The system throttles the WebView and media
-   pauses. Keeping audio alive in the background requires a **native foreground
-   media service**, which already exists for music:
-   [`JellyTauPlaybackService`](../../src-tauri/android/src/main/java/com/dtourolle/jellytau/player/JellyTauPlaybackService.kt)
-   +
-   [`JellyTauPlayer`](../../src-tauri/android/src/main/java/com/dtourolle/jellytau/player/JellyTauPlayer.kt)
-   (ExoPlayer) + `MediaSessionCompat`.
-
-**Therefore the design is a handoff**, not "keep the WebView alive": on
-background, stop the WebView `<video>` and start audio-only playback of the same
-item through the existing native ExoPlayer audio service; on foreground, hand
-back to the WebView `<video>`.
-
-This also aligns with the project's one-directional playback rule
-(`CLAUDE.md` → "Playback state is one-directional"): the currently-authoritative
-player (WebView element **or** native audio service) drives position; the UI and
-MediaSession consume it. The handoff is a change of *which* player is
-authoritative, and must transfer position cleanly.
-
-## User-facing behavior
-
-### The toggle
-
-- A toggle button in the video player controls (next to the existing PiP /
-  fullscreen buttons in
-  [VideoPlayer.svelte](../../src/lib/components/player/VideoPlayer.svelte)).
-- Icon: headphones / "audio-only" glyph. Two visual states (on/off).
-- **Visible only when** `isPipSupported()`-equivalent conditions hold — i.e.
-  Android with a native audio service available. Hidden on Linux in v1.
-- State is a UI preference on the player. Consider persisting the last choice
-  per user (see Open Questions) — v1 may default OFF each session.
-
-### When toggle is ON and the app goes to background / screen locks
-
-1. Auto-PiP is suppressed (see "Interaction with PiP").
-2. The WebView `<video>` is paused and its decode stopped (release the media
-   source so the decoder is freed, not merely `pause()`).
-3. Native audio-only playback of the same item starts at the current position,
-   through `JellyTauPlaybackService` (foreground notification + lockscreen
-   controls via the existing `MediaSessionCompat`).
-4. Lockscreen / notification shows the item with play/pause/seek, driven by the
-   native player (existing music behavior — reused, not rebuilt).
-
-### When toggle is ON and the app returns to foreground
-
-1. Native audio playback stops; its final position is captured.
-2. WebView `<video>` reloads/resumes at that position and continues as normal
-   audiovisual playback.
-3. Playback state (playing/paused) is preserved across the handoff.
-
-### When toggle is OFF (default)
-
-Current behavior is unchanged: backgrounding video auto-enters PiP
-(`onUserLeaveHint` → `PictureInPictureManager.enterPip`).
-
-## Interaction with PiP
-
-The toggle chooses one behavior or the other:
-
-- Toggle **ON** → call `AndroidPictureInPicture.setAutoEnterEnabled(false)` (the
-  bridge already exists,
-  [pictureInPicture.ts](../../src/lib/utils/pictureInPicture.ts) →
-  `setAutoEnterEnabled`). Background → audio handoff instead of PiP.
-- Toggle **OFF** → `setAutoEnterEnabled(true)`. Background → PiP (status quo).
-
-The frontend must also call `setAutoEnterEnabled(false)` on unmount if it left
-it enabled, and re-assert the correct value whenever the toggle changes, so a
-stale setting can't leak into the next player.
-
-> Note: `canEnterPip()` today requires `isPlayingVideo()` on the *native*
-> ExoPlayer, but video plays via the WebView, so native `isPlayingVideo()` is
-> false during normal playback. Confirm during implementation how auto-PiP is
-> actually triggering today (it may rely on a different signal), because the
-> background-audio handoff needs the same "is a local video active" signal to
-> know it should fire. **This is a load-bearing unknown — resolve it first
-> (Phase 0).**
-
-## Technical design
-
-### The audio-only stream
-
-Jellyfin can transcode/stream a video item as audio-only. Add a repository
-method (mirroring
-[`get_video_stream_url`](../../src-tauri/src/repository/online.rs) and
-[`get_audio_stream_url`](../../src-tauri/src/repository/mod.rs)) that returns an
-**audio-only stream URL for a video item** at a given audio-stream index — so
-the currently-selected audio track (`selectedAudioTrackIndex` in the player)
-carries over. Prefer direct-play of the audio stream where the container/codec
-allows; transcode to a broadly-supported audio codec otherwise.
-
-Position semantics must match between the WebView `<video>` timeline and the
-audio stream (account for the transcoded-HLS `seekOffset` model already in the
-player — see the `seekOffset` handling in `VideoPlayer.svelte`).
-
-### Backend command surface (Rust)
-
-New/extended `#[tauri::command]`s in `src-tauri/src/commands/player/` (follow the
-camelCase param rule and `Result<T, String>` convention):
-
-- `player_enter_background_audio(item_id, position_seconds, audio_stream_index)`
-  — stop WebView authority, start native audio-only playback at position; makes
-  the native player authoritative. Emits state via the existing player-event
-  channel so MediaSession/UI stay consumers.
-- `player_exit_background_audio() -> position_seconds` — stop native audio,
-  return final position for the WebView to resume from; restores WebView
-  authority.
-
-Reuse existing `player_play_*` / `player_stop` plumbing where possible rather
-than adding a parallel path.
-
-### Android native
-
-- Reuse `JellyTauPlaybackService` + `JellyTauPlayer` audio path
-  (`MediaSessionCompat`, foreground notification, audio-becoming-noisy, etc. —
-  all already implemented for music).
-- Add a bridge method (alongside `AndroidPictureInPicture`) or reuse an existing
-  one so the frontend can signal "prepare for background audio handoff" tied to
-  the Activity lifecycle (`onPause`/`onStop`/`onUserLeaveHint`).
-- On `onUserLeaveHint` / screen-off with background-audio enabled: **do not**
-  enter PiP; instead trigger the handoff command.
-- Respect the deadlock gotchas in `CLAUDE.md` (no sync/blocking calls from
-  player event callbacks; bind locked `AutoplayDecision` to a `let` before
-  matching).
-
-### Frontend (VideoPlayer.svelte)
-
-- Add toggle state + button. On change, call `setAutoEnterEnabled(!on)`.
-- Listen for Android lifecycle background/foreground signals (via a bridge event
-  or existing visibility hooks) and:
-  - background + ON → `player_enter_background_audio(...)`, pause + tear down the
-    `<video>`/HLS decode (reuse the existing HLS teardown sequence to avoid dual
-    audio).
-  - foreground + ON → `player_exit_background_audio()`, reload `<video>` at the
-    returned position, restore play/pause state.
-- **Follow the native-mode pitfall** (memory:
-  `videoplayer-native-mode-pitfalls`): no lifecycle calls after an `await` in
-  `onMount`. Keep the handoff logic out of that window.
-- Dual-audio is the key regression risk: at every handoff exactly one of
-  {WebView `<video>`, native ExoPlayer} produces audio. Tear the other down
-  *before* starting the next, mirroring the existing HLS cleanup discipline.
-
-## Phasing
-
-- **Phase 0 — De-risk (do first):**
-  - Confirm what actually triggers today's auto-PiP given video is on the
-    WebView (resolve the `canEnterPip`/`isPlayingVideo` question).
-  - Spike: obtain an audio-only stream URL for a video item and play it through
-    the native audio service; measure position accuracy and that WebView audio
-    is fully silenced (no dual audio).
-- **Phase 1 — Backend:** repository audio-only-URL method + the two player
-  commands + events.
-- **Phase 2 — Native:** lifecycle wiring, PiP suppression, handoff trigger.
-- **Phase 3 — Frontend:** toggle UI, lifecycle listeners, handoff calls,
-  teardown discipline.
-- **Phase 4 — Polish:** persist toggle preference, subtitle/audio-track
-  carry-over, edge cases (calls, headphone unplug, autoplay-next during
-  background audio).
-
-## Testing
-
-- Rust: unit tests for the audio-only URL builder and the two commands
-  (`cargo test`, `bun run test:rust`).
-- IPC param-naming integration tests for any new commands
-  (`bun run test -- tauriIntegration.test.ts`).
-- Frontend: `bun run check`, `bun run test`, plus a VideoPlayer logic test for
-  the handoff state machine (mirror the existing
-  `VideoPlayer.logic.test.ts`).
-- Manual on-device matrix:
-  - toggle ON: home button → audio continues, video stops decoding; return →
-    video resumes at position; playing/paused preserved.
-  - toggle ON: screen lock → audio continues; lockscreen controls work; unlock →
-    resumes.
-  - toggle OFF: background → PiP (unchanged).
-  - No dual audio at any transition. No audio leak after leaving the player.
-  - Transcoded (HEVC/10-bit) item — verify position with `seekOffset`.
-  - Autoplay-next fires correctly if an episode ends during background audio.
-
-## Open questions
-
-1. **Persist the toggle per user/series, or default OFF each session?**
-   (Recommend: remember last choice; series-level like the audio-track
-   preference is a nice-to-have.)
-2. **Autoplay-next during background audio** — should the next episode start as
-   audio-only and stay audio until foreground, or pause at episode end? (Recommend:
-   continue as audio-only.)
-3. **Subtitles** are irrelevant in audio-only mode but must restore on
-   foreground — confirm they survive the `<video>` teardown/reload.
-4. Exact **Android lifecycle signal** for "screen locked" vs "app backgrounded"
-   — `onUserLeaveHint` covers Home but not lock; may need a screen-off receiver.
-
-## Non-goals (v1)
-
-- Linux background audio (desktop windows keep running unfocused; low value).
-- Replacing or removing PiP — it stays as the toggle-OFF behavior.
-- Re-enabling the native ExoPlayer *video* surface path.

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

@@ -0,0 +1,213 @@
+# Spec: Windows native audio backend
+
+**Status:** Proposed — not started. Windows still runs on
+`WebviewAudioBackend`. Blocked on [libmpv2-migration.md](libmpv2-migration.md),
+whose crate swap has not landed either.
+**Requirements:** UR-003, UR-027, UR-032, UR-033 → DR-030, DR-035, DR-036;
+⚠️ the suggested id **IR-030 has since been allocated** to the scheduled catalog
+crawl — allocate a fresh id (IR-033 or later) on implementation
+**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

@@ -15,7 +15,7 @@ The CI/CD pipeline automatically validates that code changes are properly traced
 Traceability validation lives in `.gitea/workflows/traceability-check.yml`:
 
 - ✅ Automatic trace extraction
-- ✅ Coverage validation against minimum threshold (50%)
+- ✅ Coverage validation against minimum threshold (88%, ratcheted)
 - ✅ Modified file checking
 - ✅ Artifact preservation
 - ✅ Summary reports
@@ -43,14 +43,58 @@ 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:** 88% (`MIN_THRESHOLD`)
 
-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
+> The fix derives the denominators from `requirements.md` at run time.
+
+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 the threshold —
+or if it computes above 100%, which can only mean the gate is miscounting.
+
+#### Ratchet policy
+
+`MIN_THRESHOLD` **only ever goes up.** It is deliberately set a few points below
+the coverage actually achieved (88 against a real ~90%), so a genuine regression
+trips it. It previously sat at 50 while true coverage was 86%: nearly half the
+matrix could have rotted before CI objected. It was ratcheted 50 → 82 when that
+was found, and 82 → 88 once coverage had held above 88% for several releases.
+
+When coverage rises durably, raise the threshold to just under the new figure.
+**Never lower it to make a red build pass** — add the missing TRACES comments
+instead. The same number lives in `MIN_COVERAGE_PERCENT` in
+`scripts/extract-traces.ts` (so `bun run traces:coverage` gates locally on the
+same bar); `scripts/extract-traces.test.ts` fails if the two drift apart.
+
+### 2b. Dangling requirement IDs
+
+```bash
+bun run traces:validate
+```
+
+Every ID named by a `TRACES:` comment must be defined as a table row in
+`docs/requirements.md`. The extractor used to accept any well-formed ID
+silently, so a typo or a rename that missed a call site passed unnoticed —
+`DR-189` and `UT-188` were referenced from three source files, defined nowhere,
+for months.
+
+This check spans **all six** ID types (UR/IR/DR/JA/UT/IT), unlike the coverage
+`orphaned` list above, which considers only the four requirement types so that
+UT/IT noise cannot bury a real typo in the ratio's reporting. The workflow step
+**fails the build** on any dangling ID and prints each offender with the files
+that reference it.
 
 ### 3. Modified File Checking
 On pull requests, the workflow:
@@ -108,13 +152,13 @@ TRACES: [UR-###, ...] | [IR-###, ...] | [DR-###, ...] | [JA-###, ...]
 
 ### On Push to Main Branch
 1. ✅ Extracts all traces from code
-2. ✅ Validates coverage is >= 50%
+2. ✅ Validates coverage is >= 88%
 3. ✅ Generates full traceability report
 4. ✅ Saves report as artifact
 
 ### On Pull Request
 1. ✅ Extracts all traces
-2. ✅ Validates coverage >= 50%
+2. ✅ Validates coverage >= 88%
 3. ✅ Checks modified files for TRACES
 4. ✅ Warns if new code lacks TRACES
 5. ✅ Suggests proper format
@@ -122,7 +166,8 @@ TRACES: [UR-###, ...] | [IR-###, ...] | [DR-###, ...] | [JA-###, ...]
 
 ### Failure Scenarios
 The workflow **fails** (blocks merge) if:
-- Coverage drops below 50%
+- Coverage drops below 88%
+- A `TRACES:` comment names an ID `docs/requirements.md` does not define
 - JSON extraction fails
 - Invalid trace format
 
@@ -153,16 +198,18 @@ 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 August 2026 overall coverage is ~90%.
 
 ### Targets
-- **Short term** (Sprint): Maintain ≥50% overall
-- **Medium term** (Month): Reach 70% overall coverage
-- **Long term** (Release): Reach 90% coverage with focus on:
+- **Short term** (Sprint): Maintain ≥88% overall (the current ratchet)
+- **Medium term** (Month): Hold above 90% and ratchet the gate to match
+- **Long term** (Release): Reach 95% coverage with focus on:
   - IR requirements (API clients)
   - JA requirements (Jellyfin API endpoints)
   - Remaining UR/DR requirements
@@ -195,14 +242,14 @@ When submitting a pull request:
 
 - [ ] All new code has TRACES comments linking to requirements
 - [ ] TRACES format is correct: `// TRACES: UR-001 | DR-002`
-- [ ] Workflow passes (coverage ≥ 50%)
+- [ ] Workflow passes (coverage ≥ 88%)
 - [ ] No coverage regressions
 - [ ] Artifact traceability report was generated
 
 ## Troubleshooting
 
 ### "Coverage below minimum threshold"
-**Problem:** Workflow fails with coverage < 50%
+**Problem:** Workflow fails with coverage < 88%
 
 **Solution:**
 1. Run `bun run traces:json` locally

docs/traces-quick-ref.md

@@ -52,10 +52,10 @@ fn test_queue_next() {
 
 ## Where to Find Requirements
 
-1. **User Requirements (UR):** [README.md](README.md#1-user-requirements)
-2. **Integration Requirements (IR):** [README.md](README.md#21-integration-requirements)
-3. **Development Requirements (DR):** [README.md](README.md#23-development-requirements)
-4. **Jellyfin API (JA):** [README.md](README.md#22-jellyfin-api-requirements)
+1. **User Requirements (UR):** [requirements.md](requirements.md#1-user-requirements)
+2. **Integration Requirements (IR):** [requirements.md](requirements.md#21-integration-requirements)
+3. **Development Requirements (DR):** [requirements.md](requirements.md#23-development-requirements)
+4. **Jellyfin API (JA):** [requirements.md](requirements.md#22-jellyfin-api-requirements)
 
 ## How to Add TRACES
 
@@ -133,18 +133,19 @@ bun run traces:json | jq '.requirements."UR-005"'
 ### Before Committing
 1. Ensure all new code has TRACES
 2. Format is correct: `// TRACES: ...`
-3. Requirements exist in README.md
-4. No typos in requirement IDs
+3. Requirements exist in `docs/requirements.md` — `bun run traces:validate`
+4. No typos in requirement IDs (same command catches them)
 
 ## CI/CD Validation
 
 The workflow automatically checks:
-- ✅ Coverage stays >= 50%
+- ✅ Coverage stays >= 88% (a ratchet — raise it, never lower it)
+- ✅ Every traced ID is defined in `docs/requirements.md`
 - ✅ New files have TRACES
 - ✅ JSON format is valid
 - ✅ Reports are generated
 
-See [traceability-ci.md](docs/traceability-ci.md) for details.
+See [traceability-ci.md](traceability-ci.md) for details.
 
 ## Tips & Tricks
 
@@ -198,10 +199,10 @@ A: Yes! TRACES show your implementation plan.
 
 ## See Also
 
-- [Full Traceability Matrix](docs/traceability.md)
-- [CI/CD Pipeline Guide](docs/traceability-ci.md)
-- [Requirements Specification](README.md)
-- [Extraction Script](scripts/README.md#extract-tracests)
+- [Full Traceability Matrix](traceability.md)
+- [CI/CD Pipeline Guide](traceability-ci.md)
+- [Requirements Specification](requirements.md)
+- [Extraction Script](../scripts/README.md#extract-tracests)
 
 ---
 

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)
@@ -632,7 +634,7 @@ episode strip.
 │   │  S2E4 • 48m • ★8.1                        │ │
 │   │  Overview…                                │ │
 │   │  ▓▓▓▓▓░░░░░  32m left                     │ │
-│   │  [▶ Play]                                 │ │
+│   │  [▶ Play]  [⬇]  [♡]                       │ │
 │   └───────────────────────────────────────────┘ │
 │                                                 │
 │  More Episodes                                  │  ← 2. EPISODE STRIP
@@ -686,10 +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
 ```
@@ -698,6 +700,29 @@ 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
@@ -732,6 +757,130 @@ 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 [architecture/01-rust-backend.md](architecture/01-rust-backend.md#favorites-system) 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
 
 Search is **context-scoped**: what you are looking at when you start a search

e2e/.env.example

@@ -1,24 +0,0 @@
-# E2E Test Configuration
-# Copy this file to .env and fill in your test credentials
-
-# Jellyfin Server Configuration
-TEST_SERVER_URL=https://demo.jellyfin.org/stable
-TEST_SERVER_NAME=Demo Server
-
-# Test User Credentials
-TEST_USERNAME=demo
-TEST_PASSWORD=
-
-# Optional: Specific test data IDs (for testing playback, etc.)
-# You can find these IDs in your Jellyfin server
-TEST_MUSIC_LIBRARY_ID=
-TEST_MOVIE_LIBRARY_ID=
-TEST_ARTIST_ID=
-TEST_ALBUM_ID=
-TEST_TRACK_ID=
-TEST_MOVIE_ID=
-TEST_EPISODE_ID=
-
-# Test Timeouts (milliseconds)
-TEST_TIMEOUT=60000
-TEST_WAIT_TIMEOUT=15000

e2e/README.md

@@ -1,376 +0,0 @@
-# E2E Testing with WebdriverIO
-
-End-to-end tests for JellyTau using WebdriverIO and tauri-driver. These tests run against a real Tauri app instance with an **isolated test database**.
-
-## Quick Start
-
-```bash
-# 1. Configure test credentials (first time only)
-cp e2e/.env.example e2e/.env
-# Edit e2e/.env with your Jellyfin server details
-
-# 2. Build the frontend
-bun run build
-
-# 3. Run E2E tests
-bun run test:e2e
-```
-
-## Configuration
-
-### Test Credentials
-
-E2E tests use credentials from `e2e/.env` (gitignored). Copy the example file to get started:
-
-```bash
-cp e2e/.env.example e2e/.env
-```
-
-**e2e/.env** (your private file):
-```bash
-# Your Jellyfin test server
-TEST_SERVER_URL=https://your-jellyfin.example.com
-TEST_SERVER_NAME=My Test Server
-
-# Test user credentials
-TEST_USERNAME=testuser
-TEST_PASSWORD=yourpassword
-
-# Optional: Specific test data IDs
-TEST_MUSIC_LIBRARY_ID=abc123
-TEST_ALBUM_ID=xyz789
-# ... etc
-```
-
-**Important:**
-- ✅ `.env` is gitignored - your credentials stay private
-- ✅ Tests fall back to Jellyfin demo server if `.env` doesn't exist
-- ✅ Share `.env.example` with your team so they can set up their own
-
-### Isolated Test Database
-
-**Your production data is safe!** E2E tests use a completely separate database:
-
-- **Production:** `~/.local/share/com.dtourolle.jellytau/` - Your real data ✅
-- **E2E Tests:** `/tmp/jellytau-test-data/` - Isolated test data ✅
-
-This is configured via the `JELLYTAU_DATA_DIR` environment variable in `wdio.conf.ts`.
-
-## Architecture
-
-### Test Structure
-
-```
-e2e/
-├── .env.example           # Template for test credentials
-├── .env                   # Your credentials (gitignored)
-├── specs/                 # Test specifications
-│   ├── app-launch.e2e.ts  # App initialization tests
-│   ├── auth.e2e.ts        # Authentication flow
-│   └── navigation.e2e.ts  # Navigation and routing
-├── pageobjects/           # Page Object Model (POM)
-│   ├── BasePage.ts        # Base class with common methods
-│   ├── LoginPage.ts       # Login page interactions
-│   └── HomePage.ts        # Home page interactions
-└── helpers/               # Test utilities
-    ├── testConfig.ts      # Load .env configuration
-    └── testSetup.ts       # Setup helpers
-```
-
-### Page Object Model
-
-Tests use the Page Object Model pattern for maintainability:
-
-```typescript
-// Good: Using page objects
-import LoginPage from "../pageobjects/LoginPage";
-
-await LoginPage.waitForLoginPage();
-await LoginPage.connectToServer(testConfig.serverUrl);
-await LoginPage.login(testConfig.username, testConfig.password);
-
-// Bad: Direct selectors in tests
-await $("#server-url").setValue("https://...");
-await $("button").click();
-```
-
-## Writing Tests
-
-### Using Test Configuration
-
-Always use `testConfig` for credentials and server details:
-
-```typescript
-import { testConfig } from "../helpers/testConfig";
-
-describe("My Feature", () => {
-  it("should test something", async () => {
-    // Use testConfig instead of hardcoded values
-    await LoginPage.connectToServer(testConfig.serverUrl);
-    await LoginPage.login(testConfig.username, testConfig.password);
-
-    // Access optional test data
-    if (testConfig.albumId) {
-      // Test with specific album
-    }
-  });
-});
-```
-
-### Test Data IDs
-
-For tests that need specific content (albums, tracks, etc.):
-
-1. Find the ID in your Jellyfin server (check the URL when viewing an item)
-2. Add it to your `e2e/.env`:
-   ```bash
-   TEST_ALBUM_ID=abc123def456
-   ```
-3. Use it in tests:
-   ```typescript
-   if (testConfig.albumId) {
-     await browser.url(`/album/${testConfig.albumId}`);
-   }
-   ```
-
-### Example Test
-
-```typescript
-import { expect } from "@wdio/globals";
-import LoginPage from "../pageobjects/LoginPage";
-import { testConfig } from "../helpers/testConfig";
-
-describe("Album Playback", () => {
-  beforeEach(async () => {
-    // Login before each test
-    await LoginPage.waitForLoginPage();
-    await LoginPage.fullLoginFlow(
-      testConfig.serverUrl,
-      testConfig.username,
-      testConfig.password
-    );
-  });
-
-  it("should play an album", async () => {
-    // Skip if no test album configured
-    if (!testConfig.albumId) {
-      console.log("Skipping - no TEST_ALBUM_ID configured");
-      return;
-    }
-
-    // Navigate to album
-    await browser.url(`/album/${testConfig.albumId}`);
-
-    // Click play
-    const playButton = await $('[aria-label="Play"]');
-    await playButton.click();
-
-    // Verify playback started
-    const miniPlayer = await $(".mini-player");
-    expect(await miniPlayer.isDisplayed()).toBe(true);
-  });
-});
-```
-
-## Running Tests
-
-### Commands
-
-```bash
-# Run all E2E tests
-bun run test:e2e
-
-# Run in watch mode (development)
-bun run test:e2e:dev
-
-# Run specific test file
-bun run test:e2e -- e2e/specs/auth.e2e.ts
-```
-
-### Before Running
-
-**Always build the frontend first:**
-
-```bash
-bun run build
-cd src-tauri && cargo build
-```
-
-The debug binary expects built frontend files in the `build/` directory.
-
-## Test Files
-
-### app-launch.e2e.ts
-Basic app initialization tests:
-- App launches successfully
-- UI renders correctly
-- Unauthenticated users redirect to login
-
-**Status:** ✅ Working (no credentials needed)
-
-### auth.e2e.ts
-Full authentication flow:
-- Server connection (2-step process)
-- Login form validation
-- Error handling
-- Complete auth flow
-
-**Status:** ✅ Working with any Jellyfin server
-
-### navigation.e2e.ts
-Routing and navigation:
-- Protected routes
-- Redirects
-- Navigation after login
-
-**Status:** ⚠️ Needs valid credentials (configure `.env`)
-
-## Configuration Reference
-
-### wdio.conf.ts
-
-Main WebdriverIO configuration:
-
-```typescript
-{
-  port: 4444,                    // tauri-driver port
-  maxInstances: 1,               // Run tests sequentially
-  logLevel: "warn",              // Reduce noise
-  framework: "mocha",
-  timeout: 60000,                // 60s test timeout
-
-  capabilities: [{
-    "tauri:options": {
-      application: "path/to/app",
-      env: {
-        JELLYTAU_DATA_DIR: "/tmp/jellytau-test-data"  // Isolated DB
-      }
-    }
-  }]
-}
-```
-
-### Environment Variables
-
-| Variable | Description | Default |
-|----------|-------------|---------|
-| `TEST_SERVER_URL` | Jellyfin server URL | `https://demo.jellyfin.org/stable` |
-| `TEST_SERVER_NAME` | Server display name | `Demo Server` |
-| `TEST_USERNAME` | Test user username | `demo` |
-| `TEST_PASSWORD` | Test user password | `` (empty) |
-| `TEST_MUSIC_LIBRARY_ID` | Music library ID | undefined |
-| `TEST_ALBUM_ID` | Album ID for playback tests | undefined |
-| `TEST_TRACK_ID` | Track ID for tests | undefined |
-| `TEST_TIMEOUT` | Mocha test timeout (ms) | `60000` |
-| `TEST_WAIT_TIMEOUT` | Element wait timeout (ms) | `15000` |
-
-## Debugging
-
-### View Application During Tests
-
-Tests run with a visible window. To pause and inspect:
-
-```typescript
-it("debug test", async () => {
-  await LoginPage.waitForLoginPage();
-
-  // Pause for 10 seconds to inspect
-  await browser.pause(10000);
-
-  await LoginPage.enterServerUrl(testConfig.serverUrl);
-});
-```
-
-### Check Logs
-
-- **WebdriverIO logs:** Console output (set `logLevel: "info"` in config)
-- **tauri-driver logs:** Stdout/stderr from driver process
-- **App logs:** Check app console (if running with dev tools)
-
-### Common Issues
-
-**"Connection refused" in browser body**
-- Frontend not built: Run `bun run build`
-- Solution: Always build before testing
-
-**"Element not found" errors**
-- Selector might be wrong
-- Element not loaded yet - add wait: `await element.waitForDisplayed()`
-
-**"Invalid session id"**
-- Normal when app closes between tests
-- Each test file gets a fresh app instance
-
-**Tests fail with "no .env file"**
-- Copy `e2e/.env.example` to `e2e/.env`
-- Configure your Jellyfin server details
-
-**Database still using production data**
-- Check `wdio.conf.ts` has `JELLYTAU_DATA_DIR` env var
-- Rebuild app: `cd src-tauri && cargo build`
-
-## Platform Support
-
-### Supported
-
-- ✅ **Linux** - Primary development platform
-- ✅ **Windows** - Supported (paths auto-detected)
-- ✅ **macOS** - Supported (paths auto-detected)
-
-### Not Supported
-
-- ❌ **Android** - E2E testing requires Appium + emulators (out of scope)
-- Desktop tests cover 90% of app logic anyway
-
-## Team Collaboration
-
-### Sharing Test Configuration
-
-**DO:**
-- ✅ Commit `e2e/.env.example` with template values
-- ✅ Update README when adding new test data requirements
-- ✅ Use descriptive variable names in `.env.example`
-
-**DON'T:**
-- ❌ Commit `e2e/.env` with real credentials
-- ❌ Hardcode server URLs in test files
-- ❌ Skip authentication in tests (always test full flows)
-
-### Setting Up for a New Team Member
-
-1. **Clone repo**
-2. **Copy env template:** `cp e2e/.env.example e2e/.env`
-3. **Configure credentials:** Edit `e2e/.env` with your Jellyfin server
-4. **Build frontend:** `bun run build`
-5. **Run tests:** `bun run test:e2e`
-
-That's it! No shared credentials needed.
-
-## Best Practices
-
-1. **Use testConfig:** Never hardcode credentials
-2. **Use Page Objects:** Keep selectors out of test specs
-3. **Wait for Elements:** Always use `.waitForDisplayed()`
-4. **Independent Tests:** Each test should work standalone
-5. **Skip Gracefully:** Check for optional test data before using
-6. **Build First:** Always `bun run build` before running tests
-7. **Clear Names:** Use descriptive `describe` and `it` blocks
-
-## Future Enhancements
-
-- [ ] Add more page objects (Player, Library, Queue, Settings)
-- [ ] Create test data fixtures
-- [ ] Add visual regression testing
-- [ ] Mock Jellyfin API for faster, more reliable tests
-- [ ] CI/CD integration (GitHub Actions)
-- [ ] Test report generation
-- [ ] Screenshot capture on failure
-- [ ] Video recording of test runs
-
-## Resources
-
-- [WebdriverIO Documentation](https://webdriver.io/)
-- [Tauri Testing Guide](https://v2.tauri.app/develop/tests/webdriver/)
-- [tauri-driver GitHub](https://github.com/tauri-apps/tauri/tree/dev/tooling/webdriver)
-- [Mocha Documentation](https://mochajs.org/)
-- [Page Object Model Pattern](https://webdriver.io/docs/pageobjects/)

e2e/helpers/testConfig.ts

@@ -1,105 +0,0 @@
-import fs from "node:fs";
-import path from "node:path";
-
-/**
- * Test configuration loaded from .env file
- */
-export interface TestConfig {
-  serverUrl: string;
-  serverName: string;
-  username: string;
-  password: string;
-  musicLibraryId?: string;
-  movieLibraryId?: string;
-  artistId?: string;
-  albumId?: string;
-  trackId?: string;
-  movieId?: string;
-  episodeId?: string;
-  timeout: number;
-  waitTimeout: number;
-}
-
-/**
- * Load test configuration from .env file
- * Falls back to demo server if .env doesn't exist
- */
-export function loadTestConfig(): TestConfig {
-  const envPath = path.join(__dirname, "..", ".env");
-  const config: TestConfig = {
-    serverUrl: "https://demo.jellyfin.org/stable",
-    serverName: "Demo Server",
-    username: "demo",
-    password: "",
-    timeout: 60000,
-    waitTimeout: 15000,
-  };
-
-  // Try to load .env file
-  if (fs.existsSync(envPath)) {
-    const envContent = fs.readFileSync(envPath, "utf-8");
-    const lines = envContent.split("\n");
-
-    for (const line of lines) {
-      // Skip comments and empty lines
-      if (line.trim().startsWith("#") || !line.trim()) continue;
-
-      const [key, ...valueParts] = line.split("=");
-      const value = valueParts.join("=").trim();
-
-      switch (key.trim()) {
-        case "TEST_SERVER_URL":
-          if (value) config.serverUrl = value;
-          break;
-        case "TEST_SERVER_NAME":
-          if (value) config.serverName = value;
-          break;
-        case "TEST_USERNAME":
-          if (value) config.username = value;
-          break;
-        case "TEST_PASSWORD":
-          config.password = value; // Can be empty
-          break;
-        case "TEST_MUSIC_LIBRARY_ID":
-          if (value) config.musicLibraryId = value;
-          break;
-        case "TEST_MOVIE_LIBRARY_ID":
-          if (value) config.movieLibraryId = value;
-          break;
-        case "TEST_ARTIST_ID":
-          if (value) config.artistId = value;
-          break;
-        case "TEST_ALBUM_ID":
-          if (value) config.albumId = value;
-          break;
-        case "TEST_TRACK_ID":
-          if (value) config.trackId = value;
-          break;
-        case "TEST_MOVIE_ID":
-          if (value) config.movieId = value;
-          break;
-        case "TEST_EPISODE_ID":
-          if (value) config.episodeId = value;
-          break;
-        case "TEST_TIMEOUT":
-          if (value) config.timeout = parseInt(value, 10);
-          break;
-        case "TEST_WAIT_TIMEOUT":
-          if (value) config.waitTimeout = parseInt(value, 10);
-          break;
-      }
-    }
-  } else {
-    console.warn(
-      "⚠️  No e2e/.env file found. Using demo server credentials."
-    );
-    console.warn(
-      "   Copy e2e/.env.example to e2e/.env and configure your test server."
-    );
-  }
-
-  return config;
-}
-
-// Export a singleton instance
-export const testConfig = loadTestConfig();

e2e/helpers/testSetup.ts

@@ -1,53 +0,0 @@
-import fs from "node:fs";
-import path from "node:path";
-import os from "node:os";
-
-/**
- * Clears the JellyTau database and cache before tests
- * This ensures each test run starts with a fresh state
- */
-export function clearAppData() {
-  const appDataDir = path.join(
-    os.homedir(),
-    ".local/share/com.dtourolle.jellytau"
-  );
-
-  try {
-    if (fs.existsSync(appDataDir)) {
-      // Remove database file
-      const dbPath = path.join(appDataDir, "jellytau.db");
-      if (fs.existsSync(dbPath)) {
-        fs.unlinkSync(dbPath);
-        console.log("Cleared test database");
-      }
-
-      // Clear any cache files if needed
-      // Add more cleanup as needed
-    }
-  } catch (error) {
-    console.warn("Failed to clear app data:", error);
-    // Don't fail tests if cleanup fails
-  }
-}
-
-/**
- * Wait for element with retries
- * Useful for elements that might take time to appear
- */
-export async function waitForElement(
-  selector: string,
-  timeout: number = 15000,
-  retries: number = 3
-): Promise<WebdriverIO.Element> {
-  for (let i = 0; i < retries; i++) {
-    try {
-      const element = await $(selector);
-      await element.waitForDisplayed({ timeout });
-      return element;
-    } catch (error) {
-      if (i === retries - 1) throw error;
-      await browser.pause(1000);
-    }
-  }
-  throw new Error(`Element ${selector} not found after ${retries} retries`);
-}

e2e/pageobjects/BasePage.ts

@@ -1,31 +0,0 @@
-export default class BasePage {
-  async waitForElement(selector: string, timeout: number = 10000) {
-    const element = await $(selector);
-    await element.waitForDisplayed({ timeout });
-    return element;
-  }
-
-  async clickElement(selector: string) {
-    const element = await this.waitForElement(selector);
-    await element.click();
-  }
-
-  async enterText(selector: string, text: string) {
-    const element = await this.waitForElement(selector);
-    await element.setValue(text);
-  }
-
-  async getText(selector: string): Promise<string> {
-    const element = await this.waitForElement(selector);
-    return await element.getText();
-  }
-
-  async isElementDisplayed(selector: string): Promise<boolean> {
-    try {
-      const element = await $(selector);
-      return await element.isDisplayed();
-    } catch (error) {
-      return false;
-    }
-  }
-}

e2e/pageobjects/HomePage.ts

@@ -1,55 +0,0 @@
-import BasePage from "./BasePage";
-
-class HomePage extends BasePage {
-  // Selectors
-  get loadingSpinner() {
-    return $(".animate-spin");
-  }
-
-  get browseLibrariesButton() {
-    return $("button*=Browse all libraries");
-  }
-
-  get offlineBanner() {
-    return $(".bg-amber-600\\/90");
-  }
-
-  // Carousel sections
-  get heroSection() {
-    return $("div"); // Hero banner would need specific selector
-  }
-
-  // Actions
-  async waitForHomePageLoad(timeout: number = 15000) {
-    // Wait for loading spinner to disappear
-    try {
-      await this.loadingSpinner.waitForDisplayed({ timeout: 5000 });
-      await this.loadingSpinner.waitForDisplayed({ timeout, reverse: true });
-    } catch {
-      // Spinner might not appear if page loads quickly
-    }
-  }
-
-  async isOffline(): Promise<boolean> {
-    try {
-      return await this.offlineBanner.isDisplayed();
-    } catch {
-      return false;
-    }
-  }
-
-  async clickBrowseLibraries() {
-    await this.browseLibrariesButton.click();
-  }
-
-  async hasContent(): Promise<boolean> {
-    // Check if browse button exists (indicates loaded state)
-    try {
-      return await this.browseLibrariesButton.isExisting();
-    } catch {
-      return false;
-    }
-  }
-}
-
-export default new HomePage();

e2e/pageobjects/LoginPage.ts

@@ -1,116 +0,0 @@
-import BasePage from "./BasePage";
-
-class LoginPage extends BasePage {
-  // Selectors
-  get pageTitle() {
-    return $("h1");
-  }
-
-  get serverUrlInput() {
-    return $("#server-url");
-  }
-
-  get connectButton() {
-    return $('button[type="submit"]');
-  }
-
-  get usernameInput() {
-    return $("#username");
-  }
-
-  get passwordInput() {
-    return $("#password");
-  }
-
-  get signInButton() {
-    return $('button[type="submit"]');
-  }
-
-  get errorMessage() {
-    return $(".bg-red-900\\/50");
-  }
-
-  get backButton() {
-    return $("button*=Back");
-  }
-
-  get serverNameDisplay() {
-    return $('p.text-\\[var\\(--color-jellyfin\\)\\]');
-  }
-
-  // Actions
-  async waitForLoginPage(timeout: number = 10000) {
-    await this.serverUrlInput.waitForDisplayed({ timeout });
-  }
-
-  async enterServerUrl(url: string) {
-    await this.serverUrlInput.setValue(url);
-  }
-
-  async clickConnect() {
-    await this.connectButton.click();
-  }
-
-  async connectToServer(url: string) {
-    await this.enterServerUrl(url);
-    await this.clickConnect();
-
-    // Wait for transition to login form
-    await this.usernameInput.waitForDisplayed({ timeout: 10000 });
-  }
-
-  async enterUsername(username: string) {
-    await this.usernameInput.setValue(username);
-  }
-
-  async enterPassword(password: string) {
-    await this.passwordInput.setValue(password);
-  }
-
-  async clickSignIn() {
-    await this.signInButton.click();
-  }
-
-  async login(username: string, password: string) {
-    await this.enterUsername(username);
-    await this.enterPassword(password);
-    await this.clickSignIn();
-  }
-
-  async fullLoginFlow(serverUrl: string, username: string, password: string) {
-    await this.waitForLoginPage();
-    await this.connectToServer(serverUrl);
-    await this.login(username, password);
-  }
-
-  async isOnServerStep(): Promise<boolean> {
-    try {
-      return await this.serverUrlInput.isDisplayed();
-    } catch {
-      return false;
-    }
-  }
-
-  async isOnLoginStep(): Promise<boolean> {
-    try {
-      return await this.usernameInput.isDisplayed();
-    } catch {
-      return false;
-    }
-  }
-
-  async getErrorMessage(): Promise<string> {
-    await this.errorMessage.waitForDisplayed({ timeout: 5000 });
-    return await this.errorMessage.getText();
-  }
-
-  async hasError(): Promise<boolean> {
-    try {
-      return await this.errorMessage.isDisplayed();
-    } catch {
-      return false;
-    }
-  }
-}
-
-export default new LoginPage();

e2e/specs/app-launch.e2e.ts

@@ -1,39 +0,0 @@
-import { expect } from "@wdio/globals";
-
-describe("Application Launch", () => {
-  it("should launch the application", async () => {
-    // Wait for body element to appear
-    const body = await $("body");
-    await body.waitForDisplayed({ timeout: 15000 });
-
-    // Verify app launched successfully
-    expect(await body.isDisplayed()).toBe(true);
-  });
-
-  it("should render the main app container", async () => {
-    // The app has a root div with specific classes
-    const appContainer = await $("div.h-screen.bg-\\[var\\(--color-background\\)\\]");
-
-    // Verify the main container exists
-    expect(await appContainer.isExisting()).toBe(true);
-    expect(await appContainer.isDisplayed()).toBe(true);
-  });
-
-  it("should show JellyTau branding", async () => {
-    // The app should show JellyTau title on login page (default state)
-    const title = await $("h1");
-    await title.waitForDisplayed({ timeout: 10000 });
-
-    const titleText = await title.getText();
-    expect(titleText).toContain("JellyTau");
-  });
-
-  it("should redirect unauthenticated users to login", async () => {
-    // Wait for login page elements to appear
-    const serverUrlInput = await $("#server-url");
-    await serverUrlInput.waitForDisplayed({ timeout: 10000 });
-
-    // Verify we're on the login page
-    expect(await serverUrlInput.isDisplayed()).toBe(true);
-  });
-});

e2e/specs/auth.e2e.ts

@@ -1,145 +0,0 @@
-import { expect } from "@wdio/globals";
-import LoginPage from "../pageobjects/LoginPage";
-import { testConfig } from "../helpers/testConfig";
-
-describe("Authentication Flow", () => {
-  beforeEach(async () => {
-    // Each test starts fresh - app should redirect to login
-    await LoginPage.waitForLoginPage();
-  });
-
-  describe("Server Connection", () => {
-    it("should display the server connection form", async () => {
-      expect(await LoginPage.isOnServerStep()).toBe(true);
-      expect(await LoginPage.pageTitle.getText()).toContain("JellyTau");
-    });
-
-    it("should show server URL input field", async () => {
-      const serverInput = await LoginPage.serverUrlInput;
-
-      expect(await serverInput.isDisplayed()).toBe(true);
-      expect(await serverInput.getAttribute("placeholder")).toContain("jellyfin");
-    });
-
-    it("should have a disabled connect button when URL is empty", async () => {
-      const connectButton = await LoginPage.connectButton;
-
-      // Button should be disabled when input is empty
-      expect(await connectButton.isEnabled()).toBe(false);
-    });
-
-    it("should enable connect button when URL is entered", async () => {
-      await LoginPage.enterServerUrl(testConfig.serverUrl);
-
-      const connectButton = await LoginPage.connectButton;
-      expect(await connectButton.isEnabled()).toBe(true);
-    });
-
-    it("should show error for invalid server URL", async () => {
-      await LoginPage.enterServerUrl("not-a-valid-url");
-      await LoginPage.clickConnect();
-
-      // Wait for error to appear
-      await browser.pause(2000);
-
-      expect(await LoginPage.hasError()).toBe(true);
-    });
-
-    it("should transition to login form on successful connection", async () => {
-      // Using configured test server
-      await LoginPage.connectToServer(testConfig.serverUrl);
-
-      // Should now be on login step
-      expect(await LoginPage.isOnLoginStep()).toBe(true);
-      expect(await LoginPage.isOnServerStep()).toBe(false);
-    });
-  });
-
-  describe("User Login", () => {
-    beforeEach(async () => {
-      // Connect to configured test server before each login test
-      await LoginPage.connectToServer(testConfig.serverUrl);
-    });
-
-    it("should display login form after server connection", async () => {
-      expect(await LoginPage.usernameInput.isDisplayed()).toBe(true);
-      expect(await LoginPage.passwordInput.isDisplayed()).toBe(true);
-      expect(await LoginPage.signInButton.isDisplayed()).toBe(true);
-    });
-
-    it("should show server information", async () => {
-      // Server name and URL should be displayed
-      const serverName = await LoginPage.serverNameDisplay;
-      expect(await serverName.isDisplayed()).toBe(true);
-    });
-
-    it("should have back button to return to server selection", async () => {
-      expect(await LoginPage.backButton.isDisplayed()).toBe(true);
-
-      await LoginPage.backButton.click();
-      await browser.pause(500);
-
-      // Should be back on server step
-      expect(await LoginPage.isOnServerStep()).toBe(true);
-    });
-
-    it("should disable sign in button when username is empty", async () => {
-      const signInButton = await LoginPage.signInButton;
-      expect(await signInButton.isEnabled()).toBe(false);
-    });
-
-    it("should enable sign in button when username is entered", async () => {
-      await LoginPage.enterUsername("demo");
-
-      const signInButton = await LoginPage.signInButton;
-      expect(await signInButton.isEnabled()).toBe(true);
-    });
-
-    it("should show error for invalid credentials", async () => {
-      await LoginPage.login("invalid-user", "wrong-password");
-
-      // Wait for error
-      await browser.pause(2000);
-
-      expect(await LoginPage.hasError()).toBe(true);
-    });
-
-    // Enable this test by configuring e2e/.env with valid credentials
-    it.skip("should successfully login with valid credentials", async () => {
-      await LoginPage.login(testConfig.username, testConfig.password);
-
-      // Wait for redirect to home page
-      await browser.pause(3000);
-
-      // Should redirect away from login page
-      const currentUrl = await browser.getUrl();
-      expect(currentUrl).not.toContain("/login");
-    });
-  });
-
-  describe("Full Authentication Flow", () => {
-    it("should complete full auth flow with test server", async () => {
-      // Test the complete flow
-      await LoginPage.waitForLoginPage();
-
-      // Step 1: Enter server URL
-      expect(await LoginPage.isOnServerStep()).toBe(true);
-      await LoginPage.enterServerUrl(testConfig.serverUrl);
-      await LoginPage.clickConnect();
-
-      // Wait for transition
-      await browser.pause(2000);
-
-      // Step 2: Should be on login form
-      expect(await LoginPage.isOnLoginStep()).toBe(true);
-
-      // Step 3: Enter credentials
-      await LoginPage.enterUsername(testConfig.username);
-      await LoginPage.enterPassword(testConfig.password);
-
-      // Verify form is filled
-      const username = await LoginPage.usernameInput.getValue();
-      expect(username).toBe(testConfig.username);
-    });
-  });
-});

e2e/specs/navigation.e2e.ts

@@ -1,39 +0,0 @@
-import { expect } from "@wdio/globals";
-import LoginPage from "../pageobjects/LoginPage";
-import HomePage from "../pageobjects/HomePage";
-import { testConfig } from "../helpers/testConfig";
-
-describe("Navigation", () => {
-  it("should redirect unauthenticated users to login", async () => {
-    // App should automatically redirect to login when not authenticated
-    await LoginPage.waitForLoginPage();
-
-    expect(await LoginPage.isOnServerStep()).toBe(true);
-  });
-
-  it("should prevent direct access to protected routes", async () => {
-    // Try to navigate to a protected route
-    await browser.url("http://localhost:4444/session/fake-session-id/url");
-    await browser.pause(1000);
-
-    // Should redirect back to login
-    await LoginPage.waitForLoginPage(5000);
-    expect(await LoginPage.isOnServerStep()).toBe(true);
-  });
-
-  // This test requires valid authentication - configure e2e/.env to enable
-  it.skip("should allow navigation after login", async () => {
-    // Login first
-    await LoginPage.fullLoginFlow(
-      testConfig.serverUrl,
-      testConfig.username,
-      testConfig.password
-    );
-
-    // Wait for home page
-    await HomePage.waitForHomePageLoad();
-
-    // Should be able to navigate
-    expect(await HomePage.hasContent()).toBe(true);
-  });
-});

eslint.config.js

@@ -0,0 +1,199 @@
+// ESLint flat config for the JellyTau frontend (Svelte 5 + TypeScript strict).
+//
+// TRACES: | DR-205
+//
+// Scope: `src/` (the presentation layer), `scripts/` (build tooling), and the
+// root config files. The Rust backend is linted by clippy, not by this config.
+//
+// Formatting is NOT ESLint's job here — `eslint-config-prettier` is applied last
+// and switches off every stylistic rule that would fight `prettier`. Run
+// `bun run format` / `bun run format:check` for layout.
+import js from "@eslint/js";
+import ts from "typescript-eslint";
+import svelte from "eslint-plugin-svelte";
+import globals from "globals";
+import prettier from "eslint-config-prettier";
+import svelteConfig from "./svelte.config.js";
+
+export default ts.config(
+  {
+    // Kept in one place so `npx eslint .` and editor integrations agree.
+    ignores: [
+      "node_modules/",
+      ".svelte-kit/",
+      // Scratch worktrees (git-ignored) hold full checkouts of this repo,
+      // including their own generated .svelte-kit trees. Without this, `eslint .`
+      // lints every in-flight branch and reports its generated code as ours.
+      ".claude/",
+      "build/",
+      "dist/",
+      "coverage/",
+      "package/",
+      "src-tauri/",
+      // Generated by tauri-specta on every Rust build — never hand-edited, and
+      // its shape is dictated by the Rust command definitions.
+      "src/lib/api/bindings.ts",
+    ],
+  },
+
+  js.configs.recommended,
+  ...ts.configs.recommended,
+  ...svelte.configs.recommended,
+  prettier,
+  ...svelte.configs.prettier,
+
+  {
+    languageOptions: {
+      globals: {
+        ...globals.browser,
+        ...globals.es2021,
+      },
+    },
+    rules: {
+      // The logger-facade migration this rule was waiting on is done: the ~468
+      // `console.*` calls that used to live in `src/` are gone, replaced by
+      // `createLogger(...)` from src/lib/utils/logger.ts (DR-204), which is now
+      // the single sink. Nothing is allowed through — not even warn/error —
+      // because the facade's own `warn`/`error` levels are always emitted, so a
+      // raw call has no capability the facade lacks. It only loses the scope tag
+      // and the runtime level control.
+      //
+      // The sink itself is exempted below, as are tests (a test that asserts on
+      // logging has to be able to talk about `console`).
+      "no-console": "error",
+
+      // Unused values are a real signal, but `_`-prefixed args are the
+      // established way to say "this parameter exists for the signature".
+      //
+      // ⚠️ warn, not error: the tree carries ~94 genuinely dead bindings (stale
+      // imports, `$state` left over from refactors, unused `catch (e)`). Every
+      // one is a real finding, but fixing them here would mean ~50 unrelated
+      // files in this tooling commit. Clear the backlog, then promote to
+      // "error".
+      "@typescript-eslint/no-unused-vars": [
+        "warn",
+        {
+          argsIgnorePattern: "^_",
+          varsIgnorePattern: "^_",
+          caughtErrorsIgnorePattern: "^_",
+          destructuredArrayIgnorePattern: "^_",
+        },
+      ],
+
+      // Warn-only rules: each flags something real, but the existing tree has
+      // more instances than can be fixed without swamping unrelated diffs.
+      // Drive these to zero and promote them to "error" — do not delete them.
+      //
+      // `any` at the Tauri IPC boundary, mostly in code predating the
+      // tauri-specta bindings (~25 sites outside tests).
+      "@typescript-eslint/no-explicit-any": "warn",
+      // Empty catch/if bodies that swallow an error.
+      "no-empty": ["warn", { allowEmptyCatch: true }],
+
+      // Prefer `import type` so type-only imports are erased cleanly by the
+      // bundler instead of pulling a module in at run time.
+      "@typescript-eslint/consistent-type-imports": "off",
+
+      // Not applicable to this app (~130 hits, all no-ops). SvelteKit's
+      // `resolve()` exists so hrefs keep working under a non-empty
+      // `kit.paths.base`; JellyTau is an adapter-static SPA served from the
+      // Tauri webview root and svelte.config.js sets no `base`. Re-enable this
+      // the day a base path is introduced — the rule is otherwise correct.
+      // (Declared here, not in the *.svelte block: `goto()` is also called from
+      // plain .ts modules such as src/lib/utils/navigation.ts.)
+      "svelte/no-navigation-without-resolve": "off",
+    },
+  },
+
+  {
+    // Svelte components: the parser needs the project's svelte.config.js so it
+    // resolves preprocessors and Svelte 5 runes the same way the build does.
+    files: ["**/*.svelte", "**/*.svelte.ts", "**/*.svelte.js"],
+    languageOptions: {
+      parserOptions: {
+        parser: ts.parser,
+        svelteConfig,
+      },
+    },
+    rules: {
+      // Warn-only — real findings, but each fix is a behavioural refactor that
+      // does not belong in a tooling commit:
+      //   require-each-key      keyed {#each} changes DOM reuse semantics
+      //   prefer-svelte-reactivity  Set/Map -> SvelteSet/SvelteMap changes
+      //                             reactivity, not just syntax
+      //   prefer-writable-derived   $state + $effect -> writable $derived
+      //   no-at-html-tags           {@html} sites need an XSS review each
+      "svelte/require-each-key": "warn",
+      "svelte/prefer-svelte-reactivity": "warn",
+      "svelte/prefer-writable-derived": "warn",
+      "svelte/no-at-html-tags": "warn",
+
+      // Warn-only: this rule cannot see the Svelte *compiler's* warning set, so
+      // it reports `<!-- svelte-ignore a11y_… -->` as unused when the compiler
+      // may still be emitting the warning it suppresses. Verify against a real
+      // `bun run check` before deleting any of them.
+      "svelte/no-unused-svelte-ignore": "warn",
+    },
+  },
+
+  {
+    // The logging facade is the one place allowed to touch `console` — it *is*
+    // the sink every other module reaches it through (see the `no-console`
+    // comment above). `createLogger`'s `console[method](...)` dispatch is a
+    // computed member access, which the rule flags like any other.
+    files: ["src/lib/utils/logger.ts"],
+    rules: {
+      "no-console": "off",
+    },
+  },
+
+  {
+    // Node-side tooling: build/test scripts and root config files run under
+    // Bun/Node, not in the webview.
+    files: [
+      "scripts/**/*.{ts,js}",
+      "*.config.{ts,js}",
+      "*.config.*.{ts,js}",
+      "svelte.config.js",
+      "eslint.config.js",
+    ],
+    languageOptions: {
+      globals: {
+        ...globals.node,
+      },
+    },
+    rules: {
+      // These are command-line tools (extract-traces, release-notes, ...) whose
+      // stdout IS the product — `bun run traces:markdown > docs/traceability.md`
+      // depends on it. The logging facade is a webview concern; a CLI printing
+      // its result is not a stray debug statement.
+      "no-console": "off",
+    },
+  },
+
+  {
+    // Test files: vitest globals are enabled in vitest.config.ts.
+    files: ["**/*.{test,spec}.{ts,js}", "src/test/**/*.{ts,js}"],
+    languageOptions: {
+      globals: {
+        ...globals.node,
+        ...globals.vitest,
+      },
+    },
+    rules: {
+      // Tests are allowed to talk about `console` — several spy on it to assert
+      // what the logging facade emits, and scripts/ tooling tests capture output.
+      "no-console": "off",
+      // Test doubles legitimately use `any` for partial mocks.
+      "@typescript-eslint/no-explicit-any": "off",
+      // `vi.mock` factories are hoisted above the import graph, so a lazy
+      // `require()` inside one is the documented escape hatch.
+      "@typescript-eslint/no-require-imports": "off",
+      // Several tests deliberately replay a production assignment sequence
+      // (`currentStreamUrl = newStreamUrl; hasSeeked = false;`) to document the
+      // `$effect` they stand in for. The "useless" write is the subject under
+      // test, not dead code.
+      "no-useless-assignment": "off",
+    },
+  },
+);

package.json

@@ -1,7 +1,14 @@
 {
   "name": "jellytau",
-  "version": "0.1.1",
-  "description": "",
+  "version": "0.10.1",
+  "description": "A cross-platform Jellyfin client built with Tauri, SvelteKit and Rust.",
+  "author": "Duncan Tourolle <duncan@tourolle.paris>",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "https://gitea.tourolle.paris/dtourolle/jellytau"
+  },
+  "private": true,
   "type": "module",
   "packageManager": "bun@1.3.5",
   "scripts": {
@@ -10,16 +17,24 @@
     "preview": "vite preview",
     "check": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json",
     "check:watch": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json --watch",
-    "test": "vitest",
+    "test": "vitest run",
+    "test:watch": "vitest",
     "test:ui": "vitest --ui",
-    "test:coverage": "vitest --coverage",
-    "test:e2e": "wdio run ./wdio.conf.ts",
-    "test:e2e:dev": "wdio run ./wdio.conf.ts --watch",
+    "test:coverage": "vitest run --coverage",
     "test:all": "./scripts/test-all.sh",
     "test:rust": "./scripts/test-rust.sh",
+    "lint": "eslint .",
+    "lint:fix": "eslint . --fix",
+    "format": "prettier --write .",
+    "format:check": "prettier --check .",
     "check:boundary": "bash scripts/check-frontend-boundary.sh",
+    "check:links": "bash scripts/check-doc-links.sh",
+    "check:tooling": "bash scripts/check-tooling.sh",
+    "hooks:install": "./scripts/install-hooks.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",
@@ -36,37 +51,44 @@
     "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",
+    "traces:validate": "bun run scripts/extract-traces.ts --format validate",
     "release:notes": "bun run scripts/release-notes.ts"
   },
-  "license": "MIT",
   "dependencies": {
-    "@tauri-apps/api": "^2",
-    "@tauri-apps/plugin-opener": "^2",
+    "@tauri-apps/api": "^2.11.1",
+    "@tauri-apps/plugin-log": "2.9.0",
+    "@tauri-apps/plugin-opener": "^2.5.4",
     "@tauri-apps/plugin-os": "^2.3.2",
+    "@tauri-apps/plugin-process": "^2.3.1",
+    "@tauri-apps/plugin-updater": "2.10.1",
     "hls.js": "^1.6.15",
     "svelte-dnd-action": "^0.9.69"
   },
   "devDependencies": {
+    "@eslint/js": "^10.0.1",
     "@sveltejs/adapter-static": "^3.0.6",
     "@sveltejs/kit": "^2.9.0",
     "@sveltejs/vite-plugin-svelte": "^6.2.4",
     "@tailwindcss/vite": "^4.1.18",
-    "@tauri-apps/cli": "^2",
+    "@tauri-apps/cli": "^2.11.4",
     "@testing-library/svelte": "^5.3.1",
     "@vitest/coverage-v8": "^4.0.18",
     "@vitest/ui": "^4.0.16",
-    "@wdio/cli": "^9.5.0",
-    "@wdio/local-runner": "^9.5.0",
-    "@wdio/mocha-framework": "^9.5.0",
-    "@wdio/spec-reporter": "^9.5.0",
+    "eslint": "^10.8.1",
+    "eslint-config-prettier": "^10.1.8",
+    "eslint-plugin-svelte": "^3.23.0",
+    "globals": "^17.11.0",
     "happy-dom": "^20.0.11",
     "jsdom": "^27.4.0",
+    "prettier": "^3.9.6",
+    "prettier-plugin-svelte": "^4.1.1",
     "svelte": "^5.47.1",
     "svelte-check": "^4.0.0",
     "tailwindcss": "^4.1.18",
     "typescript": "~5.6.2",
+    "typescript-eslint": "^8.67.0",
     "vite": "^6.0.3",
-    "vitest": ">=1.0.0 <5.0.0",
-    "webdriverio": "^9.5.0"
+    "vitest": "^4.1.10"
   }
 }

packaging/arch/PKGBUILD

@@ -8,7 +8,7 @@
 # tarball/VCS URL and drop the local-copy prepare() step.
 
 pkgname=jellytau
-pkgver=0.0.18
+pkgver=0.10.1
 pkgrel=1
 pkgdesc="A cross-platform Jellyfin client"
 arch=('x86_64')
@@ -29,7 +29,45 @@ build() {
   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)
+  #
+  # 🔴 `tauri/custom-protocol` is not optional. `tauri build` passes it for you;
+  # a bare `cargo build` does not, and without it Tauri loads the frontend from
+  # `devUrl` rather than the assets embedded from `frontendDist`. The result
+  # builds and installs cleanly and then cannot load its own UI. check() guards
+  # this.
+  (cd src-tauri && cargo build --release --locked --features tauri/custom-protocol)
+}
+
+check() {
+  cd "$_srcdir"
+
+  # A Tauri binary built without `custom-protocol` does not embed the frontend;
+  # it serves it from `devUrl` (http://localhost:1420) instead. It compiles,
+  # links and installs perfectly, then launches into "Could not connect to
+  # localhost: Connection refused" — which is what this package did for its
+  # entire existence, because `tauri build` adds that feature for you and a bare
+  # `cargo build` does not.
+  #
+  # Test for the *assets*, not for the dev URL: `devUrl` is part of the config
+  # blob that generate_context!() embeds either way, so its presence proves
+  # nothing. A content-hashed filename from the vite build can only be in the
+  # binary if the bundle was embedded — the with-feature binary is ~400 KB
+  # larger for exactly this reason.
+  local _binary="src-tauri/target/release/jellytau"
+  local _asset
+  _asset="$(basename "$(ls -1 build/_app/immutable/entry/*.js | head -n1)")"
+
+  if [ -z "$_asset" ]; then
+    echo "==> ERROR: no frontend build found — 'bun run build' did not produce build/_app." >&2
+    return 1
+  fi
+
+  if ! grep -qa "$_asset" "$_binary"; then
+    echo "==> ERROR: the frontend bundle is not embedded in the binary." >&2
+    echo "    Build with --features tauri/custom-protocol, or the packaged app" >&2
+    echo "    will start up unable to load its own UI." >&2
+    return 1
+  fi
 }
 
 package() {
@@ -42,6 +80,11 @@ package() {
   install -Dm644 "packaging/arch/jellytau.desktop" \
     "$pkgdir/usr/share/applications/jellytau.desktop"
 
+  # MIT is not in /usr/share/licenses/common, so Arch packaging requires the
+  # licence text to ship with the package.
+  install -Dm644 "LICENSE" \
+    "$pkgdir/usr/share/licenses/$pkgname/LICENSE"
+
   # Icons (hicolor)
   install -Dm644 "src-tauri/icons/32x32.png" \
     "$pkgdir/usr/share/icons/hicolor/32x32/apps/jellytau.png"

scripts/README.md

@@ -13,11 +13,26 @@ Run all tests (frontend + Rust backend).
 ### `test-frontend.sh`
 Run frontend tests only.
 ```bash
-./scripts/test-frontend.sh          # Run all tests
+./scripts/test-frontend.sh          # Single pass (same as `bun run test`)
 ./scripts/test-frontend.sh --watch  # Watch mode
 ./scripts/test-frontend.sh --ui     # Open UI
 ```
 
+`bun run test` is `vitest run` — one pass, exit code, done. It used to be bare
+`vitest`, which parked in watch mode; CLAUDE.md's "Before Committing" list tells
+people to run it, so it had to terminate. The interactive modes moved to their
+own entry points:
+
+| Command | Runs |
+|---------|------|
+| `bun run test` | `vitest run` — single pass |
+| `bun run test:watch` | `vitest` — watch mode |
+| `bun run test:ui` | `vitest --ui` |
+| `bun run test:coverage` | `vitest run --coverage` |
+
+`test-frontend.sh` forwards any extra arguments to vitest and switches to the
+long-running form automatically when it sees `--watch`, `-w`, or `--ui`.
+
 ### `test-rust.sh`
 Run Rust tests only.
 ```bash
@@ -69,13 +84,37 @@ 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 the ratchet
+bun run traces:validate           # Dangling-ID gate — every traced ID must be defined
 ```
 
-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).
+
+**`bun run traces:validate` is the dangling-ID gate.** It fails if any traced ID
+— including `UT`/`IT`, which coverage deliberately ignores — is not defined as a
+table row in `requirements.md`, printing each offender with the files that
+reference it. Without it the extractor accepted any well-formed ID silently, so
+typos and renames that missed a call site went unreported for months.
+
+> **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
+> it reported `Total Requirements: 1` and then "All requirements have
+> implementations!". Nothing referenced it. Use `bun run traces:coverage`.
+
 Example TRACES comment in code:
 ```typescript
 // TRACES: UR-005, UR-026 | DR-029
@@ -88,7 +127,8 @@ See [docs/traceability.md](../docs/traceability.md) for the latest generated map
 
 The traceability system is integrated with Gitea Actions CI/CD:
 - Automatically validates TRACES on every push and pull request
-- Enforces minimum 50% coverage threshold
+- Enforces a minimum coverage threshold (a ratchet: raise it, never lower it)
+- Fails on dangling IDs — traced but undefined in `requirements.md`
 - Warns if new code lacks TRACES comments
 - Generates traceability reports automatically
 
@@ -96,6 +136,59 @@ For details, see:
 - [Traceability CI Guide](../docs/traceability-ci.md) - Full CI/CD documentation
 - [TRACES Quick Reference](../docs/traces-quick-ref.md) - Quick guide for adding TRACES
 
+## Linting & Formatting
+
+There is no script wrapper for these — they are plain package.json entries:
+
+```bash
+bun run lint          # eslint .
+bun run lint:fix      # eslint . --fix
+bun run format        # prettier --write .
+bun run format:check  # prettier --check .
+```
+
+Config lives in `eslint.config.js` (flat config: typescript-eslint +
+eslint-plugin-svelte, tuned for Svelte 5 and TS `strict`), `.prettierrc`, and
+`.prettierignore`. `src/lib/api/bindings.ts` is excluded from both — it is
+generated by tauri-specta on every Rust build.
+
+`bun run lint` is currently **error-clean but not warning-clean**: several rules
+are deliberately set to `warn` because the existing tree has more hits than a
+tooling change should touch (unused bindings, `any` at the IPC boundary, unkeyed
+`{#each}`). Each one is annotated in `eslint.config.js` with why, and the
+intended end state is `error`. Drive them down; do not delete them.
+
+`no-console` is switched **off** for now — see the note in `eslint.config.js`.
+
+## Git Hooks
+
+### `install-hooks.sh`
+Point git at the repo's tracked hooks directory (`core.hooksPath`).
+```bash
+bun run hooks:install      # or: ./scripts/install-hooks.sh
+```
+
+### `hooks/pre-commit`
+Runs the fast half of CLAUDE.md's "Before Committing" list so it is enforced
+rather than remembered:
+
+- `bun run check` (svelte-check)
+- `bun run test` (vitest, single pass)
+- `scripts/check-frontend-boundary.sh`
+- `cargo fmt --all -- --check`, **only when staged files touch `src-tauri/`**
+
+`cargo clippy` and `cargo test` are deliberately *not* in the hook — minutes per
+commit is how you teach people to reach for `--no-verify`. They run in CI, and
+locally via `bun run test:all`.
+
+```bash
+git commit --no-verify     # skip the hook for one commit
+git config --unset core.hooksPath   # uninstall
+```
+
+The hook skips itself during a merge, rebase, or cherry-pick, and when nothing
+is staged.
+
 ## Utility Scripts
 
 ### `clean.sh`
@@ -108,8 +201,12 @@ Clean all build artifacts.
 
 You can also run these via npm/bun:
 ```bash
+bun run test               # Frontend tests (single pass)
 bun run test:all           # All tests
 bun run test:rust          # Rust tests
+bun run lint               # ESLint
+bun run format:check       # Prettier (check only)
+bun run hooks:install      # Install the git hooks
 bun run android:build      # Build Android APK
 bun run android:deploy     # Deploy to device
 bun run android:dev        # Build + deploy debug

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

@@ -18,20 +18,81 @@ 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)..."
     rm -rf node_modules/.vite dist .svelte-kit .next build target src-tauri/target 2>/dev/null || true
-    npm install > /dev/null 2>&1
+    # `bun install`, NOT `npm install`. This is a bun project (see packageManager
+    # in package.json) and bun.lock is the lockfile that is committed; npm
+    # ignores it, re-resolves the tree from package.json alone, and writes a
+    # package-lock.json that .gitignore then hides.
+    #
+    # That is not cosmetic. The Tauri CLI refuses to build when a plugin's Rust
+    # crate and npm package differ by minor version, so the JS side is pinned
+    # exactly to match Cargo.lock; a re-resolve is precisely how those halves
+    # drift apart again. A clean build must not be able to change what gets
+    # installed.
+    bun install > /dev/null 2>&1
 fi
 
 # Step 1: Sync Android source files
@@ -43,17 +104,34 @@ echo "🎨 Building frontend..."
 bun run build
 
 # Step 2: Build Android APK
-if [ "$BUILD_TYPE" = "release" ]; then
+# `--apk` is a boolean flag, NOT `--apk true`.
+#
+# tauri-cli took a value here until 2.10; from 2.11 it is a plain flag and the
+# stray `true` is parsed as a positional argument, failing with
+# "error: unexpected argument 'true' found" before the build starts. Found by
+# deploying to a device after the Tauri 2.9.5 -> 2.11.5 upgrade.
+if [ "$BUILD_TYPE" = "release" ] && [ "$SIDE_BY_SIDE" = "1" ]; then
+    # A release build in the debug slot: R8 still runs, but the applicationId is
+    # suffixed and the debug keystore signs it (read by build.gradle.kts from
+    # JT_SIDE_BY_SIDE), so the real key is not needed and it replaces any other
+    # .debug install cleanly. Deliberately does NOT write keystore.properties.
+    echo "📦 Building side-by-side release APK (com.dtourolle.jellytau.debug)..."
+    JT_SIDE_BY_SIDE=1 bun run tauri android build --apk "${TARGET_ARGS[@]}"
+elif [ "$BUILD_TYPE" = "release" ]; then
     # Configure release signing from .env (single source of truth). Must run
     # after sync-android-sources.sh, since gen/android is (re)generated there.
     ./scripts/write-keystore-properties.sh
     echo "📦 Building release APK..."
-    bun run tauri android build --apk true
+    bun run tauri android build --apk "${TARGET_ARGS[@]}"
 else
     echo "📦 Building debug APK..."
-    bun run tauri android build --apk true --debug
+    bun run tauri android build --apk --debug "${TARGET_ARGS[@]}"
 fi
 
 echo ""
 echo "✅ APK build complete!"
 echo "📱 APK location: src-tauri/gen/android/app/build/outputs/apk/"
+
+# Containerised builds run as root against a bind-mounted tree; hand the
+# artifacts back to the host user. No-op when not root. See DR-213.
+"$(dirname "$0")/restore-ownership.sh"

scripts/build-builder-image.sh

@@ -26,19 +26,62 @@ echo "🏷️  Tagging for registry..."
 docker tag ${IMAGE_NAME}:${IMAGE_TAG} ${FULL_IMAGE_NAME}
 
 # Step 3: Login to registry (if not already logged in)
+#
+# `docker info | grep Username` only ever reports a Docker Hub session, so for a
+# private registry it never matched — meaning this branch fired on every push and
+# dropped into an interactive `docker login`, which hangs any non-interactive run
+# (a scripted release, or CI). Check the credential store for this specific
+# registry instead, and refuse rather than prompt when there is no TTY to
+# prompt on.
 echo "🔐 Checking registry authentication..."
-if ! docker info | grep -q "Username"; then
-    echo "Not authenticated to Docker. Logging in to ${REGISTRY_HOST}..."
-    docker login ${REGISTRY_HOST}
+DOCKER_CFG="${DOCKER_CONFIG:-$HOME/.docker}/config.json"
+if ! grep -q "\"${REGISTRY_HOST}\"" "$DOCKER_CFG" 2>/dev/null; then
+    if [ -t 0 ]; then
+        echo "Not authenticated to ${REGISTRY_HOST}. Logging in..."
+        docker login "${REGISTRY_HOST}"
+    else
+        echo "❌ Not authenticated to ${REGISTRY_HOST}, and stdin is not a TTY."
+        echo "   Run this first:  docker login ${REGISTRY_HOST}"
+        exit 1
+    fi
+else
+    echo "   Using stored credentials for ${REGISTRY_HOST}."
 fi
 
 # Step 4: Push to registry
+#
+# Two tags, on purpose:
+#
+#   <date>   what the workflows pin (e.g. :2026.08). CI must name an immutable
+#            tag -- while every job said :latest, rebuilding the image silently
+#            changed what every build, including a rebuild of an old release
+#            tag, compiled against. That is the opposite of reproducible.
+#   latest   convenience for local `docker compose` runs and for anyone pulling
+#            the image by hand.
+#
+# Date tags rather than per-commit SHA tags: the Gitea runner shares a 74 GB
+# disk with two other projects, and SHA-tagged images accumulated there until it
+# filled. Keep at most a couple of dated tags live and prune the rest
+# (`docker image prune -a` on the runner).
+#
+# To bump: build+push a new dated tag, then update the `image:` lines in
+# .gitea/workflows/*.yml in the same commit as whatever needed the new tool.
 echo "📤 Pushing image to registry..."
 docker push ${FULL_IMAGE_NAME}
 
+if [ "$IMAGE_TAG" != "latest" ]; then
+    echo "🏷️  Also tagging as :latest for local use..."
+    LATEST_IMAGE_NAME="${REGISTRY_HOST}/${REGISTRY_USER}/${IMAGE_NAME}:latest"
+    docker tag ${IMAGE_NAME}:${IMAGE_TAG} ${LATEST_IMAGE_NAME}
+    docker push ${LATEST_IMAGE_NAME}
+fi
+
 echo ""
 echo "✅ Successfully built and pushed: ${FULL_IMAGE_NAME}"
 echo ""
-echo "Update your workflow to use:"
+echo "Workflows must pin the dated tag, not :latest --"
 echo "  container:"
 echo "    image: ${FULL_IMAGE_NAME}"
+echo ""
+echo "Currently pinned in .gitea/workflows/:"
+grep -ho "jellytau-builder:[A-Za-z0-9._-]*" "$(git rev-parse --show-toplevel)"/.gitea/workflows/*.yml 2>/dev/null | sort -u | sed "s/^/    /"

scripts/build-desktop-linux.sh

@@ -26,7 +26,25 @@ 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 "$BUNDLES"
+# TRACES: | DR-221
+#
+# 🔴 NO_STRIP=true is required for the AppImage bundle.
+#
+# linuxdeploy (which Tauri downloads and runs to build the AppImage) carries its
+# own `strip`, and that copy is too old to parse the `.relr.dyn` section modern
+# toolchains emit for RELR relocations. It fails on essentially every bundled
+# library:
+#
+#   strip: libzstd.so.1: unknown type [0x13] section `.relr.dyn'
+#   failed to bundle project `failed to run linuxdeploy-x86_64.AppImage`
+#
+# Ubuntu 23.10+ links with -z pack-relative-relocs by default, so the CI builder
+# image hits this exactly as a modern Arch host does. Skipping the strip step is
+# linuxdeploy's own documented escape hatch; the cost is an unstripped, larger
+# AppImage (~153 MB for a build that bundles libmpv and its ffmpeg stack).
+#
+# Remove this only after confirming a linuxdeploy release that understands RELR.
+NO_STRIP=true bun run tauri build --bundles "$BUNDLES"
 
 BUNDLE_ROOT="src-tauri/target/release/bundle"
 echo ""
@@ -42,3 +60,7 @@ if [[ -n "${OUTPUT_DIR:-}" ]]; then
   echo ""
   echo "📦 Copied bundles to $OUTPUT_DIR"
 fi
+
+# Containerised builds run as root against a bind-mounted tree; hand the
+# artifacts back to the host user. No-op when not root. See DR-213.
+"$(dirname "$0")/restore-ownership.sh"

scripts/build-windows-cross.sh

@@ -8,7 +8,7 @@
 # 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.
+# <audio> backend (WebviewAudioBackend) — see docs/build/build-windows.md.
 #
 # Requirements (present in the Docker windows-cross target / unified builder):
 #   - rustup target x86_64-pc-windows-msvc
@@ -46,6 +46,28 @@ bun run build
 # 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.
+# TRACES: | DR-221
+#
+# 🔴 Clear the bundle output before building.
+#
+# The bundle directory is not versioned and is never cleaned by cargo, and the
+# CI runner reuses src-tauri/target between builds. The copy step below globs
+# `bundle/**/*-setup.exe`, so every stale installer left there was picked up and
+# attached to the release: v0.8.2 shipped sixteen Windows installers, thirteen
+# of them from earlier versions, and v0.5.0 offered users a download list going
+# back to 0.1.0. Every release from v0.1.0 to v0.8.2 did this. It stopped only
+# because an unrelated change wiped the runner's target dir, so it is dormant
+# rather than fixed.
+#
+# Filtering the copy by version would hide it; removing the directory means a
+# stale file cannot exist to be copied. scripts/check-release-artifacts.sh is
+# the backstop if some other path reintroduces one.
+BUNDLE_DIR="src-tauri/target/$TARGET/release/bundle"
+if [[ -d "$BUNDLE_DIR" ]]; then
+  echo "🧹 Clearing previous bundle output at $BUNDLE_DIR"
+  rm -rf "$BUNDLE_DIR"
+fi
+
 if [[ "$WIN_BUNDLES" == "none" ]]; then
   bun run tauri build --runner cargo-xwin --target "$TARGET" --no-bundle
 else
@@ -62,8 +84,18 @@ 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 \( -name '*-setup.exe' -o -name '*.msi' \) \
+  #
+  # The .sig files come along too: when TAURI_SIGNING_PRIVATE_KEY is set the
+  # bundler writes `<installer>.sig` beside each installer, and that signature is
+  # what the updater verifies before installing anything. Leaving it behind
+  # produces a release whose manifest references a signature that was never
+  # published, which fails only on the user's machine.
+  find "$BIN_DIR/bundle" -type f \( -name '*-setup.exe' -o -name '*.msi' -o -name '*.sig' \) \
     -exec cp -v {} "$OUTPUT_DIR/" \; 2>/dev/null || true
   echo ""
   echo "📦 Copied Windows artifacts to $OUTPUT_DIR"
 fi
+
+# Containerised builds run as root against a bind-mounted tree; hand the
+# artifacts back to the host user. No-op when not root. See DR-213.
+"$(dirname "$0")/restore-ownership.sh"

scripts/check-doc-links.sh

@@ -0,0 +1,198 @@
+#!/usr/bin/env bash
+# Documentation link integrity: every relative markdown link must point at a
+# file that exists.
+#
+# Implements DR-208 (see docs/requirements.md).
+#
+# Why this exists: docs/traceability.md is generated into docs/ while its file
+# links were emitted repo-root-relative, so all ~2,800 of them resolved to
+# docs/src-tauri/… and 404'd — in the Gitea repo browser and on the published
+# mdBook site alike. Nobody clicks 2,800 links, so it went unnoticed for months.
+# Several hand-written docs had the same defect at smaller scale: links to files
+# that had been deleted, and links written as if the doc lived at the repo root.
+# A link that does not resolve is a documentation defect of the same kind as a
+# compile error, and a grep is enough to catch the whole class.
+#
+# What it checks: for every tracked `.md` file, every inline markdown link
+# `[text](target)` whose target is a *path* — the target is resolved relative to
+# the directory of the file containing it, and must exist on disk.
+#
+# ⚠️ It validates PATHS, NOT ANCHORS. A green run does not mean the links land
+# where the text claims.
+#
+# 🔴 What it deliberately CANNOT see (do not read a green run as proof):
+#   - **Anchor fragments.** `foo.md#some-heading` is checked only as `foo.md`.
+#     Resolving the fragment needs a markdown renderer's heading-slug rules
+#     (which differ between Gitea, GitHub and mdBook), so a link to a heading
+#     that was renamed still passes here. That is a deliberate scope cut, not an
+#     oversight.
+#   - **External URLs.** http(s):// and mailto: are skipped. Checking them means
+#     network I/O in a gate, which makes the gate flaky and slow; link rot in an
+#     external URL is also not something a commit can break.
+#   - **Reference-style links** (`[text][ref]` with a separate `[ref]: target`
+#     definition) and bare autolinks. This project writes inline links; add the
+#     pattern here if that changes.
+#   - **Links inside fenced code blocks**, which are intentionally skipped —
+#     a template being *shown* to the reader (e.g. the release-notes template in
+#     docs/release-checklist.md) is sample text, not a live link, and its targets
+#     are resolved wherever it is eventually pasted, not from the docs tree.
+#   - **A link that resolves to the wrong existing file.** Existence is not
+#     correctness.
+#
+# Usage:  bash scripts/check-doc-links.sh
+# Exits non-zero, listing file:line and the unresolved target, on any failure.
+
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+# Generated, vendored or build-output trees. Their markdown is not authored here
+# and their link targets are not ours to fix.
+#
+# Only consulted when this is NOT a git checkout — inside one, the tracked-file
+# list does this job and does not need maintaining. Kept for the tarball case.
+EXCLUDES=(
+  "./node_modules/*"
+  "./.svelte-kit/*"
+  "./build/*"
+  "./dist/*"
+  "./src-tauri/gen/*"
+  "./src-tauri/target/*"
+  "./.git/*"
+  # Agent/dev scratch worktrees (.claude/worktrees is itself git-ignored). These
+  # are full checkouts of the repo, so without this the checker walks every
+  # in-flight branch and reports its links as if they were ours.
+  "./.claude/*"
+)
+
+# Targets that do not exist in the repo *by design* because the publish-docs job
+# writes them into docs/ at build time (see .gitea/workflows/publish-docs.yml).
+# Keep this list to genuinely generated pages — anything else here is a broken
+# link being hidden.
+GENERATED_TARGETS=(
+  "./docs/README.md"        # the site's landing page, written by publish-docs
+  "./docs/api-redirect.md"  # the rustdoc redirect stub, likewise
+)
+
+is_generated() {
+  local candidate="$1"
+  for generated in "${GENERATED_TARGETS[@]}"; do
+    [[ "$candidate" == "$generated" ]] && return 0
+  done
+  return 1
+}
+
+echo "🔎 Checking relative markdown links resolve to files on disk…"
+
+# Ask git which markdown files are ours, rather than walking the filesystem.
+#
+# This started as a find(1) with a hand-maintained prune list, and that list was
+# wrong three times in a row: it walked the scratch worktrees under .claude/,
+# then makepkg's vendored cargo registry under packaging/arch/src/ — each time
+# reporting a dependency's broken README as if it were ours. Every one of those
+# directories is already git-ignored, so the tracked-file list is the exclusion
+# rule, and it cannot drift out of date the way EXCLUDES did. It also matches
+# what this script always claimed to do.
+#
+# Untracked-but-not-ignored files are deliberately included: a new doc added in
+# a working tree should be checked before it is committed, not after.
+if git rev-parse --git-dir >/dev/null 2>&1; then
+  mapfile -t md_files < <(
+    { git ls-files -z --cached --others --exclude-standard -- '*.md' | tr '\0' '\n'; } \
+      | sed 's|^|./|' | sort -u
+  )
+else
+  # Not a git checkout (an exported tarball, say): fall back to walking, with
+  # the prune list below as the only defence.
+  find_args=(. )
+  for pattern in "${EXCLUDES[@]}"; do
+    find_args+=(-path "$pattern" -prune -o)
+  done
+  find_args+=(-name "*.md" -type f -print)
+  mapfile -t md_files < <(find "${find_args[@]}" | sort)
+fi
+
+echo "   ${#md_files[@]} markdown files"
+
+broken=""
+checked=0
+
+for md in "${md_files[@]}"; do
+  dir="$(dirname "$md")"
+
+  # One documented exception: docs-site/SUMMARY.md is mdBook's table of
+  # contents, and the publish-docs job copies it *into* docs/ before rendering
+  # (book.toml sets src = "../docs"). Its links are therefore written relative
+  # to docs/, not to the directory the file is stored in. Resolving it from
+  # docs/ is what actually validates it — and it is the check that catches a
+  # SUMMARY entry pointing at a page that does not exist, which mdBook itself
+  # only warns about.
+  if [[ "$md" == "./docs-site/SUMMARY.md" ]]; then
+    dir="./docs"
+  fi
+
+  # Strip fenced code blocks (``` and ~~~) before extracting links, so sample
+  # markdown shown to the reader is not checked as if it were a live link.
+  # Line numbers are preserved by blanking the lines rather than deleting them.
+  #
+  # Then emit "lineno<TAB>target" for each inline link on each surviving line.
+  while IFS=$'\t' read -r lineno target; do
+    [[ -z "${target:-}" ]] && continue
+
+    # Skip external schemes and pure-anchor links.
+    case "$target" in
+      http://*|https://*|mailto:*|ftp://*|"#"*|"") continue ;;
+      # A protocol-relative or scheme-ish target we do not resolve.
+      //*) continue ;;
+    esac
+
+    # Drop any anchor fragment and query string — we check the path only.
+    path="${target%%#*}"
+    path="${path%%\?*}"
+    [[ -z "$path" ]] && continue
+
+    # Percent-decode: SvelteKit route directories are literally named `[id]`,
+    # which docs link as `%5Bid%5D`, and spaces appear as `%20`.
+    if [[ "$path" == *%* ]]; then
+      path="$(printf '%b' "${path//%/\\x}")"
+    fi
+
+    checked=$((checked + 1))
+
+    if is_generated "$dir/$path"; then
+      continue
+    fi
+
+    if [[ ! -e "$dir/$path" ]]; then
+      broken+="${md}:${lineno}  ->  ${target}"$'\n'
+    fi
+  done < <(
+    awk '
+      /^[[:space:]]*(```|~~~)/ { fence = !fence; print ""; next }
+      fence { print ""; next }
+      { print }
+    ' "$md" |
+    grep -noE '\]\([^)[:space:]]+' |
+    sed -E 's/^([0-9]+):\]\(/\1\t/'
+  )
+done
+
+echo "   $checked relative links checked"
+
+if [[ -n "$broken" ]]; then
+  echo ""
+  echo "❌ Broken documentation links — these targets do not exist on disk:"
+  echo ""
+  echo "$broken" | sed 's/^/   /'
+  echo "   Each link is resolved relative to the directory of the file it is in."
+  echo "   The usual causes:"
+  echo "     • the target file was moved or deleted — update or drop the link;"
+  echo "     • the link was written as if the doc lived at the repo root — a doc"
+  echo "       in docs/ needs '../' to reach src/, scripts/ or CHANGELOG.md;"
+  echo "     • a generated doc emits repo-root-relative hrefs — fix the"
+  echo "       generator, not the output (see scripts/extract-traces.ts)."
+  exit 1
+fi
+
+echo "✅ All relative documentation links resolve."
+echo "   (Reminder: paths only — anchors and external URLs are NOT checked.)"

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 +11,30 @@
 #
 # ⚠️ 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 (DR-094). The pattern below is the hardened one: it
+# matches an item-type array literal anywhere, not just at a query site.
 #
 # Escaping a genuine exception: add the file+reason to the ALLOWLIST below.
 
@@ -28,7 +42,7 @@ 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 +50,31 @@ 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
+  # a backend-supplied flag; deferred rather than bundled with the tripwire work.
+  "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 +83,28 @@ 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 +123,11 @@ 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-release-artifacts.sh

@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+# Refuse to publish a release whose artifacts are not all from this release.
+#
+# TRACES: | DR-220
+#
+#   ./scripts/check-release-artifacts.sh <version> <dir> [<dir>...]
+#
+# e.g.
+#   ./scripts/check-release-artifacts.sh v0.9.2 artifacts/linux artifacts/windows
+#
+# ## The defect this exists for
+#
+# Every JellyTau release from v0.1.0 to v0.8.2 shipped every Windows installer
+# ever built. `src-tauri/target/*/release/bundle/` is not versioned, cargo never
+# cleans it, and the CI runner reuses the target directory between builds — so
+# the copy step's `bundle/**/*-setup.exe` glob collected the whole history. By
+# v0.8.2 that was sixteen installers, thirteen of them stale. v0.5.0 offered
+# users a download list going back to 0.1.0.
+#
+# Nobody noticed for eight months. There was nothing to notice with: the upload
+# loop reported success, the assets were real files, and the release page looked
+# busy rather than wrong.
+#
+# The builds now clear the bundle directory first, which removes the cause. This
+# is the backstop for the next thing that reintroduces a stale file by a route
+# nobody predicted — a cached directory, a restored artifact, a hand-copied fix.
+#
+# ## What it checks
+#
+# Every file whose name embeds a semantic version must embed *this* version.
+# Files with no version in the name (jellytau-release.apk, jellytau.exe,
+# SHA256SUMS, latest.json) are accepted: they are produced fresh each build and
+# have no version to disagree with.
+
+set -euo pipefail
+
+if [ "$#" -lt 2 ]; then
+  echo "usage: $0 <version> <dir> [<dir>...]" >&2
+  exit 2
+fi
+
+VERSION_RAW="$1"
+shift
+# Accept the tag form (v0.9.2) or the bare form (0.9.2).
+VERSION="${VERSION_RAW#v}"
+
+echo "🔎 Checking release artifacts are all version ${VERSION}…"
+
+FOUND=0
+STALE=0
+UNVERSIONED=0
+
+for dir in "$@"; do
+  if [ ! -d "$dir" ]; then
+    echo "   (no $dir — skipping)"
+    continue
+  fi
+
+  # -print0/read -d '' so a filename with a space cannot split into two.
+  while IFS= read -r -d '' file; do
+    name="$(basename "$file")"
+    FOUND=$((FOUND + 1))
+
+    # First x.y.z in the filename, if any.
+    embedded="$(printf '%s' "$name" | grep -oE '[0-9]+\.[0-9]+\.[0-9]+' | head -1 || true)"
+
+    if [ -z "$embedded" ]; then
+      UNVERSIONED=$((UNVERSIONED + 1))
+      continue
+    fi
+
+    if [ "$embedded" != "$VERSION" ]; then
+      echo "   ❌ $name carries version $embedded"
+      STALE=$((STALE + 1))
+    fi
+  done < <(find "$dir" -type f -print0)
+done
+
+echo ""
+echo "   $FOUND file(s) checked; $UNVERSIONED carry no version in the name."
+
+if [ "$FOUND" -eq 0 ]; then
+  echo "❌ No artifacts found at all. A release with no files is a failed build," >&2
+  echo "   not an empty one." >&2
+  exit 1
+fi
+
+if [ "$STALE" -gt 0 ]; then
+  echo ""
+  echo "❌ $STALE artifact(s) belong to a different version than ${VERSION}." >&2
+  echo "" >&2
+  echo "   This is how every release from v0.1.0 to v0.8.2 came to ship its" >&2
+  echo "   predecessors' Windows installers: src-tauri/target/*/release/bundle/" >&2
+  echo "   is never cleaned and the runner reuses it, so a glob picks up" >&2
+  echo "   whatever was left behind." >&2
+  echo "" >&2
+  echo "   The builds clear that directory first, so seeing this means a stale" >&2
+  echo "   file arrived by some other route. Find it before publishing — do not" >&2
+  echo "   delete the file and re-run." >&2
+  exit 1
+fi
+
+echo "✅ Every versioned artifact is ${VERSION}."

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/check-tooling.sh

@@ -0,0 +1,73 @@
+#!/usr/bin/env bash
+# Refuse tooling that contradicts what this project actually uses.
+#
+# TRACES: | DR-222
+#
+#   ./scripts/check-tooling.sh
+#
+# ## Why
+#
+# This is a bun project: `packageManager` in package.json says so, bun.lock is
+# the committed lockfile, and .gitignore hides the other package managers'
+# lockfiles precisely so they cannot be committed by accident.
+#
+# scripts/build-android.sh nonetheless ran `npm install` on its clean-build
+# path. npm ignores bun.lock, re-resolves the whole tree from package.json, and
+# writes a package-lock.json that .gitignore then hides from view.
+#
+# That is not a style preference. The Tauri CLI refuses to build when a plugin's
+# Rust crate and npm package differ by minor version, so the JS side is pinned
+# exactly against Cargo.lock -- and a silent re-resolve is exactly how those
+# halves drift apart again. The drift already cost one release build.
+#
+# It survived because the clean-build path runs rarely. That is the shape of
+# nearly every defect found while preparing v0.10.0: the code that runs on every
+# commit was fine, and the code that runs on a release, a clean build or a tag
+# had no guard at all.
+
+set -uo pipefail
+
+REPO_ROOT="$(git rev-parse --show-toplevel)"
+cd "$REPO_ROOT" || exit 1
+
+FAILED=0
+
+echo "🔎 Checking build tooling is consistent with packageManager…"
+
+# Only the *invocations* matter. A comment explaining why npm is wrong, or a
+# .gitignore entry naming package-lock.json, is not a violation -- so match a
+# command at the start of a line or after a shell separator.
+PATTERN='(^|[;&|(]|&&|\|\||\bthen |\bdo |[[:space:]]{4,})(npm|yarn|pnpm)[[:space:]]+(install|ci|add|run|exec)\b'
+
+MATCHES="$(grep -rInE "$PATTERN" \
+  --include='*.sh' --include='*.yml' --include='*.yaml' \
+  scripts/ .gitea/ 2>/dev/null | grep -v '^\s*#' || true)"
+
+if [ -n "$MATCHES" ]; then
+  echo "❌ A non-bun package manager is invoked:"
+  echo "$MATCHES" | sed 's/^/   /'
+  echo ""
+  echo "   This project uses bun (packageManager in package.json, bun.lock"
+  echo "   committed). npm/yarn/pnpm ignore that lockfile and re-resolve the"
+  echo "   dependency tree, which is how the Tauri plugin crate/package"
+  echo "   versions drifted apart and broke a release build."
+  echo ""
+  echo "   Use: bun install / bun run / bunx"
+  FAILED=1
+fi
+
+# A lockfile from another manager should never exist here; .gitignore hides
+# them, so one can sit in a working tree unnoticed and change what installs.
+for stray in package-lock.json yarn.lock pnpm-lock.yaml; do
+  if [ -f "$stray" ]; then
+    echo "❌ $stray exists. Another package manager has run here."
+    echo "   Delete it and run: bun install"
+    FAILED=1
+  fi
+done
+
+if [ "$FAILED" -eq 0 ]; then
+  echo "✅ Only bun is used, and no foreign lockfile is present."
+fi
+
+exit "$FAILED"

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 +1,406 @@
+/**
+ * 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
+ * @req-test: UT-202 - Generated matrix links resolve from docs/
+ */
+
+import { describe, it, expect } from "vitest";
+import * as fs from "fs";
+import * as path from "path";
+import {
+  countDefinedRequirements,
+  computeCoverage,
+  findDanglingIds,
+  formatMatrixFileLink,
+  generateMarkdown,
+  isTracedSourceFile,
+  MIN_COVERAGE_PERCENT,
+  type TracesData,
+} from "./extract-traces";
+
+// import.meta.dir is Bun-only; derive from import.meta.url under vitest.
+const HERE = path.dirname(new URL(import.meta.url).pathname);
+
+describe("isTracedSourceFile", () => {
+  // The extractor used to accept only .ts/.svelte/.rs under src/, src-tauri/src/
+  // and scripts/. Every requirement implemented by *configuration* was therefore
+  // invisible to the matrix that measures it: eslint.config.js (DR-205), the
+  // pre-commit hook (DR-207), rust-toolchain.toml (DR-206) and deny.toml
+  // (DR-216) all carry TRACES comments that were never read. Each one counted
+  // against coverage as an uncovered requirement while being, in fact, covered.
+  it("accepts the source extensions it always did", () => {
+    expect(isTracedSourceFile("src/lib/utils/logger.ts")).toBe(true);
+    expect(isTracedSourceFile("src/routes/settings/+page.svelte")).toBe(true);
+    expect(isTracedSourceFile("src-tauri/src/lib.rs")).toBe(true);
+  });
+
+  it("accepts tooling files that implement a requirement", () => {
+    expect(isTracedSourceFile("eslint.config.js")).toBe(true);
+    expect(isTracedSourceFile("src-tauri/deny.toml")).toBe(true);
+    expect(isTracedSourceFile("src-tauri/rust-toolchain.toml")).toBe(true);
+    expect(isTracedSourceFile("scripts/hooks/pre-commit")).toBe(true);
+    // Shell tooling is listed individually, not globbed: most scripts/*.sh
+    // implement nothing, and adding one should be a decision.
+    expect(isTracedSourceFile("scripts/check-release-artifacts.sh")).toBe(true);
+    expect(isTracedSourceFile("scripts/build-desktop-linux.sh")).toBe(true);
+    expect(isTracedSourceFile("scripts/logcat.sh")).toBe(false);
+  });
+
+  it("does not scan CI workflows, whose comments discuss TRACES in prose", () => {
+    // .gitea/workflows/traceability-check.yml explains the gate, so it contains
+    // lines like "a `TRACES:` comment ... (DR-189 and UT-188 lived in three
+    // source files, defined nowhere)". The extractor's pattern would read that
+    // as a trace and manufacture references to IDs that do not exist, failing
+    // traces:validate. A file that *describes* traceability is not a file that
+    // implements a requirement.
+    expect(isTracedSourceFile(".gitea/workflows/traceability-check.yml")).toBe(false);
+    expect(isTracedSourceFile(".gitea/workflows/build-and-test.yml")).toBe(false);
+  });
+
+  it("rejects files that merely mention a requirement in prose", () => {
+    // requirements.md defines IDs; traceability.md is generated *from* traces.
+    // Scanning either would make every requirement trace to itself.
+    expect(isTracedSourceFile("docs/requirements.md")).toBe(false);
+    expect(isTracedSourceFile("docs/traceability.md")).toBe(false);
+    expect(isTracedSourceFile("README.md")).toBe(false);
+  });
+
+  it("rejects generated and vendored trees", () => {
+    expect(isTracedSourceFile("node_modules/foo/index.ts")).toBe(false);
+    expect(isTracedSourceFile("src-tauri/target/debug/build/x.rs")).toBe(false);
+    expect(isTracedSourceFile("src-tauri/gen/android/app/build.gradle.kts")).toBe(false);
+  });
+});
+
+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);
+  });
+
+  it("collects UT/IT rows separately, out of the coverage denominator", () => {
+    // §4 defines the test taxonomy. Those rows must be known (so a TRACES
+    // comment may name them) without ever moving the coverage ratio.
+    const md = `
+| UR-001 | A | High | Done |
+| UT-001 | Player state transitions | DR-001 | Pending |
+| IT-004 | Playback end-to-end | DR-002 | Pending |
+`;
+    const defined = countDefinedRequirements(md);
+    expect(defined.total).toBe(1);
+    expect(defined.ids.has("UT-001")).toBe(false);
+    expect(defined.testIds.has("UT-001")).toBe(true);
+    expect(defined.testIds.has("IT-004")).toBe(true);
+  });
+});
+
+describe("findDanglingIds", () => {
+  const defined = {
+    UR: 1,
+    IR: 0,
+    DR: 1,
+    JA: 0,
+    total: 2,
+    ids: new Set(["UR-001", "DR-001"]),
+    testIds: new Set(["UT-001"]),
+  };
+
+  it("flags a requirement ID that requirements.md does not define", () => {
+    expect(findDanglingIds(["UR-001", "DR-189"], defined)).toEqual(["DR-189"]);
+  });
+
+  it("flags an undefined UT/IT id, which the coverage orphan list cannot", () => {
+    // The gap this closes: computeCoverage deliberately ignores UT/IT, so
+    // UT-188 sat in three source files, defined nowhere, entirely unreported.
+    expect(computeCoverage(["UT-188"], defined).orphaned).toEqual([]);
+    expect(findDanglingIds(["UT-188"], defined)).toEqual(["UT-188"]);
+  });
+
+  it("accepts every ID that is defined, requirement or test", () => {
+    expect(findDanglingIds(["UR-001", "DR-001", "UT-001"], defined)).toEqual([]);
+  });
+
+  it("deduplicates and sorts, so one typo is reported once", () => {
+    expect(findDanglingIds(["DR-189", "DR-189", "UR-999", "DR-189"], defined)).toEqual([
+      "DR-189",
+      "UR-999",
+    ]);
+  });
+
+  it("ignores IDs whose prefix is not a known trace type", () => {
+    // e.g. an unrelated "AB-123" caught by the loose ID regex.
+    expect(findDanglingIds(["AB-123"], defined)).toEqual([]);
+  });
+});
+
+describe("coverage threshold", () => {
+  it("matches MIN_THRESHOLD in the Gitea traceability workflow", () => {
+    // Two files must agree on the gate: the script (local `traces:coverage`)
+    // and the workflow. Drift means the local gate and CI disagree about what
+    // passes, which is how the 50%-while-actually-86% slack went unnoticed.
+    const workflow = fs.readFileSync(
+      path.resolve(HERE, "../.gitea/workflows/traceability-check.yml"),
+      "utf-8",
+    );
+    const match = workflow.match(/^\s*MIN_THRESHOLD=(\d+)\s*$/m);
+    expect(match).not.toBeNull();
+    expect(Number(match![1])).toBe(MIN_COVERAGE_PERCENT);
+  });
+
+  it("is a ratchet: never lower it to make a red build pass", () => {
+    // Sanity bound. If coverage genuinely climbs, raise both numbers together.
+    expect(MIN_COVERAGE_PERCENT).toBeGreaterThanOrEqual(82);
+    expect(MIN_COVERAGE_PERCENT).toBeLessThanOrEqual(100);
+  });
+});
+
+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"]),
+    testIds: new Set<string>(),
+  };
+
+  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>(),
+      testIds: 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("generated matrix file links", () => {
+  // Regression: the generator emitted the repo-root-relative path as the href
+  // (`](src-tauri/src/…)`), but writes its output to docs/traceability.md — so
+  // every one of the ~2,800 links resolved to docs/src-tauri/… and 404'd, in
+  // the repo browser and on the published mdBook site. The markdown generator
+  // had no test at all, which is why it survived. UT-202.
+  //
+  // @req-test: UT-202
+
+  /** A minimal TracesData whose single entry points at a file that really exists. */
+  function fixture(file: string, line = 12): TracesData {
+    return {
+      timestamp: new Date().toISOString(),
+      totalFiles: 1,
+      totalTraces: 1,
+      requirements: {
+        "DR-093": [{ file, line, context: "export function x() {}" }],
+      },
+      byType: { UR: [], IR: [], DR: ["DR-093"], JA: [] },
+    } as TracesData;
+  }
+
+  /** Pull the href out of the first `- **File:** [`x`](href)` line. */
+  function firstHref(md: string): string {
+    const m = md.match(/^- \*\*File:\*\* \[`[^`]+`\]\(([^)]+)\)/m);
+    expect(m).not.toBeNull();
+    return m![1];
+  }
+
+  it("emits an href that resolves, from docs/, to a file that exists", () => {
+    // Use a real repo file so "exists on disk" is a genuine assertion.
+    const target = "scripts/extract-traces.ts";
+    const md = generateMarkdown(fixture(target));
+
+    const href = firstHref(md);
+    const [relPath] = href.split("#");
+
+    // traceability.md is written to docs/, so links resolve from there.
+    const resolved = path.resolve(HERE, "../docs", relPath);
+    expect(fs.existsSync(resolved)).toBe(true);
+    expect(resolved).toBe(path.resolve(HERE, "..", target));
+  });
+
+  it("keeps the repo-root-relative path as the visible link text", () => {
+    // The text is what a developer copies into an editor or a grep; only the
+    // href is rewritten for the docs/ location.
+    const md = generateMarkdown(fixture("src-tauri/src/lib.rs"));
+    expect(md).toContain("[`src-tauri/src/lib.rs`]");
+    expect(md).not.toContain("[`../src-tauri/src/lib.rs`]");
+  });
+
+  it("keeps the #Lnn line anchor on the href", () => {
+    const link = formatMatrixFileLink("scripts/extract-traces.ts", 427);
+    expect(link).toBe("[`scripts/extract-traces.ts`](../scripts/extract-traces.ts#L427)");
+  });
+
+  it("does not produce a bare repo-root href, which resolves to docs/<path>", () => {
+    const md = generateMarkdown(fixture("scripts/extract-traces.ts"));
+    const href = firstHref(md);
+    expect(href.startsWith("../")).toBe(true);
+    // The pre-fix output — the exact shape that produced docs/scripts/….
+    expect(href.startsWith("scripts/")).toBe(false);
+  });
+});
+
+describe("live requirements.md", () => {
+  it("parses the real file into a self-consistent denominator", () => {
+    // Guards the original regression: CI hardcoded UR/39, IR/24, DR/48, JA/3
+    // (total 114) while the real file had grown past 200, so the gate compared
+    // live traces against a frozen denominator and reported 158% coverage.
+    //
+    // Deliberately asserts *invariants*, not exact totals. Pinning the counts
+    // was tried and turned this test into a merge-conflict magnet: every
+    // requirement added on any branch had to edit the numbers here too, and the
+    // comment above them grew into a ledger of which branch contributed which
+    // row. Worse, the pins never guarded the actual defect — a stale denominator
+    // is caught by the sum-consistency check below, and the >100% ratio it
+    // produced is covered directly by the computeCoverage tests, on fixtures.
+    const md = fs.readFileSync(path.resolve(HERE, "../docs/requirements.md"), "utf-8");
+    const defined = countDefinedRequirements(md);
+
+    // The parser found real rows of every type: a section silently failing to
+    // parse would shrink the denominator and inflate coverage.
+    expect(defined.UR).toBeGreaterThan(0);
+    expect(defined.IR).toBeGreaterThan(0);
+    expect(defined.DR).toBeGreaterThan(0);
+    expect(defined.JA).toBeGreaterThan(0);
+
+    // The denominator is the sum of its parts, and every counted id is unique —
+    // double-counting one section is the other way a ratio breaks.
+    expect(defined.total).toBe(defined.UR + defined.IR + defined.DR + defined.JA);
+    expect(defined.ids.size).toBe(defined.total);
+
+    // The file is live, not frozen: it is well past the 114 the stale gate used.
+    expect(defined.total).toBeGreaterThan(200);
+  });
+});

scripts/extract-traces.ts

@@ -23,7 +23,7 @@ interface RequirementMapping {
   [reqId: string]: TraceEntry[];
 }
 
-interface TracesData {
+export interface TracesData {
   timestamp: string;
   totalFiles: number;
   totalTraces: number;
@@ -34,11 +34,38 @@ 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;
+  /** Traced IDs of any type that requirements.md does not define. */
+  dangling?: string[];
 }
 
+/**
+ * Minimum overall requirement coverage the traceability gate accepts.
+ *
+ * **Ratchet policy: this number only ever goes up.** It is set a few points
+ * below the coverage actually achieved, so a real regression trips it instead of
+ * being absorbed by slack. It sat at 50 while true coverage was 86%, which meant
+ * half the matrix could rot before CI noticed. When coverage rises durably,
+ * raise this to sit just under the new figure. Do **not** lower it to make a
+ * failing build pass — add the missing TRACES comments instead.
+ *
+ * `.gitea/workflows/traceability-check.yml` carries the same number as
+ * `MIN_THRESHOLD`; `scripts/extract-traces.test.ts` fails if the two drift.
+ *
+ * TRACES: | DR-093
+ */
+export const MIN_COVERAGE_PERCENT = 89;
+
 // 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;
@@ -48,9 +75,85 @@ function extractRequirementIds(tracesString: string): string[] {
   return matches.map((m) => `${m[1]}-${m[2]}`);
 }
 
+/**
+ * Tooling files that implement a requirement.
+ *
+ * The walker below only visits `src/`, `src-tauri/src/` and `scripts/`, and only
+ * picks up `.ts`/`.svelte`/`.rs`. That made every requirement implemented by
+ * *configuration* invisible to the matrix that measures it — DR-205
+ * (eslint.config.js), DR-206 (rust-toolchain.toml), DR-207 (the pre-commit
+ * hook) and DR-216 (deny.toml) all carry TRACES comments that nothing read, so
+ * each was counted as uncovered while being covered.
+ *
+ * An explicit list rather than "also scan .toml/.js/.yml": most config files in
+ * this repo implement nothing, and one class of file is actively dangerous to
+ * scan — see `isTracedSourceFile`.
+ */
+const TOOLING_FILES = new Set([
+  "eslint.config.js",
+  "vitest.config.ts",
+  "scripts/hooks/pre-commit",
+  "src-tauri/deny.toml",
+  "src-tauri/rust-toolchain.toml",
+  // Shell tooling that implements a requirement. Named individually rather than
+  // globbing scripts/*.sh: most of these scripts implement nothing, and the
+  // point of the list is that adding a file is a decision.
+  "scripts/install-hooks.sh",
+  "scripts/check-release-artifacts.sh",
+  "scripts/build-desktop-linux.sh",
+  "scripts/build-windows-cross.sh",
+  "scripts/restore-ownership.sh",
+]);
+
+/** Directory names that never contain hand-written traced source. */
+const EXCLUDED_SEGMENTS = new Set([
+  "node_modules",
+  "target",
+  "build",
+  ".git",
+  ".svelte-kit",
+  "docs-site",
+  // Tauri regenerates src-tauri/gen/ on every android/desktop init; the
+  // canonical Android sources live in src-tauri/android/ and are synced into it.
+  "gen",
+]);
+
+/**
+ * Decide whether a repo-relative path should be scanned for TRACES comments.
+ *
+ * Exported for scripts/extract-traces.test.ts — the file-walking half needs a
+ * filesystem, this half is a pure decision and is where the mistakes live.
+ *
+ * Deliberately excluded:
+ *  - `docs/requirements.md` *defines* IDs and `docs/traceability.md` is
+ *    generated from traces; scanning either would make requirements trace to
+ *    themselves.
+ *  - `.gitea/workflows/*.yml` — traceability-check.yml explains the gate in
+ *    prose, quoting "a `TRACES:` comment" on the same line as example IDs that
+ *    are deliberately undefined. The extractor would read those as real traces
+ *    and then fail its own dangling-ID check.
+ */
+export function isTracedSourceFile(relativePath: string): boolean {
+  const p = relativePath.split(path.sep).join("/");
+  if (p.split("/").some((segment) => EXCLUDED_SEGMENTS.has(segment))) {
+    return false;
+  }
+  if (TOOLING_FILES.has(p)) {
+    return true;
+  }
+  const isSourceExtension = p.endsWith(".ts") || p.endsWith(".svelte") || p.endsWith(".rs");
+  if (!isSourceExtension) {
+    return false;
+  }
+  return p.startsWith("src/") || p.startsWith("src-tauri/src/") || p.startsWith("scripts/");
+}
+
 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) {
@@ -60,23 +163,16 @@ function getAllSourceFiles(): string[] {
         const fullPath = path.join(dir, entry.name);
         const relativePath = path.relative(baseDir, fullPath);
 
-        // Skip node_modules, target, build
-        if (
-          relativePath.includes("node_modules") ||
-          relativePath.includes("target") ||
-          relativePath.includes("build") ||
-          relativePath.includes(".git")
-        ) {
+        // Directory pruning still happens here so the walk does not descend
+        // into node_modules/target at all; isTracedSourceFile repeats the rule
+        // for individual files (and is the version under test).
+        if (entry.isDirectory() && !isTracedSourceFile(path.join(relativePath, "x.ts"))) {
           continue;
         }
 
         if (entry.isDirectory()) {
           walkDir(fullPath);
-        } else if (
-          entry.name.endsWith(".ts") ||
-          entry.name.endsWith(".svelte") ||
-          entry.name.endsWith(".rs")
-        ) {
+        } else if (isTracedSourceFile(relativePath)) {
           files.push(fullPath);
         }
       }
@@ -92,6 +188,15 @@ function getAllSourceFiles(): string[] {
     }
   }
 
+  // eslint.config.js, deny.toml and rust-toolchain.toml sit at the repo root or
+  // in src-tauri/ rather than under a walked root, so they are added by name.
+  for (const toolingFile of TOOLING_FILES) {
+    const fullPath = path.join(baseDir, toolingFile);
+    if (fs.existsSync(fullPath) && !files.includes(fullPath)) {
+      files.push(fullPath);
+    }
+  }
+
   return files;
 }
 
@@ -192,7 +297,167 @@ function extractTraces(): TracesData {
   };
 }
 
-function generateMarkdown(data: TracesData): string {
+// ---------------------------------------------------------------------------
+// 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/traceability-ci.md, "Coverage Thresholds".
+//
+// TRACES: | DR-093
+// ---------------------------------------------------------------------------
+
+export interface DefinedRequirements {
+  UR: number;
+  IR: number;
+  DR: number;
+  JA: number;
+  total: number;
+  /** Requirement IDs (UR/IR/DR/JA) — the coverage denominator. */
+  ids: Set<string>;
+  /** Test IDs (UT/IT) from §4. A separate taxonomy: never part of coverage. */
+  testIds: 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 testIds = new Set<string>();
+  const ROW_ID = /^\|\s*(UR|IR|DR|JA|UT|IT)-(\d{3})\s*\|/;
+
+  for (const line of markdown.split("\n")) {
+    const match = line.match(ROW_ID);
+    if (!match) continue;
+    const id = `${match[1]}-${match[2]}`;
+    // UT/IT rows live in §4 and are collected separately: they must not enter
+    // the coverage denominator, but they still need to exist for a `TRACES:`
+    // comment to be allowed to name them (see findDanglingIds).
+    if (match[1] === "UT" || match[1] === "IT") testIds.add(id);
+    else ids.add(id);
+  }
+
+  const countOf = (type: string) => [...ids].filter((id) => id.startsWith(`${type}-`)).length;
+
+  return {
+    UR: countOf("UR"),
+    IR: countOf("IR"),
+    DR: countOf("DR"),
+    JA: countOf("JA"),
+    total: ids.size,
+    ids,
+    testIds,
+  };
+}
+
+/**
+ * Every traced ID that requirements.md defines nowhere — a typo, a rename that
+ * missed a call site, or a reference to a deleted requirement.
+ *
+ * This is broader than `CoverageResult.orphaned`, which only ever considers the
+ * four requirement types because a UT/IT entry among the orphans would corrupt
+ * the coverage ratio's reporting. Dangling detection has no such constraint, so
+ * it checks all six ID types against both defined sets. Before it existed, the
+ * extractor accepted any well-formed ID silently: `DR-189` and `UT-188` were
+ * referenced from `controlsVisibility.ts` and `VideoPlayer.svelte` for months
+ * without being defined anywhere, and nothing reported it.
+ *
+ * TRACES: | DR-093
+ */
+export function findDanglingIds(tracedIds: string[], defined: DefinedRequirements): string[] {
+  const KNOWN_TYPE = /^(UR|IR|DR|JA|UT|IT)-\d{3}$/;
+
+  const dangling = new Set(
+    tracedIds
+      .filter((id) => KNOWN_TYPE.test(id))
+      .filter((id) => !defined.ids.has(id) && !defined.testIds.has(id)),
+  );
+
+  return [...dangling].sort();
+}
+
+/**
+ * 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}$/.test(id);
+
+  const traced = new Set(tracedIds.filter(isRequirement));
+  const covered = [...traced].filter((id) => defined.ids.has(id));
+  const orphaned = [...traced].filter((id) => !defined.ids.has(id)).sort();
+
+  return {
+    covered: covered.length,
+    total: defined.total,
+    percent: defined.total === 0 ? 0 : Math.round((covered.length / defined.total) * 100),
+    orphaned,
+  };
+}
+
+/** Read requirements.md from the repo and count what it defines. */
+export function readDefinedRequirements(): DefinedRequirements {
+  const reqPath = path.join(BASE_DIR, "docs", "requirements.md");
+  return countDefinedRequirements(fs.readFileSync(reqPath, "utf-8"));
+}
+
+/**
+ * Path prefix that turns a repo-root-relative file path into a link target that
+ * resolves from `docs/traceability.md`, where this markdown is written.
+ *
+ * The generated matrix lives one directory below the repo root, so a bare
+ * `src-tauri/src/player/mod.rs` href resolves to `docs/src-tauri/…` and 404s —
+ * in the repo browser and on the published mdBook site alike. Every file link
+ * in the matrix was dead for this reason. The *display text* stays
+ * repo-root-relative (that is the path a developer types and greps for); only
+ * the href is rewritten.
+ *
+ * TRACES: | DR-093 | UT-202
+ */
+export const MATRIX_LINK_PREFIX = "../";
+
+/**
+ * Build the ``[`path`](href#Lnn)`` link used for one trace entry in the matrix.
+ *
+ * Exported so extract-traces.test.ts can resolve a generated href against
+ * `docs/` and assert the target exists on disk.
+ *
+ * TRACES: | DR-093 | UT-202
+ */
+export function formatMatrixFileLink(file: string, line: number): string {
+  return `[\`${file}\`](${MATRIX_LINK_PREFIX}${file}#L${line})`;
+}
+
+export function generateMarkdown(data: TracesData): string {
   let md = `# Code Traceability Matrix
 
 **Generated:** ${new Date(data.timestamp).toLocaleString()}
@@ -250,7 +515,7 @@ ${data.byType.JA.join(", ")}
     md += `**Locations:** ${entries.length} file(s)\n\n`;
 
     for (const entry of entries) {
-      md += `- **File:** [\`${entry.file}\`](${entry.file}#L${entry.line})\n`;
+      md += `- **File:** ${formatMatrixFileLink(entry.file, entry.line)}\n`;
       md += `  - **Line:** ${entry.line}\n`;
       const contextPreview = entry.context.substring(0, 70);
       md += `  - **Context:** \`${contextPreview}${entry.context.length > 70 ? "..." : ""}\`\n`;
@@ -265,21 +530,116 @@ 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.");
+  }
+
+  if (data.dangling && data.dangling.length > 0) {
+    console.log("");
+    console.log(
+      `⚠️  Dangling IDs (incl. UT/IT): ${data.dangling.join(", ")} — run \`bun run traces:validate\`.`,
+    );
+  }
+
+  // A ratio above 100% means the computation is broken (the condition that hid
+  // the stale-denominator bug for so long). Fail loudly rather than report it.
+  if (cov.percent > 100) {
+    console.log("");
+    console.log(`❌ Coverage > 100% — the gate is miscomputing.`);
+    return 1;
+  }
+
+  if (cov.percent < minThreshold) {
+    console.log("");
+    console.log(`❌ Coverage (${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`
-);
+/**
+ * Hard gate on dangling IDs: a `TRACES:` comment may only name an ID that
+ * requirements.md actually defines. Prints every offender with the files that
+ * reference it, so the fix is mechanical.
+ *
+ * TRACES: | DR-093
+ */
+function reportDangling(data: TracesData): number {
+  const dangling = data.dangling ?? [];
+
+  if (dangling.length === 0) {
+    console.log("✅ All traced IDs are defined in docs/requirements.md");
+    return 0;
+  }
+
+  console.log("❌ TRACES reference IDs that docs/requirements.md does not define:");
+  console.log("");
+  for (const id of dangling) {
+    const files = [...new Set((data.requirements[id] ?? []).map((e) => e.file))].sort();
+    console.log(`   ${id}`);
+    for (const file of files) console.log(`      ${file}`);
+  }
+  console.log("");
+  console.log("Fix each one by either:");
+  console.log("   • correcting the ID in the TRACES comment (typo/rename), or");
+  console.log("   • adding the requirement as a table row in docs/requirements.md.");
+  return 1;
+}
+
+// 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);
+  data.dangling = findDanglingIds(allTraced, defined);
+
+  if (format === "json") {
+    console.log(generateJson(data));
+  } else if (format === "coverage") {
+    process.exit(reportCoverage(data, MIN_COVERAGE_PERCENT));
+  } else if (format === "validate") {
+    process.exit(reportDangling(data));
+  } 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/hooks/pre-commit

@@ -0,0 +1,93 @@
+#!/usr/bin/env bash
+# JellyTau pre-commit hook — the fast half of CLAUDE.md's "Before Committing"
+# list, enforced instead of remembered.
+#
+# TRACES: | DR-207
+#
+# Install with:   bun run hooks:install      (sets core.hooksPath=scripts/hooks)
+# Skip once with: git commit --no-verify
+#
+# What runs here is deliberately limited to gates that finish in seconds:
+#
+#   bun run check                        svelte-check (types)
+#   bun run test                         vitest, single pass
+#   scripts/check-frontend-boundary.sh   domain-taxonomy tripwire (DR-094)
+#   bun run format:check                 prettier
+#   bun run lint --max-warnings=N        eslint, warning-count ratchet
+#   cargo fmt --all -- --check           only when src-tauri/ is staged
+#
+# NOT here, on purpose: `cargo clippy` and `cargo test`. Both take minutes on a
+# cold target dir, which turns every commit into a coffee break and trains
+# people to reach for --no-verify. CI (.gitea/workflows/build-and-test.yml) is
+# where those run; `bun run test:all` is the local equivalent.
+
+set -uo pipefail
+
+# Merge and rebase commits carry someone else's changes, and conflict resolution
+# is exactly when a slow gate is least welcome. Let them through — CI still
+# gates the merge result.
+GIT_DIR_PATH="$(git rev-parse --git-dir 2>/dev/null)" || exit 0
+if [ -e "$GIT_DIR_PATH/MERGE_HEAD" ] ||
+  [ -d "$GIT_DIR_PATH/rebase-merge" ] ||
+  [ -d "$GIT_DIR_PATH/rebase-apply" ] ||
+  [ -e "$GIT_DIR_PATH/CHERRY_PICK_HEAD" ]; then
+  echo "pre-commit: merge/rebase in progress — skipping checks (CI still gates the result)."
+  exit 0
+fi
+
+# Nothing staged (e.g. `git commit --amend` that only edits the message): nothing
+# to check.
+STAGED="$(git diff --cached --name-only --diff-filter=ACMR)"
+if [ -z "$STAGED" ]; then
+  exit 0
+fi
+
+REPO_ROOT="$(git rev-parse --show-toplevel)"
+cd "$REPO_ROOT" || exit 1
+
+FAILED=0
+
+run_gate() {
+  label="$1"
+  shift
+  echo ""
+  echo "🔎 pre-commit: $label"
+  if ! "$@"; then
+    echo "❌ pre-commit: $label failed"
+    FAILED=1
+  fi
+}
+
+run_gate "svelte-check (bun run check)" bun run check
+run_gate "frontend tests (bun run test)" bun run test
+run_gate "frontend/backend boundary" bash scripts/check-frontend-boundary.sh
+run_gate "formatting (bun run format:check)" bun run format:check
+# Same ratchet as the CI step in build-and-test.yml — keep the two numbers equal,
+# or a commit passes here and fails there.
+run_gate "lint (bun run lint)" bun run lint --max-warnings=159
+
+# rustfmt only matters when Rust actually changed, and `cargo fmt --check` is
+# cheap (no compilation) whenever it does.
+if printf '%s\n' "$STAGED" | grep -q '^src-tauri/'; then
+  if command -v cargo >/dev/null 2>&1; then
+    echo ""
+    echo "🔎 pre-commit: rustfmt (src-tauri/ is staged)"
+    if ! (cd src-tauri && cargo fmt --all -- --check); then
+      echo "❌ pre-commit: cargo fmt --all -- --check failed"
+      echo "   fix with: cd src-tauri && cargo fmt"
+      FAILED=1
+    fi
+  else
+    echo "⚠️  pre-commit: src-tauri/ staged but cargo is not on PATH — skipping rustfmt."
+  fi
+fi
+
+if [ "$FAILED" -ne 0 ]; then
+  echo ""
+  echo "🛑 pre-commit checks failed. Fix them, or bypass deliberately with:"
+  echo "     git commit --no-verify"
+  exit 1
+fi
+
+echo ""
+echo "✅ pre-commit checks passed."

scripts/install-hooks.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# Point git at the repo's tracked hooks directory.
+#
+# TRACES: | DR-207
+#
+#   bun run hooks:install        # or: ./scripts/install-hooks.sh
+#
+# `core.hooksPath` is used rather than copying files into .git/hooks so the
+# hooks stay version-controlled: an update to scripts/hooks/pre-commit reaches
+# everyone on their next pull instead of needing a re-install.
+#
+# The setting is local to this clone (git config, not committed). To undo:
+#   git config --unset core.hooksPath
+
+set -euo pipefail
+
+REPO_ROOT="$(git rev-parse --show-toplevel)"
+cd "$REPO_ROOT"
+
+HOOKS_DIR="scripts/hooks"
+
+if [ ! -d "$HOOKS_DIR" ]; then
+  echo "❌ $HOOKS_DIR does not exist — are you in the JellyTau repo?" >&2
+  exit 1
+fi
+
+# Git refuses to run a hook that is not executable, and the bit is easy to lose
+# on a fresh checkout on some filesystems.
+chmod +x "$HOOKS_DIR"/* 2>/dev/null || true
+
+git config core.hooksPath "$HOOKS_DIR"
+
+echo "✅ core.hooksPath = $(git config core.hooksPath)"
+echo ""
+echo "Installed hooks:"
+for hook in "$HOOKS_DIR"/*; do
+  [ -f "$hook" ] || continue
+  echo "  - $(basename "$hook")"
+done
+echo ""
+echo "pre-commit runs: bun run check, bun run test, check-frontend-boundary.sh,"
+echo "and cargo fmt --check when src-tauri/ is staged."
+echo "Bypass a single commit with: git commit --no-verify"

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/release-notes.test.ts

@@ -0,0 +1,58 @@
+/**
+ * Tests for release-note derivation.
+ *
+ * TRACES: | DR-219 | UT-210
+ *
+ * The bug these were written against: `bun run release:notes v0.9.1..HEAD`
+ * listed *every user requirement in the project* as a feature of the release.
+ * The range contained a repo-wide `prettier --write` sweep, so `git diff
+ * --name-only` reported 199 files, their TRACES comments resolved to nearly the
+ * whole matrix, and the result claimed one release had added the entire
+ * application.
+ *
+ * That mattered more than it looked: build-release.yml now generates the
+ * published release body from this script, so the noise would have shipped.
+ */
+
+import { describe, it, expect } from "vitest";
+import { isCosmeticCommit } from "./release-notes";
+
+describe("isCosmeticCommit", () => {
+  it("treats a formatting sweep as cosmetic", () => {
+    // The actual commit that triggered this.
+    expect(isCosmeticCommit("chore(format): run prettier over src/ and scripts/")).toBe(true);
+    expect(isCosmeticCommit("style: reindent the player module")).toBe(true);
+    expect(isCosmeticCommit("style(player): reindent")).toBe(true);
+  });
+
+  it("treats a lockfile-only dependency bump as cosmetic", () => {
+    // Touches package.json/bun.lock, which carry no TRACES, but a `chore(deps)`
+    // that also edits source would still be caught by that source file.
+    expect(isCosmeticCommit("chore(deps): bump vitest to 4.1.11")).toBe(true);
+  });
+
+  it("does NOT treat ordinary work as cosmetic", () => {
+    expect(isCosmeticCommit("fix(player): restart the hero banner timer")).toBe(false);
+    expect(isCosmeticCommit("feat(updater): in-app update on desktop")).toBe(false);
+    expect(isCosmeticCommit("ci: make the frontend gates real")).toBe(false);
+    expect(isCosmeticCommit("docs: add SECURITY.md")).toBe(false);
+  });
+
+  it("does not mistake a chore that is not formatting for a formatting one", () => {
+    // `chore(release)` bumps versions and must still be attributable; a bare
+    // `chore:` could be anything, so it is NOT skipped by default.
+    expect(isCosmeticCommit("chore(release): v0.9.2")).toBe(false);
+    expect(isCosmeticCommit("chore: tidy up the queue helper")).toBe(false);
+  });
+
+  it("is not fooled by the word format appearing later in a subject", () => {
+    // A real fix to formatting *code* is not a cosmetic commit.
+    expect(isCosmeticCommit("fix(duration): format times over 24 hours correctly")).toBe(false);
+    expect(isCosmeticCommit("feat: add a format picker to settings")).toBe(false);
+  });
+
+  it("handles an empty or malformed subject without throwing", () => {
+    expect(isCosmeticCommit("")).toBe(false);
+    expect(isCosmeticCommit("   ")).toBe(false);
+  });
+});

scripts/release-notes.ts

@@ -54,13 +54,86 @@ function loadRequirementDescriptions(): Map<string, string> {
   return map;
 }
 
+/**
+ * Commit subjects whose changes carry no requirement meaning.
+ *
+ * `chore(format)` / `style` rewrite files without changing behaviour;
+ * `chore(deps)` moves lockfiles. Anything else — including a bare `chore:` and
+ * `chore(release):` — is assumed to mean something and is kept.
+ *
+ * Anchored at the start of the subject on purpose: "fix(duration): format times
+ * over 24 hours" is a real fix to formatting *code*, not a formatting commit.
+ */
+const COSMETIC_SUBJECT = /^(chore\(format\)|chore\(deps\)|style)(\([^)]*\))?\s*:/i;
+
+/**
+ * Does this commit subject describe a change with no requirement meaning?
+ *
+ * Exported for scripts/release-notes.test.ts.
+ *
+ * TRACES: | DR-219
+ */
+export function isCosmeticCommit(subject: string): boolean {
+  return COSMETIC_SUBJECT.test(subject.trim());
+}
+
+/**
+ * Files the range changed, excluding those touched only by cosmetic commits.
+ *
+ * Why not a plain `git diff --name-only <range>`: that is what this did, and a
+ * single repo-wide `prettier --write` inside the range made it report 199 files
+ * whose TRACES comments resolved to nearly the entire requirement matrix. The
+ * generated notes for v0.9.2 claimed the release had added the whole
+ * application — and build-release.yml publishes this output, so the noise would
+ * have shipped.
+ *
+ * Walking commit by commit and skipping the cosmetic ones keeps a file that a
+ * sweep *and* a real change both touched: it is still listed by the real
+ * commit. Only files touched exclusively by cosmetic commits drop out, which is
+ * exactly the intent.
+ *
+ * Merge commits produce no output from `git diff-tree` without `-m`, and are
+ * skipped deliberately: everything they merge is already in the range as its
+ * own commit, so including them would double-count.
+ */
 function changedFiles(range: string): string[] {
-  const cmd = range
-    ? `git diff --name-only ${range}`
-    : "git ls-files"; // untagged repo: describe everything currently traced
-  return sh(cmd)
-    .split("\n")
-    .filter((f) => f && existsSync(f));
+  // Untagged repo: describe everything currently traced.
+  if (!range) {
+    return sh("git ls-files")
+      .split("\n")
+      .filter((f) => f && existsSync(f));
+  }
+
+  // NUL between hash and subject so a subject containing anything at all is safe.
+  const log = sh(`git log --no-merges --format=%H%x00%s ${range}`);
+  if (!log) return [];
+
+  const files = new Set<string>();
+  let skipped = 0;
+
+  for (const line of log.split("\n")) {
+    const [sha, ...subjectParts] = line.split("\u0000");
+    const subject = subjectParts.join("\u0000");
+    if (!sha) continue;
+
+    if (isCosmeticCommit(subject)) {
+      skipped++;
+      continue;
+    }
+
+    for (const f of sh(`git diff-tree --no-commit-id --name-only -r ${sha}`).split("\n")) {
+      if (f && existsSync(f)) files.add(f);
+    }
+  }
+
+  if (skipped > 0) {
+    // Say what was dropped rather than silently reporting a smaller set.
+    console.error(
+      `ℹ️  Skipped ${skipped} cosmetic commit(s) (formatting/deps) when deriving notes.`,
+    );
+  }
+
+  return [...files];
 }
 
 /** Collect requirement IDs referenced by TRACES comments in the given files. */
@@ -134,4 +207,7 @@ function main() {
   console.log(out.join("\n"));
 }
 
-main();
+// Guarded so this module stays importable from release-notes.test.ts.
+if (import.meta.main) {
+  main();
+}

scripts/restore-ownership.sh

@@ -0,0 +1,53 @@
+#!/usr/bin/env bash
+# Give build artifacts back to the human who owns the working tree.
+#
+# TRACES: | DR-213
+#
+# The containerised builds (docker-compose.yml: desktop-linux-build,
+# windows-cross, android-build, test, dev) bind-mount the repo at /app and run
+# as root, because their caches live at /root/.cargo and /root/.bun. Everything
+# they write into src-tauri/target and dist/ is therefore root-owned *on the
+# host* — and it accumulates: one audit found 11,124 such files, which is enough
+# to make `cargo clean` and scripts/clean.sh fail with EACCES for the developer.
+# Worse, a plain `cargo build` then dies part-way through, because build scripts
+# compile for the host and land in target/debug even during a cross-build.
+#
+# Running the containers as the host uid would be the tidier fix, but it needs
+# the cache volumes relocated off /root first. Until that happens, this restores
+# ownership at the end of each containerised build, which is self-healing and
+# needs no uid plumbing on the host side.
+#
+# Outside a container this is a no-op: it exits immediately unless it is running
+# as root, so the native build scripts can call it unconditionally.
+
+set -uo pipefail
+
+# Not root (a normal developer build) — nothing to fix, and nothing we may fix.
+[ "$(id -u)" -eq 0 ] || exit 0
+
+cd "$(dirname "$0")/.."
+REPO_ROOT="$(pwd)"
+
+# Whoever owns the checkout is who the artifacts should belong to. Reading it
+# from the tree means this works for any uid/gid without being told, including
+# CI runners whose uid we do not control.
+OWNER="$(stat -c '%u:%g' "$REPO_ROOT")"
+
+# uid 0 owning the tree means it is not a bind mount from a normal host account
+# (a root-owned checkout, or a CI image that clones as root). Nothing to give back.
+if [ "${OWNER%%:*}" = "0" ]; then
+  exit 0
+fi
+
+echo ""
+echo "🔑 Restoring ownership of build artifacts to ${OWNER}…"
+
+for target in src-tauri/target src-tauri/gen dist build node_modules .svelte-kit; do
+  [ -e "$REPO_ROOT/$target" ] || continue
+  chown -R "$OWNER" "$REPO_ROOT/$target" 2>/dev/null || {
+    echo "⚠️  Could not fully chown $target — you may need:"
+    echo "     sudo chown -R $OWNER $REPO_ROOT/$target"
+  }
+done
+
+echo "✅ Ownership restored."

scripts/set-version.sh

@@ -0,0 +1,134 @@
+#!/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
+
+# PKGBUILD — the Arch package version. Easy to miss because Arch packaging is a
+# separate path from the tauri bundler, and missing it is exactly the failure
+# this script exists to prevent: pkgver sat at 0.0.18 while the rest of the tree
+# had moved on, so `makepkg` produced a package whose version bore no relation
+# to the source it was built from. `pkgrel` resets to 1 because a new upstream
+# version starts its packaging revisions over.
+if [ -f packaging/arch/PKGBUILD ]; then
+  # Arch pkgver may not contain a hyphen (it separates pkgver from pkgrel), so a
+  # dev version like 0.9.0-3-gabc1234 becomes 0.9.0.r3.gabc1234, per the VCS
+  # package guidelines.
+  ARCH_VERSION="$(echo "$VERSION" | sed 's/-\([0-9]*\)-g/.r\1.g/; s/-/_/g')"
+  perl -0pi -e 's/^pkgver=.*$/pkgver='"$ARCH_VERSION"'/m' packaging/arch/PKGBUILD
+  perl -0pi -e 's/^pkgrel=.*$/pkgrel=1/m' packaging/arch/PKGBUILD
+fi
+
+# --- Android versionCode ----------------------------------------------------
+# Only when the generated Android project exists (i.e. after `tauri android
+# init`); on Linux/Windows jobs there is nothing to stamp.
+#
+# `tauri android init` derives a versionCode from the semver (0.0.15 -> 15).
+# That is both tiny and NOT monotonic across our history: earlier local/dev
+# builds shipped versionCode 1000 (from a 0.1.0 config), so a plain 15 is a
+# *downgrade* and Android refuses the update.
+#
+# The floor has to clear the highest code actually in the field, which is not the
+# same as the highest this formula has produced. v0.5.2 shipped versionCode
+# **5002** under an earlier `minor*1000` scheme; the `minor*100` formula that
+# replaced it yields only 1502 for that same version, and 1503 for 0.5.3 — so
+# every 0.5.x release built from it was an un-installable downgrade for anyone
+# already on v0.5.2, which is exactly the failure this block exists to prevent.
+# The multipliers are widened and the floor raised past 5002 accordingly.
+#
+#   code = 10000 + major*1000000 + minor*1000 + patch
+# e.g. 0.0.14 -> 10014, 0.1.0 -> 11000, 0.5.3 -> 15003, 1.0.0 -> 1010000.
+PROPS="src-tauri/gen/android/app/tauri.properties"
+if [ -f "$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 +1,228 @@
+/**
+ * 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",
+  );
+  fs.mkdirSync(path.join(dir, "packaging", "arch"), { recursive: true });
+  fs.writeFileSync(
+    path.join(dir, "packaging", "arch", "PKGBUILD"),
+    ["pkgname=jellytau", "pkgver=0.0.1", "pkgrel=3", 'pkgdesc="x"', ""].join("\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 "${GITHUB_REF#refs/tags/}" unconditionally; on a branch build
+    // that is still a full ref, and must not fail the job.
+    it("falls back to a dev version when handed a non-tag ref", () => {
+      const out = run("refs/heads/master");
+      expect(out).not.toMatch(/refs\/heads/);
+      expect(JSON.parse(read("package.json")).version).not.toBe("0.0.1");
+    });
+  });
+
+  // The Arch package is built by makepkg, not the tauri bundler, so its version
+  // lives in a file the rest of the release path never touches. It sat at
+  // 0.0.18 while the tree was on 0.8.x — makepkg happily produced a package
+  // whose version bore no relation to the source it was built from, which is
+  // the exact failure this script was written to prevent.
+  describe("PKGBUILD", () => {
+    it("stamps pkgver and resets pkgrel", () => {
+      run("0.9.0");
+      const pkgbuild = read("packaging/arch/PKGBUILD");
+      expect(pkgbuild).toMatch(/^pkgver=0\.9\.0$/m);
+      // A new upstream version starts its packaging revisions over.
+      expect(pkgbuild).toMatch(/^pkgrel=1$/m);
+    });
+
+    it("converts a dev version into a pkgver Arch accepts", () => {
+      // pkgver may not contain a hyphen — it is the pkgver/pkgrel separator.
+      run("0.9.0-3-gabc1234");
+      const pkgbuild = read("packaging/arch/PKGBUILD");
+      const match = pkgbuild.match(/^pkgver=(.*)$/m);
+      expect(match).not.toBeNull();
+      expect(match![1]).not.toContain("-");
+      expect(match![1]).toBe("0.9.0.r3.gabc1234");
+    });
+  });
+});

scripts/sync-android-sources.sh

@@ -23,6 +23,19 @@ rm -rf "$TARGET_DIR/player" "$TARGET_DIR/security"
 cp -r "$SOURCE_DIR/player" "$TARGET_DIR/"
 cp -r "$SOURCE_DIR/security" "$TARGET_DIR/"
 
+# JVM unit tests (src/test). Plain JUnit over the pure decision helpers — no
+# Android framework classes — run with `./gradlew :app:testDebugUnitTest` from
+# gen/android. Mirrored here so the canonical tree stays the only place tests
+# are edited.
+TEST_SOURCE_DIR="$PROJECT_ROOT/src-tauri/android/src/test/java/com/dtourolle/jellytau"
+TEST_TARGET_DIR="$PROJECT_ROOT/src-tauri/gen/android/app/src/test/java/com/dtourolle/jellytau"
+if [ -d "$TEST_SOURCE_DIR" ]; then
+    rm -rf "$TEST_TARGET_DIR"
+    mkdir -p "$TEST_TARGET_DIR"
+    cp -r "$TEST_SOURCE_DIR"/. "$TEST_TARGET_DIR/"
+    echo "  Copied unit tests: src/test"
+fi
+
 # Copy individual Kotlin files (like VideoOverlayManager.kt)
 for kt_file in "$SOURCE_DIR"/*.kt; do
     if [ -f "$kt_file" ]; then
@@ -87,6 +100,16 @@ if [ -d "$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 +131,26 @@ 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/tauri-security-config.test.ts

@@ -0,0 +1,92 @@
+/**
+ * Guards the shipped webview security configuration.
+ *
+ * `csp` was `null` and the asset protocol was scoped to the whole storage root,
+ * which is the directory holding the SQLite database and the encrypted-token
+ * fallback file. Both are one-character regressions away and neither is visible
+ * in any behavioural test, so they are asserted here instead: the restrictive
+ * half of the policy must stay restrictive, and the permissive half must keep
+ * the schemes playback actually needs.
+ *
+ * TRACES: UR-012, UR-071 | DR-198 | UT-193
+ */
+
+import { describe, it, expect } from "vitest";
+import { readFileSync } from "fs";
+import { resolve } from "path";
+
+const config = JSON.parse(
+  readFileSync(resolve(__dirname, "../src-tauri/tauri.conf.json"), "utf-8"),
+);
+
+const security = config.app.security;
+
+/** Split a CSP string into `directive -> sources`. */
+function directives(csp: string): Record<string, string[]> {
+  const map: Record<string, string[]> = {};
+  for (const part of csp.split(";")) {
+    const [name, ...sources] = part.trim().split(/\s+/);
+    if (name) map[name] = sources;
+  }
+  return map;
+}
+
+describe("tauri.conf.json CSP", () => {
+  it("is set at all — a null CSP hands any injected script the full IPC surface", () => {
+    expect(typeof security.csp).toBe("string");
+    expect(security.csp.length).toBeGreaterThan(0);
+  });
+
+  const csp = directives(security.csp as string);
+
+  it("locks down script execution", () => {
+    // Tauri injects a nonce for SvelteKit's inline bootstrap script at build
+    // time, so 'self' alone is enough and inline/eval must never be re-added.
+    expect(csp["script-src"]).toEqual(["'self'"]);
+    expect(csp["object-src"]).toEqual(["'none'"]);
+    expect(csp["frame-src"]).toEqual(["'none'"]);
+    expect(csp["base-uri"]).toEqual(["'self'"]);
+    expect(csp["default-src"]).toEqual(["'self'"]);
+  });
+
+  it("keeps the schemes playback and thumbnails depend on", () => {
+    // The asset protocol under both names convertFileSrc emits.
+    expect(csp["img-src"]).toContain("asset:");
+    expect(csp["img-src"]).toContain("http://asset.localhost");
+    expect(csp["media-src"]).toContain("asset:");
+    // hls.js: MSE object URLs, and its demuxer worker built from a blob.
+    expect(csp["media-src"]).toContain("blob:");
+    expect(csp["worker-src"]).toContain("blob:");
+    // The token-guarded loopback media server (DR-137).
+    expect(csp["media-src"]).toContain("http://127.0.0.1:*");
+    // Tauri's invoke transport.
+    expect(csp["connect-src"]).toContain("ipc:");
+    expect(csp["connect-src"]).toContain("http://ipc.localhost");
+    // The user's Jellyfin server: an arbitrary run-time origin, http on a LAN.
+    for (const directive of ["img-src", "media-src", "connect-src"]) {
+      expect(csp[directive]).toContain("http:");
+      expect(csp[directive]).toContain("https:");
+    }
+  });
+
+  it("never widens a data directive into script execution", () => {
+    for (const [name, sources] of Object.entries(csp)) {
+      if (name === "script-src" || name === "worker-src") {
+        expect(sources).not.toContain("'unsafe-eval'");
+        expect(sources).not.toContain("'unsafe-inline'");
+      }
+      // A bare `*` would re-admit every scheme, including file:.
+      expect(sources).not.toContain("*");
+    }
+  });
+});
+
+describe("tauri.conf.json asset protocol scope", () => {
+  const scope: string[] = security.assetProtocol.scope;
+
+  it("covers only the thumbnail cache, not the storage root", () => {
+    expect(scope).toEqual(["$APPDATA/thumbnails/**"]);
+    // The database and the encrypted-token fallback live directly in $APPDATA.
+    expect(scope).not.toContain("$APPDATA/**");
+  });
+});