Release 0.16.0

The manual described the review (7c9a4ee) without a picture, because the
demo library holds no duplicates. The new duplicates scene makes two:
it copies two New York frames into a bck folder beside their own,
restarts the app so the scan finds them, waits for the sidebar row the
sweep's dating brings (54aee50), opens the review from it, presses
Check and takes the page. It deletes the copies and restarts on the
library as it was; record.sh's snapshot restore would remove them too.
It is registered last, so no other scene sees the copies.

The picture shows both groups proved the same file, the camera-named
copy outside bck marked Stays, and "Move 2 copies to trash" ready.

.gitattributes

@@ -23,6 +23,9 @@ fixtures/** filter=lfs diff=lfs merge=lfs -text
 
 # The manual's pictures live in LFS for the same reason the models do: a
 # screenshot or a GIF changes wholesale when the interface it shows changes,
-# and every re-recording would otherwise stay in every clone for good. CI's
-# pulls exclude the directory; nothing built or tested reads it.
+# and every re-recording would otherwise stay in every clone for good. The
+# desktop and benchmark legs exclude the directory, since nothing they build
+# or test reads it; the Android and Windows legs fetch it, because the APK
+# and the installer carry the manual (docs/manual/index.html) with its
+# pictures, and their packagers refuse a pointer.
 docs/manual/media/** filter=lfs diff=lfs merge=lfs -text

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

@@ -7,6 +7,10 @@ name: Build and test
 on:
   push:
     branches: [main, master, develop]
+    # A release tag builds again and publishes what it built (the `release`
+    # job at the end). The master push of the same commit has usually filled
+    # the caches, so the second run is the warm one.
+    tags: ['v*']
   pull_request:
     branches: [main, master, develop]
 
@@ -96,7 +100,9 @@ jobs:
             | while read -r key; do git config --local --unset-all "$key"; done || true
           git config --local lfs.url \
             "https://x-access-token:${LFS_TOKEN}@gitea.tourolle.paris/dtourolle/DarkRoom.git/info/lfs"
-          git lfs pull --exclude="fixtures/**,docs/manual/media/**"
+          # The manual's pictures too: the APK carries the manual, and
+          # assemble-apk.sh refuses a pointer where a picture should be.
+          git lfs pull --exclude="fixtures/**"
           ls -lR models/
 
       - name: Cache cargo
@@ -154,6 +160,16 @@ jobs:
       - name: Build
         run: cargo build --workspace --release
 
+      # Only on a release tag: the binary is 150 MB and nothing but the
+      # release job wants it.
+      - name: Upload the desktop binary
+        if: startsWith(github.ref, 'refs/tags/v')
+        uses: actions/upload-artifact@v3
+        with:
+          name: darkroom-desktop-x86_64-linux
+          path: target/release/darkroom-desktop
+          if-no-files-found: error
+
       - name: Disk after
         if: always()
         run: df -h /workspace 2>/dev/null || df -h .
@@ -213,7 +229,9 @@ jobs:
             | while read -r key; do git config --local --unset-all "$key"; done || true
           git config --local lfs.url \
             "https://x-access-token:${LFS_TOKEN}@gitea.tourolle.paris/dtourolle/DarkRoom.git/info/lfs"
-          git lfs pull --exclude="fixtures/**,docs/manual/media/**"
+          # The manual's pictures too: the APK carries the manual, and
+          # assemble-apk.sh refuses a pointer where a picture should be.
+          git lfs pull --exclude="fixtures/**"
           ls -lR models/
 
       - name: Cache cargo
@@ -406,7 +424,9 @@ jobs:
             | while read -r key; do git config --local --unset-all "$key"; done || true
           git config --local lfs.url \
             "https://x-access-token:${LFS_TOKEN}@gitea.tourolle.paris/dtourolle/DarkRoom.git/info/lfs"
-          git lfs pull --exclude="fixtures/**,docs/manual/media/**"
+          # The manual's pictures too: the installer carries the manual, and
+          # package.sh refuses a pointer where a picture should be.
+          git lfs pull --exclude="fixtures/**"
           ls -l models/face models/scene
 
       - name: Cache cargo
@@ -460,6 +480,11 @@ jobs:
           WANT=$(find models/face models/scene models/inpaint -maxdepth 1 -type f ! -name README.md | wc -l)
           GOT=$(ls "$INST/models" | wc -l)
           [ "$GOT" = "$WANT" ] || { echo "FAIL: expected $WANT model files, installed $GOT"; exit 1; }
+          # The manual, and every picture it shows, counted the same way.
+          [ -f "$INST/manual/index.html" ] || { echo "FAIL: no manual installed"; exit 1; }
+          WANT=$(ls docs/manual/media | wc -l)
+          GOT=$(ls "$INST/manual/media" | wc -l)
+          [ "$GOT" = "$WANT" ] || { echo "FAIL: expected $WANT manual pictures, installed $GOT"; exit 1; }
           wine reg query 'HKCU\Software\Microsoft\Windows\CurrentVersion\Uninstall\DarkRoom' 2>/dev/null \
             | grep -q DisplayVersion || { echo "FAIL: no uninstall registry key"; exit 1; }
           wine "$INST/darkroom.exe" --version 2>/dev/null | grep -q '^darkroom-desktop ' \
@@ -519,3 +544,47 @@ jobs:
             fi
           done
           exit $FAILED
+
+  # A v* tag becomes a Gitea Release carrying the three builds and their
+  # SHA256SUMS, titled and described by the tag's message. Until this job
+  # existed every release was made by hand, and most tags never got one.
+  #
+  # It needs all three platform jobs, so a tag whose tests fail publishes
+  # nothing; re-run the failed job and this one follows. The work is
+  # tools/publish-release.sh, which is also how a release is finished by hand.
+  release:
+    if: startsWith(github.ref, 'refs/tags/v')
+    needs: [desktop, android, windows]
+    runs-on: linux/amd64
+    name: Publish the release
+    container:
+      image: catthehacker/ubuntu:act-latest
+    permissions:
+      contents: write
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Fetch the builds
+        uses: actions/download-artifact@v3
+        with:
+          path: dist
+
+      # Named for the download page, with the version in each name the way
+      # the hand-made releases had them. The installer already carries its
+      # version from package.sh.
+      - name: Publish
+        env:
+          GITEA_TOKEN: ${{ secrets.GITEA_TOKEN || github.token }}
+          TAG: ${{ github.ref_name }}
+        run: |
+          set -e
+          V="${TAG#v}"
+          ls -lR dist
+          mkdir -p out
+          cp dist/darkroom-arm64-v8a-apk/darkroom.apk "out/darkroom-${V}-arm64-v8a.apk"
+          cp dist/darkroom-desktop-x86_64-linux/darkroom-desktop "out/darkroom-desktop-${V}-x86_64-linux"
+          chmod +x "out/darkroom-desktop-${V}-x86_64-linux"
+          cp dist/darkroom-windows-x86_64-setup/DarkRoom-${V}-x86_64-setup.exe out/
+          bash tools/publish-release.sh "$TAG" out/*

.gitea/workflows/traceability-check.yml

@@ -67,6 +67,12 @@ jobs:
       # threshold: zero requirements parsed, zero files scanned, a ratio above
       # 100%, or any orphan tag all fail the build. A misconfigured run must not
       # report a plausible-looking 0%.
+      # Every picture the manual shows is made by a scene in
+      # tools/manual/scenes.py, and every picture a scene makes is shown.
+      # Two files read; no app, no display.
+      - name: Manual pictures have scenes
+        run: tools/manual/record.sh --check
+
       - name: Traceability gate
         run: cargo run -q -p traceability -- check
 
@@ -91,10 +97,19 @@ jobs:
       # they will conclude the application is broken rather than the page.
       #
       # This also fails on a malformed tag, so a typo costs a gesture its
-      # desktop half loudly rather than silently.
+      # desktop half loudly rather than silently — and on a key a Slint
+      # handler binds that no tag names, or a key a tag names that no handler
+      # binds (tools/traceability/src/keymap.rs).
       - name: Regenerate the gesture vocabulary and check it is committed
         run: cargo run -q -p traceability -- gestures-check
 
+      # The manual's page, which the packages carry and the help sheet links
+      # into. Blocking for the gesture book's reason: it is shown to the user,
+      # and a page that disagrees with the README is a manual describing an
+      # application that no longer exists.
+      - name: Regenerate the manual page and check it is committed
+        run: cargo run -q -p traceability -- manual-check
+
       # Advisory, not blocking: not every file implements a requirement, and a
       # tag on every function is noise that rots faster than it helps. Tag the
       # unit that decides.

.githooks/pre-commit

@@ -26,7 +26,7 @@ fi
 # The artefacts are generated from the tree, so regenerating them because one
 # was itself edited would be circular.
 case "$(tr -d '[:space:]' <<< "${staged}")" in
-    docs/dev/traceability.md | docs/gestures.md | ui/dr-ui/src/gesture_book.rs)
+    docs/dev/traceability.md | docs/gestures.md | ui/dr-ui/src/gesture_book.rs | docs/manual/index.html)
         exit 0
         ;;
 esac
@@ -65,3 +65,17 @@ for f in docs/gestures.md ui/dr-ui/src/gesture_book.rs; do
         echo "pre-commit: regenerated ${f} and staged it"
     fi
 done
+
+# The manual's page, when its source is part of the commit. Rendered from
+# nothing but the README, so there is no reason to pay for it otherwise.
+if grep -qx 'docs/manual/README.md' <<< "${staged}"; then
+    if ! out="$(cargo run -q -p traceability -- manual 2>&1)"; then
+        echo "pre-commit: the manual would not render" >&2
+        echo "${out}" >&2
+        exit 1
+    fi
+    if ! git diff --quiet -- docs/manual/index.html; then
+        git add docs/manual/index.html
+        echo "pre-commit: regenerated docs/manual/index.html and staged it"
+    fi
+fi

CLAUDE.md

@@ -31,6 +31,30 @@ over a result set — including `deep_count` per sidebar row, which is fine
 at sidebar scale and would not be at grid scale. Aggregate in one
 statement and look up in memory.
 
+**SQL text in a loop is a prepare in a loop.** rusqlite's `execute` and
+`query_row` compile their statement on every call. A loop that calls them
+per row pays a prepare per row even when each query is a primary-key seek:
+the merge of a synced catalog prepared four statements for each of 13,000
+incoming faces (450 ms of a pass that changed nothing), `persist` four per
+photograph a scan listed (1.5 s for a first scan), the shard sync one per
+image each way. Hoist the statement, use `prepare_cached`, or — better, when
+the loop asks the same table about every row — read that table once into a
+map. And do not rewrite a row with what it already holds: an upsert of
+identical values still dirties the page.
+
+**A `LIKE` is case-insensitive, and no index here serves that.**
+`source_ref LIKE 'stem.%'` read every name of the root per sidecar a pull
+took in. When the check that decides is exact, spell the prefix as a range
+(`>= 'stem.' AND < 'stem/'`, `/` being the byte after `.`), which the
+`(root_id, source_ref)` key answers with a seek.
+
+**`Catalog::open` is not free, and every worker thread calls it.** The
+backfill runs on every open, and the develop view opens a catalog to fetch
+each original and again for each neighbour it prefetches. Keep each
+backfill step's no-op case to a read of the small side — the unpaired
+JPEGs, not every RAW; the distinct keywords, not every assignment — and
+measure an open with `catalog_bench` after adding one.
+
 **Filter and aggregate in SQL, and aggregate the small side first.**
 `faces::people` read 19,000 rows, grouped, sorted them by name, and the
 screen threw 17,000 away (empty unnamed groups). `people_in_use` filters in
@@ -124,6 +148,15 @@ other builds are running on the machine — the wall clock doubles under
 load, the CPU figure does not. Keep the binary from before the change and
 run both back to back rather than trusting numbers taken an hour apart.
 
+`cargo run --release -p dr-catalog --example catalog_bench -- CATALOG
+[FACES_DIR]` does the same for opening the catalog (the backfill step by
+step), the upload snapshot, a merge, and the face shard export and import;
+`persist_bench`, an ignored test in `dr-ui`'s scan module, replays a scan's
+`persist` and a sidecar pull (`DR_BENCH_CATALOG=copy.sqlite cargo test
+--release -p dr-ui --lib persist_bench -- --ignored --nocapture
+--test-threads=1`). Both take copies; hand `catalog_bench` a copy of the
+face store directory too.
+
 Reference figures from the 2026-09-19 fixes, largest person (754 faces),
 24k images, 19k faces, before → after. What one click read: `load_people`
 22 ms → 12 ms, `load_faces` 316 ms → 2.4 ms, `audit` 190 ms → not run

CONTRIBUTING.md

@@ -130,6 +130,24 @@ strength of plumbing a future feature would use. So: **close a requirement
 with a test that would fail if the behaviour were removed.** Coverage that
 moves slowly and means something beats coverage that moves quickly.
 
+**Keys and gestures are held the same way.** A key handler in Slint compares
+one canonical chord, `Keys.chord(event) == "Ctrl+Z"`, under a `// KEYMAP:`
+comment naming its section of the gesture book, and every key it binds must be
+named by a `GESTURE:` block beside it. `cargo run -p traceability -- gestures`
+regenerates [`docs/gestures.md`](docs/gestures.md) and the in-app help sheet
+from those blocks; `-- gestures-check` fails when a handler binds a key no tag
+names, or a tag names a key no handler binds. A `manual:` field in a block
+links the gesture to a section of the manual, and a heading that is not there
+fails the scan.
+
+**The manual is checked too.** `docs/manual/index.html` is rendered from
+`docs/manual/README.md` by `-- manual` and `-- manual-check` fails when they
+differ; `tools/manual/record.sh --check` fails when the manual shows a picture
+no scene in `tools/manual/scenes.py` makes. If you change what a pictured
+screen looks like, [`tools/manual`](tools/manual/README.md) says how to record
+it again. The pre-commit hook regenerates the matrix, the gesture book and the
+page; CI runs all three checks.
+
 ## Two invariants the build defends
 
 Worth knowing before you trip one, because both failures name a requirement

Cargo.lock

@@ -1221,7 +1221,7 @@ checksum = "f27ae1dd37df86211c42e150270f82743308803d90a6f6e6651cd730d5e1732f"
 
 [[package]]
 name = "darkroom-android"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "android_logger",
  "dr-plat",
@@ -1234,7 +1234,7 @@ dependencies = [
 
 [[package]]
 name = "darkroom-desktop"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "anyhow",
  "dr-plat",
@@ -1408,7 +1408,7 @@ checksum = "d8b14ccef22fc6f5a8f4d7d768562a182c04ce9a3b3157b91390b52ddfdf1a76"
 
 [[package]]
 name = "dr-bench"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "anyhow",
  "dr-catalog",
@@ -1425,7 +1425,7 @@ dependencies = [
 
 [[package]]
 name = "dr-catalog"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-face",
  "dr-plat",
@@ -1440,7 +1440,7 @@ dependencies = [
 
 [[package]]
 name = "dr-decode"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-types",
  "env_logger",
@@ -1454,7 +1454,7 @@ dependencies = [
 
 [[package]]
 name = "dr-export"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-decode",
  "dr-gpu",
@@ -1473,7 +1473,7 @@ dependencies = [
 
 [[package]]
 name = "dr-face"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-inference-engine",
  "env_logger",
@@ -1486,7 +1486,7 @@ dependencies = [
 
 [[package]]
 name = "dr-film"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "log",
  "serde",
@@ -1495,7 +1495,7 @@ dependencies = [
 
 [[package]]
 name = "dr-gpu"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "bytemuck",
  "dr-decode",
@@ -1513,7 +1513,7 @@ dependencies = [
 
 [[package]]
 name = "dr-inference-engine"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "env_logger",
  "libloading",
@@ -1528,7 +1528,7 @@ dependencies = [
 
 [[package]]
 name = "dr-ingest"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-plat",
  "dr-types",
@@ -1540,7 +1540,7 @@ dependencies = [
 
 [[package]]
 name = "dr-lens"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "lensfun",
  "log",
@@ -1548,7 +1548,7 @@ dependencies = [
 
 [[package]]
 name = "dr-pano"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-decode",
  "dr-inference-engine",
@@ -1562,7 +1562,7 @@ dependencies = [
 
 [[package]]
 name = "dr-pipeline"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-types",
  "log",
@@ -1571,7 +1571,7 @@ dependencies = [
 
 [[package]]
 name = "dr-plat"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "android-native-keyring-store",
  "dr-types",
@@ -1587,7 +1587,7 @@ dependencies = [
 
 [[package]]
 name = "dr-preset-xmp"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-pipeline",
  "log",
@@ -1597,7 +1597,7 @@ dependencies = [
 
 [[package]]
 name = "dr-segment"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-inference-engine",
  "env_logger",
@@ -1610,7 +1610,7 @@ dependencies = [
 
 [[package]]
 name = "dr-sync"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "async-trait",
  "dr-plat",
@@ -1624,7 +1624,7 @@ dependencies = [
 
 [[package]]
 name = "dr-sync-folder"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "async-trait",
  "dr-sync",
@@ -1636,7 +1636,7 @@ dependencies = [
 
 [[package]]
 name = "dr-sync-nextcloud"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "async-trait",
  "dr-decode",
@@ -1658,7 +1658,7 @@ dependencies = [
 
 [[package]]
 name = "dr-thumbs"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-types",
  "jpeg-encoder",
@@ -1670,7 +1670,7 @@ dependencies = [
 
 [[package]]
 name = "dr-types"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "serde",
  "serde_json",
@@ -1679,7 +1679,7 @@ dependencies = [
 
 [[package]]
 name = "dr-ui"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -1704,6 +1704,7 @@ dependencies = [
  "dr-types",
  "dr-xmp",
  "env_logger",
+ "i-slint-backend-testing",
  "jni 0.22.4",
  "log",
  "ndk-context",
@@ -1713,16 +1714,18 @@ dependencies = [
  "rusqlite",
  "serde_json",
  "serde_norway",
+ "sha2",
  "slint",
  "slint-build",
  "thiserror 2.0.20",
  "tokio",
+ "url",
  "wgpu",
 ]
 
 [[package]]
 name = "dr-xmp"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "dr-types",
  "log",
@@ -2780,6 +2783,18 @@ dependencies = [
  "i-slint-renderer-skia",
 ]
 
+[[package]]
+name = "i-slint-backend-testing"
+version = "1.17.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "521e901e3d47ab829c0ef500c63155776208707cd93259e6a7803ed627fa2786"
+dependencies = [
+ "cfg_aliases",
+ "i-slint-common",
+ "i-slint-core",
+ "vtable",
+]
+
 [[package]]
 name = "i-slint-backend-winit"
 version = "1.17.1"
@@ -2960,8 +2975,6 @@ dependencies = [
 [[package]]
 name = "i-slint-renderer-skia"
 version = "1.17.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7b6eed7f3f0a9a3d3ca6e8b9d4ca233371d989351fdb2a7ab88ec368b99e7b57"
 dependencies = [
  "ash",
  "bytemuck",
@@ -7023,9 +7036,10 @@ checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
 
 [[package]]
 name = "traceability"
-version = "0.14.0"
+version = "0.16.0"
 dependencies = [
  "anyhow",
+ "pulldown-cmark",
  "serde",
  "serde_json",
 ]
@@ -7923,8 +7937,6 @@ dependencies = [
 [[package]]
 name = "wgpu-hal"
 version = "29.0.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "97ace1c17727311c22a46e4e3faf56ea6de81af99dcc839bdfb54857b94d448d"
 dependencies = [
  "android_system_properties",
  "arrayvec",

Cargo.toml

@@ -27,9 +27,12 @@ members = [
     "tools/bench",
     "tools/traceability",
 ]
+# Patched copies of upstream crates, not our code: see third_party/README.md.
+# Excluded so `--workspace` does not test, lint or format them as ours.
+exclude = ["third_party"]
 
 [workspace.package]
-version = "0.14.0"
+version = "0.16.0"
 edition = "2021"
 rust-version = "1.92"
 license = "GPL-3.0-or-later"
@@ -127,6 +130,10 @@ url = "2.5"
 async-trait = "0.1"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
+# The manual's HTML rendering (tools/traceability). Already in the tree as
+# Slint's Markdown parser, so this adds a dependency edge and no crate; only
+# the HTML writer is needed, not the command-line front end.
+pulldown-cmark = { version = "0.13", default-features = false, features = ["html"] }
 base64 = "0.23"
 
 # Display-server clients, for FR-DSP-8's per-display profile acquisition.
@@ -262,3 +269,12 @@ opt-level = 0
 [profile.release]
 lto = "thin"
 codegen-units = 1
+
+# Two upstream crates carry a local patch so that the Android build can draw
+# with wgpu on a rotated display (technical-debt.md TD-1). Both are exact
+# copies of the version the lockfile already resolves, plus that patch;
+# third_party/README.md says what was changed and how to carry it forward
+# when Slint or wgpu moves.
+[patch.crates-io]
+wgpu-hal = { path = "third_party/wgpu-hal-29.0.4" }
+i-slint-renderer-skia = { path = "third_party/i-slint-renderer-skia-1.17.1" }

README.md

@@ -16,18 +16,22 @@ is still missing.
 or one a Nextcloud client keeps in virtual-files mode, where a placeholder
 is treated as the photograph rather than as a one-byte file — or at a
 Nextcloud account directly. The grid is virtualised, ordered by capture
-time with a timeline beside it, and filtered by rating, flag, person and
-whether the file is here. Ratings, keywords, collections and a
-trash that survives a crash mid-operation. Card ingest. Bursts fold. Face
-detection and identity, with the index syncing between devices.
+time with a timeline beside it, and filtered by rating, flag, colour label,
+person and whether the file is here. Ratings, colour labels, keywords,
+collections and a trash that survives a crash mid-operation. Card ingest.
+Bursts fold. The same RAW catalogued twice — a dated folder and a backup
+beside it — is found, proved the same, and folded onto one copy with the
+spares in the trash. Face detection and identity, with the index syncing
+between devices.
 
 **Developing.** Eighteen declared operations fused into one compute
 dispatch, plus the neighbourhood work that cannot be: clarity, texture,
 capture sharpening, noise reduction, lens correction, spectral film
-simulation. Crop and straighten, spot repair, and local adjustments over
-masks the model draws — click a subject or a category, then paint, subtract
-a gradient, grow or shrink the edge. Focus peaking and a raw histogram for
-judging what is recoverable. Named presets; XMP sidecars other editors read.
+simulation. Crop, straighten and correct converging verticals, spot repair,
+and local adjustments over masks the model draws — click a subject or a
+category, then paint, subtract a gradient or keep only where two selections
+agree, grow or shrink the edge. Focus peaking and a raw histogram for judging
+what is recoverable. Named presets; XMP sidecars other editors read.
 
 [![Segmenting an urban scene and choosing the sky as a mask](docs/manual/media/local-segment.png)](docs/manual/README.md#local-adjustments)
 
@@ -37,13 +41,20 @@ sources as a DNG, with a sidecar recording what it was merged from.
 
 [![Twelve hand-held frames aligned on a cylinder](docs/manual/media/panorama-aligned.png)](docs/manual/README.md#merging-a-panorama)
 
+**From the keyboard, and with its manual.** Rating, flagging and labelling
+have keys in the grid and in develop, as do zoom, undo and stepping through a
+shoot in develop, and none of them is keyboard-only. The help sheet (`F1`, or
+`?` in develop) lists every key and gesture, generated from the code that
+binds it, and links them to the sections of the manual that show them — the
+manual ships with the application and opens offline.
+
 **Export.** JPEG, PNG, AVIF, JPEG XL, 8- and 16-bit TIFF, with resize, output
 sharpening, a naming template and a colour space — to a folder here or back
 into the library.
 
 **On both platforms.** The same core runs on a desktop and a 12-inch
 tablet; the interface is one layout, tuned for a wide viewport with touch
-targets throughout. On desktop the develop view draws the compute pass's
+targets throughout. On both, the develop view draws the compute pass's
 texture directly — no readback between the GPU and the screen.
 
 ## Getting it
@@ -76,27 +87,21 @@ controls, its place in the chain and its tests.
 
 ## Where it stands
 
-**0.14.0**, twenty-one tagged releases in. 188 numbered requirements in
-scope, 82% of them claimed by code and [traced to it](docs/dev/traceability.md);
+**0.16.0**, twenty-four tagged releases in. 191 numbered requirements in
+scope, 84% of them claimed by code and [traced to it](docs/dev/traceability.md);
 the rest are written down rather than merely absent.
 
 **Not built:** plugins (post-v1, [D12](docs/dev/requirements.md)), compare and
-survey culling, AI denoise, tiled and progressive rendering, HDR merge and
-focus stacking, most of the Android platform integration beyond running,
-and the Flatpak's library chooser. The performance targets are half
+survey culling, AI denoise, tiled rendering, HDR merge and
+focus stacking, importing a Lightroom or darktable catalog, translations
+beyond the launch screen, most of the Android platform integration beyond
+running, and the Flatpak's library chooser. The performance targets are half
 verified: the per-commit benchmark suite §8 requires exists for everything
 that does not need a frame — the catalog, the scan, the thumbnails — and
 not yet for the render path, so a regression there fails nothing.
 [outstanding.md](docs/dev/outstanding.md) is the list, with the reasoning for
 each.
 
-**The one deliberate compromise worth knowing about before reading
-anything else:** the Android develop view reads its frame back through the
-CPU, because zero-copy there needs wgpu's Vulkan swapchain and that tears a
-portrait window on a tablet whose panel is mounted landscape. It is debt,
-not a revision of the rule — [technical-debt.md TD-1](docs/dev/technical-debt.md)
-has the measurements and the three things any one of which would remove it.
-
 ## Documentation
 
 [docs/README.md](docs/README.md) is the index. The short version, for someone using it:

apps/darkroom-android/android/AndroidManifest.xml

@@ -141,6 +141,23 @@
             </intent-filter>
         </activity>
 
+        <!-- The manual (dr_ui::manual): a WebView over the copy the APK
+             carries in assets/manual. See ManualActivity.java for why it is
+             not the browser.
+
+             Not exported: nothing outside this app has a reason to start it,
+             and dr_ui starts it by class name, which needs no intent filter.
+             Its own task entry is not wanted either — it is a page over the
+             app, and Back returns to the photograph it was opened from.
+             configChanges so a rotation reflows the page rather than
+             reloading it at the top. -->
+        <activity
+            android:name="paris.tourolle.darkroom.ManualActivity"
+            android:exported="false"
+            android:label="DarkRoom manual"
+            android:theme="@style/ManualTheme"
+            android:configChanges="orientation|keyboardHidden|screenSize|screenLayout|uiMode" />
+
         <!-- FR-PLAT-AND-6, outbound. Android has refused file:// URIs
              between apps since API 24 — handing one out raises
              FileUriExposedException in *this* process — so an exported JPEG

apps/darkroom-android/android/java/paris/tourolle/darkroom/ManualActivity.java

@@ -0,0 +1,105 @@
+package paris.tourolle.darkroom;
+
+import android.app.Activity;
+import android.content.ActivityNotFoundException;
+import android.content.Intent;
+import android.net.Uri;
+import android.os.Bundle;
+import android.webkit.WebResourceRequest;
+import android.webkit.WebSettings;
+import android.webkit.WebView;
+import android.webkit.WebViewClient;
+
+/**
+ * The manual that ships in the APK, shown in a WebView.
+ *
+ * <h2>Why an activity of our own rather than the browser</h2>
+ *
+ * <p>The desktop hands the manual to the system browser. Android leaves no
+ * way to do the same: the page is an asset inside the APK, which is not a
+ * file; an unpacked copy in app-private storage is a file no browser may
+ * read; a {@code file:} URI handed to another app is refused since API 24;
+ * and a {@code content:} URI serves the page but leaves the browser to fetch
+ * every picture by a relative URL against the provider, which browsers do not
+ * reliably do. A WebView reads {@code file:///android_asset/} straight from
+ * the APK, pictures and section anchor included, and nothing is unpacked.
+ *
+ * <h2>What it is not</h2>
+ *
+ * <p>A browser. JavaScript stays off (the page has none), and a link that
+ * leaves the manual — the design documents are on the forge — goes to the
+ * user's browser rather than opening inside this view, so the only thing ever
+ * shown here is the page the APK carries.
+ *
+ * <p>Started by {@code dr_ui::manual} with {@code Intent.setClassName}, so the
+ * name here and there must agree; a test in lib.rs checks the manifest
+ * declares it.
+ */
+public final class ManualActivity extends Activity {
+    /** The section to open at, a heading's anchor. Absent opens the top. */
+    public static final String EXTRA_ANCHOR = "anchor";
+
+    private static final String PAGE = "file:///android_asset/manual/index.html";
+
+    private WebView web;
+
+    @Override
+    protected void onCreate(Bundle saved) {
+        super.onCreate(saved);
+        setTitle("DarkRoom manual");
+
+        web = new WebView(this);
+        WebSettings settings = web.getSettings();
+        settings.setJavaScriptEnabled(false);
+        // Pinch to zoom into a screenshot, which is 1600 pixels wide and drawn
+        // at the width of a phone.
+        settings.setBuiltInZoomControls(true);
+        settings.setDisplayZoomControls(false);
+        web.setWebViewClient(new WebViewClient() {
+            @Override
+            public boolean shouldOverrideUrlLoading(WebView view, WebResourceRequest request) {
+                Uri uri = request.getUrl();
+                if ("file".equals(uri.getScheme())) {
+                    return false;
+                }
+                try {
+                    startActivity(new Intent(Intent.ACTION_VIEW, uri));
+                } catch (ActivityNotFoundException e) {
+                    // No browser on the device: the link does nothing, which
+                    // is all it could do.
+                }
+                return true;
+            }
+        });
+        setContentView(web);
+
+        if (saved != null) {
+            web.restoreState(saved);
+        } else {
+            String anchor = getIntent().getStringExtra(EXTRA_ANCHOR);
+            web.loadUrl(anchor == null || anchor.isEmpty() ? PAGE : PAGE + "#" + anchor);
+        }
+    }
+
+    @Override
+    protected void onSaveInstanceState(Bundle out) {
+        super.onSaveInstanceState(out);
+        web.saveState(out);
+    }
+
+    /** Back walks back through the sections visited, then leaves. */
+    @Override
+    public void onBackPressed() {
+        if (web.canGoBack()) {
+            web.goBack();
+        } else {
+            super.onBackPressed();
+        }
+    }
+
+    @Override
+    protected void onDestroy() {
+        web.destroy();
+        super.onDestroy();
+    }
+}

apps/darkroom-android/android/res/values-v29/themes.xml

@@ -0,0 +1,5 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!-- Day or night as the system is; see values/themes.xml. -->
+<resources>
+    <style name="ManualTheme" parent="@android:style/Theme.DeviceDefault.DayNight" />
+</resources>

apps/darkroom-android/android/res/values/themes.xml

@@ -0,0 +1,10 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!--
+  The manual's theme (ManualActivity). Light below API 29, which has no
+  day-night theme in the platform; values-v29 follows the system from there.
+  The WebView takes prefers-color-scheme from whether this theme is light, and
+  the manual's stylesheet takes its colours from that.
+-->
+<resources>
+    <style name="ManualTheme" parent="@android:style/Theme.DeviceDefault.Light" />
+</resources>

apps/darkroom-android/src/lib.rs

@@ -581,6 +581,37 @@ mod tests {
         );
     }
 
+    /// `dr_ui::manual` starts the manual by class name. A name the manifest
+    /// does not declare is an `ActivityNotFoundException` on the device and a
+    /// Manual button that does nothing, so the three spellings — dr_ui's, the
+    /// manifest's and the Java file's — are checked to be one.
+    #[test]
+    fn the_manual_activity_dr_ui_starts_is_declared() {
+        let manifest = manifest();
+        let wanted = dr_ui::manual::ANDROID_ACTIVITY;
+        let element = manifest
+            .split("<activity")
+            .skip(1)
+            .find(|a| attribute(a, "android:name").as_deref() == Some(wanted))
+            .unwrap_or_else(|| panic!("the manifest declares no activity {wanted}"));
+        assert_eq!(
+            attribute(element, "android:exported").as_deref(),
+            Some("false"),
+            "the manual activity has no reason to be startable by another app"
+        );
+        let java = include_str!("../android/java/paris/tourolle/darkroom/ManualActivity.java");
+        let (package, class) = wanted.rsplit_once('.').expect("unqualified class name");
+        assert!(java.contains(&format!("package {package};")));
+        assert!(java.contains(&format!("class {class} ")));
+        assert!(
+            java.contains(&format!(
+                "EXTRA_ANCHOR = \"{}\"",
+                dr_ui::manual::ANDROID_EXTRA_ANCHOR
+            )),
+            "ManualActivity reads the section from a different extra than dr_ui writes"
+        );
+    }
+
     #[test]
     fn the_provider_hands_out_one_file_at_a_time_and_nothing_by_itself() {
         let manifest = manifest();

apps/darkroom-desktop/Cargo.toml

@@ -25,3 +25,6 @@ winresource = "0.1"
 
 [features]
 default = []
+# The manual's recording hook (dr-ui's `automation`); tools/manual/record.sh
+# builds with it, nothing else does.
+automation = ["dr-ui/automation"]

core/dr-catalog/examples/catalog_bench.rs

@@ -0,0 +1,117 @@
+//! What the catalog's routine reads cost on a real library, off the GUI.
+//!
+//!   cargo run --release -p dr-catalog --example catalog_bench -- CATALOG.sqlite [FACES_DIR]
+//!
+//! Times `Catalog::open` — which every worker thread pays, including the
+//! develop view's fetch of each original and each neighbour it prefetches —
+//! and the backfill that runs inside it, step by step. Run it against a
+//! *copy* of a real catalog: opening migrates and backfills, which write.
+//!
+//! The figures are for reading side by side before and after a change; they
+//! are not a gate. Compare the `cpu` column when the machine is busy.
+
+use std::path::PathBuf;
+use std::time::{Duration, Instant};
+
+use dr_catalog::{keywords, rating, schema, Catalog};
+
+fn main() {
+    let args: Vec<String> = std::env::args().skip(1).collect();
+    let Some(path) = args.first().map(PathBuf::from) else {
+        eprintln!("usage: catalog_bench CATALOG.sqlite");
+        std::process::exit(2);
+    };
+
+    // Once untimed, so a migration or a first backfill is not in the figures.
+    drop(Catalog::open(&path).expect("catalog"));
+
+    time("Catalog::open", 20, || {
+        drop(Catalog::open(&path).unwrap());
+    });
+
+    let catalog = Catalog::open(&path).unwrap();
+    let conn = catalog.connection();
+    time("schema::backfill (all steps)", 20, || {
+        schema::backfill(conn).unwrap();
+    });
+    time("  rating::ensure_default_versions", 20, || {
+        rating::ensure_default_versions(conn).unwrap();
+    });
+    time("  rating::align_default_version_uuids", 20, || {
+        rating::align_default_version_uuids(conn).unwrap();
+    });
+    time("  keywords::adopt_orphan_terms", 20, || {
+        keywords::adopt_orphan_terms(conn).unwrap();
+    });
+
+    // A sync pass: the upload snapshot, then a merge of the catalog with a
+    // copy of itself — every row a match, which is the steady state.
+    let scratch = path.with_extension("bench-snapshot");
+    time("snapshot_for_upload", 3, || {
+        let _ = std::fs::remove_file(&scratch);
+        catalog.snapshot_for_upload(&scratch).unwrap();
+    });
+    println!(
+        "  snapshot size {:.1} MB",
+        std::fs::metadata(&scratch).map(|m| m.len()).unwrap_or(0) as f64 / 1e6
+    );
+    let remote = path.with_extension("bench-remote");
+    let _ = std::fs::remove_file(&remote);
+    conn.execute("VACUUM INTO ?1", [remote.to_string_lossy().as_ref()])
+        .unwrap();
+    time("merge_remote_catalog (self)", 5, || {
+        catalog.merge_remote_catalog(&remote).unwrap();
+    });
+    let _ = std::fs::remove_file(&scratch);
+    let _ = std::fs::remove_file(&remote);
+
+    // The face half of a sync pass, against a copy of the face store: both
+    // directions in the steady state, where nothing is new either way.
+    if let Some(faces) = args.get(1).map(PathBuf::from) {
+        let model = "scrfd_10g+w600k_mbf";
+        let mut store = dr_catalog::FaceShardStore::open(&faces).unwrap();
+        println!(
+            "  first export sent {}, first import adopted {}",
+            dr_catalog::face_shard::export_to_shards(conn, &mut store, model).unwrap(),
+            dr_catalog::face_shard::import_from_shards(conn, &store, model).unwrap()
+        );
+        time("face_shard::export_to_shards (steady)", 5, || {
+            dr_catalog::face_shard::export_to_shards(conn, &mut store, model).unwrap();
+        });
+        time("face_shard::import_from_shards (steady)", 5, || {
+            dr_catalog::face_shard::import_from_shards(conn, &store, model).unwrap();
+        });
+    }
+}
+
+/// Run `f` a few times and print the best wall-clock, the median, and the
+/// best CPU time — the figure to compare across runs on a busy machine.
+fn time(label: &str, runs: usize, mut f: impl FnMut()) {
+    let mut wall: Vec<Duration> = Vec::with_capacity(runs);
+    let mut cpu: Vec<Duration> = Vec::with_capacity(runs);
+    for _ in 0..runs {
+        let c = cpu_now();
+        let t = Instant::now();
+        f();
+        wall.push(t.elapsed());
+        cpu.push(cpu_now().saturating_sub(c));
+    }
+    wall.sort();
+    cpu.sort();
+    println!(
+        "{label:42} best {:8.2} ms   median {:8.2} ms   cpu {:8.2} ms",
+        wall[0].as_secs_f64() * 1e3,
+        wall[runs / 2].as_secs_f64() * 1e3,
+        cpu[0].as_secs_f64() * 1e3
+    );
+}
+
+/// This thread's time on a CPU so far, from `/proc/self/schedstat`; zero where
+/// the file is missing, which only makes the CPU column useless.
+fn cpu_now() -> Duration {
+    std::fs::read_to_string("/proc/self/schedstat")
+        .ok()
+        .and_then(|s| s.split_whitespace().next()?.parse::<u64>().ok())
+        .map(Duration::from_nanos)
+        .unwrap_or_default()
+}

core/dr-catalog/src/error.rs

@@ -96,6 +96,13 @@ pub enum CatalogError {
 
     #[error("io: {0}")]
     Io(String),
+
+    /// TRACES: FR-CAT-11a
+    /// A duplicate group planned earlier no longer holds: a copy was trashed,
+    /// rescanned or changed since the review was drawn. The group is left
+    /// untouched rather than consolidated on a stale plan.
+    #[error("no longer a duplicate: {0}")]
+    StaleDuplicate(String),
 }
 
 impl From<rusqlite::Error> for CatalogError {

core/dr-catalog/src/face_shard.rs

@@ -30,6 +30,7 @@
 //! identity every client agrees on (FR-NC-5), and it survives a server-side
 //! move, so a shard written before a reorganisation still applies after it.
 
+use std::collections::{HashMap, HashSet};
 use std::path::{Path, PathBuf};
 
 use rusqlite::{Connection, OptionalExtension};
@@ -155,6 +156,45 @@ impl FaceShardStore {
             .flatten()
     }
 
+    /// [`indexed_at`](Self::indexed_at) for every entry at once.
+    ///
+    /// What a pass over the whole library asks instead of one lookup per image:
+    /// the export compares every marker in the catalog with this, and a lookup
+    /// each was 19,000 statements prepared and run on every sync pass that had
+    /// nothing to send.
+    fn all_indexed_at(&self) -> Result<HashMap<(u64, String), Option<i64>>, CatalogError> {
+        let mut q = self
+            .index
+            .prepare("SELECT file_id, model_id, indexed_at FROM entries")?;
+        let rows = q.query_map([], |r| {
+            Ok((
+                (r.get::<_, i64>(0)? as u64, r.get::<_, String>(1)?),
+                r.get::<_, Option<i64>>(2)?,
+            ))
+        })?;
+        Ok(rows.collect::<Result<_, _>>()?)
+    }
+
+    /// [`held_model`](Self::held_model) for every file at once: one statement,
+    /// ordered exactly as that one is, keeping the first row per file.
+    fn all_held_models(&self, model_id: &str) -> Result<HashMap<u64, String>, CatalogError> {
+        let mut q = self.index.prepare(&format!(
+            "SELECT file_id, model_id FROM entries
+              WHERE {} = ?1
+              ORDER BY file_id, indexed_at DESC NULLS LAST, model_id",
+            crate::faces::embedder_sql("model_id")
+        ))?;
+        let rows = q.query_map([crate::faces::embedder_of(model_id)], |r| {
+            Ok((r.get::<_, i64>(0)? as u64, r.get::<_, String>(1)?))
+        })?;
+        let mut out = HashMap::new();
+        for row in rows {
+            let (file, model) = row?;
+            out.entry(file).or_insert(model);
+        }
+        Ok(out)
+    }
+
     /// The pipeline this store holds an image under, among those sharing
     /// `model_id`'s embedder — the most recently indexed where a peer has
     /// sent more than one.
@@ -784,6 +824,18 @@ pub fn export_to_shards_reporting(
     /// enough that the reporting is lost in the write it accompanies.
     const REPORT_EVERY: usize = 25;
 
+    // What the store holds, read once. A put below rewrites only its own
+    // file's entries -- its generation, and siblings it supersedes -- so a
+    // file already written in this pass is asked of the store again and every
+    // other answer is the one a lookup would have given.
+    //
+    // An index that cannot be read answers as each lookup did: nothing held.
+    let held = store.all_indexed_at().unwrap_or_else(|e| {
+        log::debug!("reading the shard index: {e}");
+        HashMap::new()
+    });
+    let mut written: HashSet<u64> = HashSet::new();
+
     let total = rows.len();
     let mut exported = 0;
     for (seen, (file_id, image_id, edge, indexed_at, model_id)) in rows.into_iter().enumerate() {
@@ -800,10 +852,14 @@ pub fn export_to_shards_reporting(
         //
         // The comparison is against when the *catalog* indexed it, so a
         // re-index is visible and an unchanged image still costs nothing.
-        if store
-            .indexed_at(file_id as u64, model_id)
-            .is_some_and(|was| was >= indexed_at)
-        {
+        let was = if written.contains(&(file_id as u64)) {
+            store.indexed_at(file_id as u64, model_id)
+        } else {
+            held.get(&(file_id as u64, model_id.to_string()))
+                .copied()
+                .flatten()
+        };
+        if was.is_some_and(|was| was >= indexed_at) {
             continue;
         }
         let mut fq = conn.prepare(
@@ -840,6 +896,7 @@ pub fn export_to_shards_reporting(
             &faces,
             Some(indexed_at),
         )?;
+        written.insert(file_id as u64);
         exported += 1;
     }
     progress(total, total);
@@ -902,6 +959,18 @@ pub fn import_from_shards(
     /// the lock, waits a fraction of a second and not the whole import.
     const CHUNK: usize = 100;
 
+    // Every file's held pipeline, read once rather than asked per candidate --
+    // 23,000 prepared lookups on every pass, nearly all of them for images
+    // this device already holds. Nothing below changes which pipeline the
+    // store holds a file under (`set_indexed_at` touches only a file already
+    // decided), and each candidate is a different file, so these are the
+    // answers the lookups gave.
+    // An index that cannot be read answers as each lookup did: nothing held.
+    let held_models = store.all_held_models(model_id).unwrap_or_else(|e| {
+        log::debug!("reading the shard index: {e}");
+        HashMap::new()
+    });
+
     let mut adopted = 0;
     let mut tx = conn.unchecked_transaction()?;
     let mut in_chunk = 0;
@@ -911,7 +980,7 @@ pub fn import_from_shards(
             tx = conn.unchecked_transaction()?;
             in_chunk = 0;
         }
-        let Some(held) = store.held_model(file_id as u64, model_id) else {
+        let Some(held) = held_models.get(&(file_id as u64)).cloned() else {
             continue;
         };
         if let Some(local) = local {

core/dr-catalog/src/faces.rs

@@ -1599,6 +1599,170 @@ fn iou(a: (f32, f32, f32, f32), b: (f32, f32, f32, f32)) -> f32 {
     }
 }
 
+/// TRACES: FR-CAT-11a | FR-CULL-10
+/// What [`carry_onto_copy`] did with one byte-identical copy's faces.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub struct FaceCarry {
+    /// Faces moved onto the survivor outright, because it had none.
+    pub moved: usize,
+    /// Names and suggestions put on a survivor's face that lacked one.
+    pub named: usize,
+    /// Rejections added to a survivor's face.
+    pub rejections: usize,
+    /// Faces named one person on the copy and another on the survivor. The
+    /// survivor's name is kept; the copy keeps its own, in the trash.
+    pub conflicts: usize,
+    /// Named faces on the copy that match nothing on the survivor. Left with
+    /// the copy rather than mixed into another pipeline's faces.
+    pub unmatched_named: usize,
+}
+
+/// TRACES: FR-CAT-11a | FR-CULL-10
+/// Bring what one copy of a photograph knows about its faces onto another
+/// copy of the same bytes, inside the caller's transaction.
+///
+/// Faces are per image, so two copies of one file indexed separately hold
+/// two sets of the same boxes, and a name confirmed on one is invisible on
+/// the other. The rule is the one [`record_detections_within`] keeps: an
+/// image holds one pipeline's faces at a time, and the user's judgements
+/// are what must survive.
+///
+/// - **The survivor has no faces at all.** The copy's faces and its run
+///   markers move over wholesale — nothing is duplicated, and the survivor
+///   is spared a detection pass it would only repeat. Its own markers go
+///   first, because a marker saying "examined, nothing found" over an image
+///   that now holds faces is the V12 state.
+/// - **The survivor has faces.** Each of the copy's faces is paired with the
+///   survivor face its box overlaps most (IoU above one half — the bytes are
+///   the same, so the boxes coincide). A name or suggestion is carried onto
+///   a survivor face that has none, a confirmation outranks a suggestion,
+///   and two different confirmed names are a conflict the survivor wins.
+///   Rejections are unioned. The copy's faces stay where they are, with the
+///   copy.
+pub fn carry_onto_copy(
+    tx: &Connection,
+    from: ImageId,
+    to: ImageId,
+) -> Result<FaceCarry, CatalogError> {
+    let mut out = FaceCarry::default();
+    let boxes = |image: ImageId| -> Result<Vec<CopyFace>, CatalogError> {
+        let mut q = tx.prepare(
+            "SELECT f.id, f.x, f.y, f.w, f.h, fp.person_id, fp.probability, fp.confirmed
+               FROM faces f
+               LEFT JOIN face_person fp ON fp.face_id = f.id
+              WHERE f.image_id = ?1
+              ORDER BY f.id",
+        )?;
+        let rows = q.query_map([image.0 as i64], |r| {
+            let person: Option<i64> = r.get(5)?;
+            Ok(CopyFace {
+                id: r.get(0)?,
+                rect: (
+                    r.get::<_, f64>(1)? as f32,
+                    r.get::<_, f64>(2)? as f32,
+                    r.get::<_, f64>(3)? as f32,
+                    r.get::<_, f64>(4)? as f32,
+                ),
+                assignment: match person {
+                    Some(p) => Some((p, r.get::<_, f64>(6)?, r.get::<_, i64>(7)? != 0)),
+                    None => None,
+                },
+            })
+        })?;
+        Ok(rows.collect::<Result<Vec<_>, _>>()?)
+    };
+
+    let theirs = boxes(from)?;
+    if theirs.is_empty() {
+        return Ok(out);
+    }
+    let ours = boxes(to)?;
+
+    if ours.is_empty() {
+        tx.execute("DELETE FROM face_index WHERE image_id = ?1", [to.0 as i64])?;
+        tx.execute(
+            "UPDATE face_index SET image_id = ?2 WHERE image_id = ?1",
+            rusqlite::params![from.0 as i64, to.0 as i64],
+        )?;
+        out.moved = tx.execute(
+            "UPDATE faces SET image_id = ?2 WHERE image_id = ?1",
+            rusqlite::params![from.0 as i64, to.0 as i64],
+        )?;
+        return Ok(out);
+    }
+
+    let mut taken = vec![false; ours.len()];
+    for face in &theirs {
+        let best = ours
+            .iter()
+            .enumerate()
+            .filter(|(i, _)| !taken[*i])
+            .map(|(i, o)| (i, iou(face.rect, o.rect)))
+            .filter(|(_, overlap)| *overlap > 0.5)
+            .max_by(|a, b| a.1.total_cmp(&b.1));
+        let Some((at, _)) = best else {
+            if face.assignment.is_some_and(|(_, _, confirmed)| confirmed) {
+                out.unmatched_named += 1;
+            }
+            continue;
+        };
+        taken[at] = true;
+        let target = &ours[at];
+
+        out.rejections += tx.execute(
+            "INSERT OR IGNORE INTO face_person_rejected (face_id, person_id)
+             SELECT ?2, person_id FROM face_person_rejected WHERE face_id = ?1",
+            rusqlite::params![face.id, target.id],
+        )?;
+
+        let Some((person, probability, confirmed)) = face.assignment else {
+            continue;
+        };
+        let carry = match target.assignment {
+            None => true,
+            // The same person: only a confirmation upgrades a suggestion.
+            Some((p, _, theirs_confirmed)) if p == person => confirmed && !theirs_confirmed,
+            // Another person, only suggested there: the user's word wins.
+            Some((_, _, false)) => confirmed,
+            // Another person, confirmed there: the survivor keeps its name.
+            Some((_, _, true)) => {
+                if confirmed {
+                    out.conflicts += 1;
+                }
+                false
+            }
+        };
+        if carry {
+            tx.execute(
+                "INSERT INTO face_person (face_id, person_id, probability, confirmed)
+                 VALUES (?1, ?2, ?3, ?4)
+                 ON CONFLICT(face_id) DO UPDATE SET
+                     person_id = excluded.person_id,
+                     probability = excluded.probability,
+                     confirmed = excluded.confirmed",
+                rusqlite::params![target.id, person, probability, confirmed],
+            )?;
+            // A name the user gave outranks a rejection of the same pair made
+            // on the survivor — the same order `confirm` applies.
+            if confirmed {
+                tx.execute(
+                    "DELETE FROM face_person_rejected WHERE face_id = ?1 AND person_id = ?2",
+                    rusqlite::params![target.id, person],
+                )?;
+            }
+            out.named += 1;
+        }
+    }
+    Ok(out)
+}
+
+/// One face as [`carry_onto_copy`] pairs it.
+struct CopyFace {
+    id: i64,
+    rect: (f32, f32, f32, f32),
+    assignment: Option<(i64, f64, bool)>,
+}
+
 pub(crate) fn now_secs() -> i64 {
     std::time::SystemTime::now()
         .duration_since(std::time::UNIX_EPOCH)

core/dr-catalog/src/jobs.rs

@@ -185,7 +185,8 @@ pub fn enqueue(
     priority: Priority,
     payload: Option<&str>,
 ) -> Result<(), CatalogError> {
-    conn.execute(
+    // Cached: a scan enqueues one per photograph it lists.
+    conn.prepare_cached(
         "INSERT INTO jobs(kind, subject_id, priority, state, payload)
          VALUES (?1, ?2, ?3, 0, ?4)
          ON CONFLICT(kind, subject_id) DO UPDATE SET
@@ -195,8 +196,13 @@ pub fn enqueue(
              state      = CASE WHEN jobs.state = 2 THEN 0 ELSE jobs.state END,
              attempts   = CASE WHEN jobs.state = 2 THEN 0 ELSE jobs.attempts END,
              not_before = CASE WHEN jobs.state = 2 THEN 0 ELSE jobs.not_before END",
-        rusqlite::params![kind as i64, subject_id, priority as i64, payload],
-    )?;
+    )?
+    .execute(rusqlite::params![
+        kind as i64,
+        subject_id,
+        priority as i64,
+        payload
+    ])?;
     Ok(())
 }
 

core/dr-catalog/src/keywords.rs

@@ -491,7 +491,12 @@ pub fn adopt_orphan_terms(conn: &Connection) -> Result<usize, CatalogError> {
             // quietly readmitted to the vocabulary; it stays visible as an
             // orphan in [`for_images`] instead, which is a state someone can
             // see and act on rather than one that silently undoes a deletion.
-            "SELECT DISTINCT k.keyword FROM keywords k
+            //
+            // The distinct words first, then the check: the vocabulary has no
+            // index a tombstone-inclusive lookup can use, so checking once per
+            // assignment scanned it 10,000 times on every open. Once per word
+            // is a few dozen scans of a few dozen rows.
+            "SELECT k.keyword FROM (SELECT DISTINCT keyword FROM keywords) k
               WHERE NOT EXISTS (SELECT 1 FROM keyword_terms t
                                  WHERE t.name = k.keyword)",
         )?;

core/dr-catalog/src/lib.rs

@@ -40,6 +40,7 @@ pub mod bursts;
 pub mod cache;
 pub mod collections;
 pub mod dedup;
+pub mod duplicates;
 pub mod error;
 pub mod face_shard;
 pub mod faces;

core/dr-catalog/src/merge.rs

@@ -931,16 +931,36 @@ fn merge_people_within(tx: &Connection, report: &mut MergeReport) -> Result<(),
             .query_map([], |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?, r.get(3)?)))?
             .collect::<Result<_, _>>()?;
 
+        // Cached, not prepared per row: a steady-state pass walks every
+        // confirmed and every ignored face the other device holds -- 13,000 on
+        // the reference library -- and preparing four statements for each was
+        // most of the 780 ms a merge that changed nothing cost.
+        let mut person_of = tx.prepare_cached("SELECT id FROM people WHERE uuid = ?1")?;
+        let mut current = tx.prepare_cached(
+            "SELECT person_id, probability, confirmed FROM face_person WHERE face_id = ?1",
+        )?;
+        let mut rejected_q = tx.prepare_cached(
+            "SELECT EXISTS(SELECT 1 FROM face_person_rejected
+                            WHERE face_id = ?1 AND person_id = ?2)",
+        )?;
+        let mut assign = tx.prepare_cached(
+            "INSERT INTO face_person (face_id, person_id, probability, confirmed)
+             VALUES (?1, ?2, ?3, ?4)
+             ON CONFLICT(face_id) DO UPDATE SET
+                 person_id = excluded.person_id,
+                 probability = excluded.probability,
+                 confirmed = excluded.confirmed",
+        )?;
+
         for (remote_face, uuid, probability, confirmed) in incoming {
             let Some(&local_face) = face_map.get(&remote_face) else {
                 continue;
             };
-            let person: Option<i64> = tx
-                .query_row("SELECT id FROM people WHERE uuid = ?1", [&uuid], |r| {
-                    r.get(0)
-                })
-                .optional()?;
+            let person: Option<i64> = person_of.query_row([&uuid], |r| r.get(0)).optional()?;
             let Some(person) = person else { continue };
+            let held: Option<(i64, f64, i64)> = current
+                .query_row([local_face], |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)))
+                .optional()?;
 
             // A local confirmation is never overwritten, in either direction.
             // Two devices confirming the same face as different people is a
@@ -948,13 +968,7 @@ fn merge_people_within(tx: &Connection, report: &mut MergeReport) -> Result<(),
             // settle it with; silently taking the remote's answer would let a
             // sync undo something the user did here. It stays as it is, and the
             // user can change it on the device they are looking at.
-            let locally_confirmed: bool = tx.query_row(
-                "SELECT EXISTS(SELECT 1 FROM face_person
-                                WHERE face_id = ?1 AND confirmed = 1)",
-                [local_face],
-                |r| r.get(0),
-            )?;
-            if locally_confirmed {
+            if held.is_some_and(|(_, _, confirmed)| confirmed == 1) {
                 report.faces_kept_local += 1;
                 continue;
             }
@@ -963,25 +977,22 @@ fn merge_people_within(tx: &Connection, report: &mut MergeReport) -> Result<(),
             // this user's judgement about this pair, and re-suggesting what
             // they pushed away is the behaviour that makes the feature feel
             // broken.
-            let rejected: bool = tx.query_row(
-                "SELECT EXISTS(SELECT 1 FROM face_person_rejected
-                                WHERE face_id = ?1 AND person_id = ?2)",
-                [local_face, person],
-                |r| r.get(0),
-            )?;
+            let rejected: bool = rejected_q.query_row([local_face, person], |r| r.get(0))?;
             if rejected {
                 continue;
             }
 
-            tx.execute(
-                "INSERT INTO face_person (face_id, person_id, probability, confirmed)
-                 VALUES (?1, ?2, ?3, ?4)
-                 ON CONFLICT(face_id) DO UPDATE SET
-                     person_id = excluded.person_id,
-                     probability = excluded.probability,
-                     confirmed = excluded.confirmed",
-                rusqlite::params![local_face, person, probability, confirmed],
-            )?;
+            // Written only when it differs. Rewriting a row with the values it
+            // already holds dirtied a page per face, every pass, for nothing;
+            // the report still counts it, as it always has.
+            if held != Some((person, probability, i64::from(confirmed))) {
+                assign.execute(rusqlite::params![
+                    local_face,
+                    person,
+                    probability,
+                    confirmed
+                ])?;
+            }
             report.faces_assigned += 1;
         }
     }
@@ -997,32 +1008,30 @@ fn merge_people_within(tx: &Connection, report: &mut MergeReport) -> Result<(),
             .query_map([], |r| Ok((r.get(0)?, r.get(1)?)))?
             .collect::<Result<_, _>>()?;
 
+        let mut person_of = tx.prepare_cached("SELECT id FROM people WHERE uuid = ?1")?;
+        let mut reject = tx.prepare_cached(
+            "INSERT OR IGNORE INTO face_person_rejected (face_id, person_id)
+             VALUES (?1, ?2)",
+        )?;
+        let mut unsuggest = tx.prepare_cached(
+            "DELETE FROM face_person
+              WHERE face_id = ?1 AND person_id = ?2 AND confirmed = 0",
+        )?;
+
         for (remote_face, uuid) in incoming {
             let Some(&local_face) = face_map.get(&remote_face) else {
                 continue;
             };
-            let person: Option<i64> = tx
-                .query_row("SELECT id FROM people WHERE uuid = ?1", [&uuid], |r| {
-                    r.get(0)
-                })
-                .optional()?;
+            let person: Option<i64> = person_of.query_row([&uuid], |r| r.get(0)).optional()?;
             let Some(person) = person else { continue };
 
-            let n = tx.execute(
-                "INSERT OR IGNORE INTO face_person_rejected (face_id, person_id)
-                 VALUES (?1, ?2)",
-                [local_face, person],
-            )?;
+            let n = reject.execute([local_face, person])?;
             report.faces_rejected += n;
 
             // A rejection that lands on a face currently *suggested* to be
             // that person has to take the suggestion with it, or the screen
             // keeps offering exactly what the other device just refused.
-            tx.execute(
-                "DELETE FROM face_person
-                  WHERE face_id = ?1 AND person_id = ?2 AND confirmed = 0",
-                [local_face, person],
-            )?;
+            unsuggest.execute([local_face, person])?;
         }
     }
 

core/dr-catalog/src/rating.rs

@@ -49,6 +49,12 @@ pub struct Judgement {
     /// 0..=5. Zero means *unrated*, which is a state in its own right.
     pub rating: u8,
     pub flag: FlagState,
+    /// TRACES: FR-CAT-5
+    /// The colour label, or `None`. Not part of [`Judgement::is_judged`]:
+    /// a label sorts photographs into piles of the photographer's own
+    /// meaning — "to print", "send to Anna" — and says nothing about whether
+    /// a frame has been culled, which is the question "unjudged" asks.
+    pub label: Option<ColourLabel>,
 }
 
 impl Judgement {
@@ -267,14 +273,6 @@ pub fn align_default_version_uuids(conn: &Connection) -> Result<usize, CatalogEr
     Ok(moved)
 }
 
-/// The default version's row id for an image, creating one if it has none.
-///
-/// Every write path goes through this rather than assuming a version exists.
-/// An image can arrive without one in two ways that are not worth trying to
-/// prevent: a row inserted by a build predating this module, and a scan whose
-/// version pass was interrupted between the image insert and the commit.
-/// Failing a rating because of either would be the wrong answer — the user
-/// pressed a key and expects a star.
 /// TRACES: FR-CAT-13
 /// How `versions.label` encodes a colour label, and back.
 ///
@@ -304,6 +302,14 @@ pub fn label_from_code(code: Option<i64>) -> Option<ColourLabel> {
     })
 }
 
+/// The default version's row id for an image, creating one if it has none.
+///
+/// Every write path goes through this rather than assuming a version exists.
+/// An image can arrive without one in two ways that are not worth trying to
+/// prevent: a row inserted by a build predating this module, and a scan whose
+/// version pass was interrupted between the image insert and the commit.
+/// Failing a rating because of either would be the wrong answer — the user
+/// pressed a key and expects a star.
 pub fn default_version_id(conn: &Connection, image: ImageId) -> Result<i64, CatalogError> {
     let existing: Option<i64> = conn
         .query_row(
@@ -387,6 +393,88 @@ pub fn set_flag_many(
     apply_many(conn, images, |conn, id| set_flag(conn, id, flag))
 }
 
+/// TRACES: FR-CAT-5
+/// Set or clear the colour label for one image.
+pub fn set_label(
+    conn: &Connection,
+    image: ImageId,
+    label: Option<ColourLabel>,
+) -> Result<(), CatalogError> {
+    let version = default_version_id(conn, image)?;
+    conn.execute(
+        "UPDATE versions SET label = ?2 WHERE id = ?1",
+        rusqlite::params![version, label.map(label_code)],
+    )?;
+    Ok(())
+}
+
+/// TRACES: FR-CAT-5
+/// Set or clear a label on many images in one transaction — one keystroke
+/// over a selection is one commit, as for [`set_rating_many`].
+pub fn set_label_many(
+    conn: &Connection,
+    images: &[ImageId],
+    label: Option<ColourLabel>,
+) -> Result<usize, CatalogError> {
+    apply_many(conn, images, |conn, id| set_label(conn, id, label))
+}
+
+/// TRACES: FR-CAT-5
+/// What a label key does to a set of images: Lightroom's toggle.
+///
+/// Pressing the key for the label every one of them already carries takes it
+/// off; otherwise every one of them gets it. Decided over the whole set
+/// rather than per image, so a selection that was half red comes out all red
+/// rather than inverted — the photographer pressed "red", and a key that
+/// turned half of them red and the other half plain would be two answers to
+/// one question.
+pub fn toggled_label(
+    current: impl IntoIterator<Item = Option<ColourLabel>>,
+    pressed: ColourLabel,
+) -> Option<ColourLabel> {
+    let mut any = false;
+    for label in current {
+        any = true;
+        if label != Some(pressed) {
+            return Some(pressed);
+        }
+    }
+    if any {
+        None
+    } else {
+        Some(pressed)
+    }
+}
+
+/// TRACES: FR-CAT-5 | FR-CAT-6
+/// How the library divides by colour label, for the filter chips' counts.
+///
+/// Index 0 is unlabelled and index `n` the label whose code is `n`. One
+/// grouped statement — the same shape as [`rating_histogram`], and for the
+/// same reason it LEFT JOINs: an image without a version row is unlabelled,
+/// not missing.
+pub fn label_histogram(conn: &Connection) -> Result<[usize; 6], CatalogError> {
+    let mut out = [0usize; 6];
+    let mut stmt = conn.prepare(
+        "SELECT coalesce(v.label, 0) AS l, count(*)
+           FROM images i
+           LEFT JOIN versions v ON v.image_id = i.id AND v.is_default = 1
+          GROUP BY l",
+    )?;
+    let rows = stmt.query_map([], |r| Ok((r.get::<_, i64>(0)?, r.get::<_, i64>(1)?)))?;
+    for (code, count) in rows.flatten() {
+        // A code this build does not know counts as unlabelled, which is how
+        // `label_from_code` reads it everywhere else.
+        let slot = if label_from_code(Some(code)).is_some() {
+            code as usize
+        } else {
+            0
+        };
+        out[slot] += count as usize;
+    }
+    Ok(out)
+}
+
 /// Shared bulk wrapper, so the two axes cannot drift in their commit
 /// behaviour — a partially-committed rating and a fully-committed flag from
 /// the same keystroke would be hard to explain and harder to notice.
@@ -411,21 +499,22 @@ fn apply_many(
 /// An image with no version reads as unrated and unflagged rather than as an
 /// error: that is exactly what it is.
 pub fn judgement(conn: &Connection, image: ImageId) -> Result<Judgement, CatalogError> {
-    let row: Option<(i64, i64)> = conn
+    let row: Option<(i64, i64, Option<i64>)> = conn
         .query_row(
-            "SELECT rating, flag FROM versions
+            "SELECT rating, flag, label FROM versions
               WHERE image_id = ?1
               ORDER BY is_default DESC, id ASC
               LIMIT 1",
             [image.0 as i64],
-            |r| Ok((r.get(0)?, r.get(1)?)),
+            |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)),
         )
         .optional()?;
 
     Ok(match row {
-        Some((rating, flag)) => Judgement {
+        Some((rating, flag, label)) => Judgement {
             rating: rating.clamp(0, MAX_RATING as i64) as u8,
             flag: flag_from_code(flag),
+            label: label_from_code(label),
         },
         None => Judgement::default(),
     })
@@ -452,7 +541,7 @@ pub fn judgements(
         .collect::<Vec<_>>()
         .join(",");
     let sql = format!(
-        "SELECT image_id, rating, flag FROM versions
+        "SELECT image_id, rating, flag, label FROM versions
           WHERE image_id IN ({placeholders}) AND is_default = 1"
     );
 
@@ -467,15 +556,17 @@ pub fn judgements(
             r.get::<_, i64>(0)?,
             r.get::<_, i64>(1)?,
             r.get::<_, i64>(2)?,
+            r.get::<_, Option<i64>>(3)?,
         ))
     })?;
 
-    for (image, rating, flag) in rows.flatten() {
+    for (image, rating, flag, label) in rows.flatten() {
         out.insert(
             ImageId(image as u64),
             Judgement {
                 rating: rating.clamp(0, MAX_RATING as i64) as u8,
                 flag: flag_from_code(flag),
+                label: label_from_code(label),
             },
         );
     }
@@ -686,6 +777,72 @@ mod tests {
         assert_eq!(distinct, 200);
     }
 
+    #[test]
+    fn a_label_round_trips_and_clears() {
+        // TRACES: FR-CAT-5
+        let cat = with_images(1);
+        let id = ids(&cat)[0];
+        set_label(cat.connection(), id, Some(ColourLabel::Green)).unwrap();
+        assert_eq!(
+            judgement(cat.connection(), id).unwrap().label,
+            Some(ColourLabel::Green)
+        );
+        set_label(cat.connection(), id, None).unwrap();
+        assert_eq!(judgement(cat.connection(), id).unwrap().label, None);
+    }
+
+    #[test]
+    fn a_label_is_not_a_judgement() {
+        // "Unjudged" is the cull's resume point; a label is a pile of the
+        // photographer's own, and labelling a frame must not hide it there.
+        let cat = with_images(1);
+        let id = ids(&cat)[0];
+        set_label(cat.connection(), id, Some(ColourLabel::Red)).unwrap();
+        assert!(!judgement(cat.connection(), id).unwrap().is_judged());
+    }
+
+    #[test]
+    fn labelling_a_selection_is_one_commit_and_reaches_every_image() {
+        // TRACES: FR-CAT-5
+        let cat = with_images(4);
+        let all = ids(&cat);
+        assert_eq!(
+            set_label_many(cat.connection(), &all, Some(ColourLabel::Blue)).unwrap(),
+            4
+        );
+        let found = judgements(cat.connection(), &all).unwrap();
+        assert!(all
+            .iter()
+            .all(|id| found[id].label == Some(ColourLabel::Blue)));
+        assert_eq!(
+            label_histogram(cat.connection()).unwrap(),
+            [0, 0, 0, 0, 4, 0]
+        );
+    }
+
+    #[test]
+    fn a_label_key_toggles_only_when_every_image_already_has_it() {
+        // TRACES: FR-CAT-5
+        use ColourLabel::*;
+        assert_eq!(toggled_label([Some(Red), Some(Red)], Red), None);
+        assert_eq!(toggled_label([Some(Red), None], Red), Some(Red));
+        assert_eq!(toggled_label([Some(Blue)], Red), Some(Red));
+        assert_eq!(toggled_label([], Red), Some(Red));
+    }
+
+    #[test]
+    fn the_label_histogram_sums_to_the_library() {
+        // TRACES: FR-CAT-6
+        // Images without a version row count as unlabelled rather than
+        // vanishing, as the rating histogram's do.
+        let cat = with_images(3);
+        let first = ids(&cat)[0];
+        set_label(cat.connection(), first, Some(ColourLabel::Purple)).unwrap();
+        let h = label_histogram(cat.connection()).unwrap();
+        assert_eq!(h, [2, 0, 0, 0, 0, 1]);
+        assert_eq!(h.iter().sum::<usize>(), 3);
+    }
+
     #[test]
     fn a_rating_round_trips() {
         let cat = with_images(1);

core/dr-catalog/src/schema.rs

@@ -494,15 +494,51 @@ fn rewrite_for_attached(sql: &str, schema_name: &str) -> String {
 fn pair_raw_and_jpeg(conn: &Connection) -> Result<usize, CatalogError> {
     use std::collections::HashMap;
 
-    // (folder, lowercase stem) -> RAW id, built in one pass over the RAWs.
+    // The small side first: the JPEGs not yet paired. On a settled library
+    // these are the ones with no RAW beside them -- 1,900 of 24,000 on the
+    // reference library -- and this runs on every open, including the ones
+    // the develop view makes for each photograph it fetches. Reading every
+    // RAW to find the handful that share a folder with one of them was most
+    // of what opening the catalog cost.
+    let jpegs: Vec<(i64, Option<i64>, String)> = {
+        let mut stmt = conn.prepare(
+            "SELECT id, folder_id, source_ref FROM images
+              WHERE lower(format) IN ('jpg','jpeg') AND shadowed_by IS NULL",
+        )?;
+        let rows = stmt.query_map([], |r| {
+            Ok((
+                r.get::<_, i64>(0)?,
+                r.get::<_, Option<i64>>(1)?,
+                r.get::<_, String>(2)?,
+            ))
+        })?;
+        rows.filter_map(Result::ok).collect()
+    };
+    if jpegs.is_empty() {
+        return Ok(0);
+    }
+
+    // (folder, lowercase stem) -> RAW id, over the folders those JPEGs are in
+    // and no others: a pair is same-folder by definition. Ordered by id so
+    // that where two RAWs share a stem the later one wins, as it did when this
+    // read every RAW in table order.
+    let mut folders: Vec<i64> = jpegs.iter().filter_map(|(_, f, _)| *f).collect();
+    folders.sort_unstable();
+    folders.dedup();
+    let unfiled = jpegs.iter().any(|(_, f, _)| f.is_none());
+    let folders = serde_json::to_string(&folders).unwrap_or_else(|_| "[]".to_string());
+
     let mut raws: HashMap<(Option<i64>, String), i64> = HashMap::new();
     {
         let mut stmt = conn.prepare(
             "SELECT id, folder_id, source_ref FROM images
               WHERE lower(format) IN
-                    ('cr2','cr3','nef','arw','raf','rw2','orf','dng')",
+                    ('cr2','cr3','nef','arw','raf','rw2','orf','dng')
+                AND (folder_id IN (SELECT value FROM json_each(?1))
+                     OR (?2 AND folder_id IS NULL))
+              ORDER BY id",
         )?;
-        let rows = stmt.query_map([], |r| {
+        let rows = stmt.query_map(rusqlite::params![folders, unfiled], |r| {
             Ok((
                 r.get::<_, i64>(0)?,
                 r.get::<_, Option<i64>>(1)?,
@@ -518,25 +554,16 @@ fn pair_raw_and_jpeg(conn: &Connection) -> Result<usize, CatalogError> {
         return Ok(0);
     }
 
-    let pairs: Vec<(i64, i64)> = {
-        let mut stmt = conn.prepare(
-            "SELECT id, folder_id, source_ref FROM images
-              WHERE lower(format) IN ('jpg','jpeg') AND shadowed_by IS NULL",
-        )?;
-        let rows = stmt.query_map([], |r| {
-            Ok((
-                r.get::<_, i64>(0)?,
-                r.get::<_, Option<i64>>(1)?,
-                r.get::<_, String>(2)?,
-            ))
-        })?;
-        rows.filter_map(|row| {
-            let (id, folder, path) = row.ok()?;
-            let raw = raws.get(&(folder, stem_of(&path).to_ascii_lowercase()))?;
-            Some((id, *raw))
+    let pairs: Vec<(i64, i64)> = jpegs
+        .iter()
+        .filter_map(|(id, folder, path)| {
+            let raw = raws.get(&(*folder, stem_of(path).to_ascii_lowercase()))?;
+            Some((*id, *raw))
         })
-        .collect()
-    };
+        .collect();
+    if pairs.is_empty() {
+        return Ok(0);
+    }
 
     let tx = conn.unchecked_transaction()?;
     for (jpeg, raw) in &pairs {
@@ -1582,6 +1609,36 @@ mod tests {
         assert_eq!(backfilled(&c, "shadowed_by"), 0);
     }
 
+    #[test]
+    fn a_pair_is_found_among_other_folders_and_unfiled_images() {
+        // The RAWs are read only from the folders an unpaired JPEG is in,
+        // plus the unfiled ones when an unfiled JPEG is waiting: each JPEG
+        // must still find its own sibling, and only its own.
+        let c = with_root();
+        let raw_a = image(&c, Some(1), "a/IMG_7.CR2", "cr2");
+        image(&c, Some(2), "b/IMG_7.CR2", "cr2");
+        image(&c, Some(2), "b/IMG_8.CR2", "cr2");
+        let raw_unfiled = image(&c, None, "IMG_9.DNG", "dng");
+        let jpeg_a = image(&c, Some(1), "a/IMG_7.JPG", "jpg");
+        let jpeg_unfiled = image(&c, None, "IMG_9.jpg", "jpg");
+        image(&c, Some(1), "a/IMG_9.jpg", "jpg");
+
+        assert_eq!(backfilled(&c, "shadowed_by"), 2);
+        let of = |id: i64| -> Option<i64> {
+            c.query_row("SELECT shadowed_by FROM images WHERE id = ?1", [id], |r| {
+                r.get(0)
+            })
+            .unwrap()
+        };
+        assert_eq!(of(jpeg_a), Some(raw_a));
+        assert_eq!(of(jpeg_unfiled), Some(raw_unfiled));
+        assert_eq!(
+            backfilled(&c, "shadowed_by"),
+            0,
+            "settled on the second pass"
+        );
+    }
+
     #[test]
     fn a_raw_is_never_shadowed_by_a_jpeg() {
         // The relationship is one-way: the RAW is the photograph.

core/dr-catalog/src/trash.rs

@@ -129,28 +129,40 @@ pub fn record_trashed(
         return Ok(0);
     }
     let tx = conn.unchecked_transaction()?;
-    let mut n = 0;
-
-    {
-        let mut stmt = tx.prepare(
-            "UPDATE images
-                SET trashed_from = CASE
-                        WHEN trashed_at IS NULL THEN source_ref
-                        ELSE trashed_from
-                    END,
-                    source_ref   = ?2,
-                    trashed_at   = coalesce(trashed_at, ?3)
-              WHERE id = ?1",
-        )?;
-        for (image, path) in moved {
-            n += stmt.execute(rusqlite::params![image.0 as i64, path, now])?;
-        }
-    }
-
+    let n = record_trashed_within(&tx, moved, now)?;
     tx.commit()?;
     Ok(n)
 }
 
+/// [`record_trashed`] inside a transaction the caller owns.
+///
+/// For a caller whose trash is one half of a larger write that must land
+/// whole or not at all — consolidating duplicates (`crate::duplicates`)
+/// merges a copy's judgements onto the survivor and trashes the copy in one
+/// commit. `unchecked_transaction` cannot nest, so this is offered here
+/// rather than wrapped from above.
+pub fn record_trashed_within(
+    tx: &Connection,
+    moved: &[(ImageId, String)],
+    now: i64,
+) -> Result<usize, CatalogError> {
+    let mut n = 0;
+    let mut stmt = tx.prepare(
+        "UPDATE images
+            SET trashed_from = CASE
+                    WHEN trashed_at IS NULL THEN source_ref
+                    ELSE trashed_from
+                END,
+                source_ref   = ?2,
+                trashed_at   = coalesce(trashed_at, ?3)
+          WHERE id = ?1",
+    )?;
+    for (image, path) in moved {
+        n += stmt.execute(rusqlite::params![image.0 as i64, path, now])?;
+    }
+    Ok(n)
+}
+
 /// Record that images have been moved back out of the trash.
 ///
 /// Call after the move succeeds, for the same reason as [`record_trashed`].

core/dr-decode/src/decoder.rs

@@ -0,0 +1,121 @@
+//! TRACES: FR-RAW-2
+//! The seam a second decoder plugs into.
+//!
+//! D2 keeps LibRaw as the fallback for bodies rawler does not cover. Adding
+//! it later should be a new `impl Decoder`, not an edit to every caller that
+//! reads a header, cuts a thumbnail or opens a photograph for export — which
+//! is what the free functions alone would have made it. So the callers take a
+//! `&dyn Decoder`, and only the places that start a job name [`default`].
+//!
+//! Bytes in, always. Nothing here takes a path or a `SourceRef`: resolving a
+//! file to bytes is `Storage`'s job at the caller, so the same decoder serves a
+//! local file, an Android document and a range fetched from Nextcloud. The
+//! decoder's part in that is to say how much of a file it needs
+//! ([`Decoder::header_bytes`]) and where its preview sits
+//! ([`Decoder::locate_preview`]); the storage layer fetches exactly that.
+//!
+//! What stays a free function is what is not a decoder's to vary: recognising
+//! a JPEG ([`crate::probe`]), decoding one ([`crate::decode_jpeg`]) and
+//! checking one is whole ([`crate::is_complete_jpeg`]). A second RAW decoder
+//! would not read a JPEG differently.
+
+use dr_types::Orientation;
+
+use crate::{DecodeError, Metadata, Preview, PreviewLocation, PreviewSize, RawImage};
+
+/// TRACES: FR-RAW-2
+/// A RAW decoder, over bytes.
+///
+/// Object-safe so a caller can hold `&dyn Decoder` without becoming generic,
+/// `Send + Sync` because the callers that need one most — the thumbnail
+/// lanes, the export worker — run off the UI thread, and `Debug` so a job
+/// description that carries one can still be printed.
+pub trait Decoder: Send + Sync + std::fmt::Debug {
+    /// How much of the start of a file [`Self::metadata`] and
+    /// [`Self::locate_preview`] need. A caller reading over a network fetches
+    /// this range and no more.
+    fn header_bytes(&self) -> u64;
+
+    /// Capture metadata, from a header or a whole file, without touching
+    /// sensor data.
+    fn metadata(&self, bytes: &[u8]) -> Result<Metadata, DecodeError>;
+
+    /// How the stored pixels are turned, from a header. `None` where the file
+    /// does not say, which callers take as upright.
+    fn orientation(&self, header: &[u8]) -> Option<Orientation>;
+
+    /// Where the embedded preview best suited to a thumbnail sits in the file,
+    /// from its header, so a remote caller can fetch that range alone.
+    fn locate_preview(&self, header: &[u8], file_len: u64) -> Option<PreviewLocation>;
+
+    /// The embedded preview at the size asked for, falling through the ladder
+    /// to the next size where the file lacks it.
+    fn preview(&self, bytes: &[u8], size: PreviewSize) -> Result<Preview, DecodeError>;
+
+    /// Sensor data, for develop and export. The expensive path.
+    fn decode(&self, bytes: &[u8]) -> Result<RawImage, DecodeError>;
+}
+
+/// TRACES: FR-RAW-2
+/// The decoder the application ships: rawler for sensor data and the
+/// previews it knows, DarkRoom's own container walk for headers and ranges.
+///
+/// Its methods are the crate's free functions, unchanged. They stay public
+/// for the tools and examples that read one file and have no caller to keep
+/// decoder-agnostic.
+#[derive(Debug, Clone, Copy, Default)]
+pub struct Rawler;
+
+impl Decoder for Rawler {
+    fn header_bytes(&self) -> u64 {
+        crate::HEADER_BYTES
+    }
+
+    fn metadata(&self, bytes: &[u8]) -> Result<Metadata, DecodeError> {
+        crate::metadata(bytes)
+    }
+
+    fn orientation(&self, header: &[u8]) -> Option<Orientation> {
+        crate::orientation(header)
+    }
+
+    fn locate_preview(&self, header: &[u8], file_len: u64) -> Option<PreviewLocation> {
+        crate::locate_preview(header, file_len)
+    }
+
+    fn preview(&self, bytes: &[u8], size: PreviewSize) -> Result<Preview, DecodeError> {
+        crate::extract_preview(bytes, size)
+    }
+
+    fn decode(&self, bytes: &[u8]) -> Result<RawImage, DecodeError> {
+        crate::decode(bytes)
+    }
+}
+
+/// TRACES: FR-RAW-2
+/// The decoder a job uses unless it was handed another.
+///
+/// Named by the places that start work — a thread, a UI handler — and by
+/// nothing below them. Returning `&'static dyn Decoder` rather than `Rawler`
+/// is the point: a caller that only has this cannot reach past the trait.
+pub fn default() -> &'static dyn Decoder {
+    static RAWLER: Rawler = Rawler;
+    &RAWLER
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The default is the shipped decoder, reached through the trait: same
+    /// header budget, and the same answer to bytes neither can read.
+    #[test]
+    fn the_default_is_rawler_behind_the_trait() {
+        let d = default();
+        assert_eq!(d.header_bytes(), crate::HEADER_BYTES);
+        let junk = [0u8; 64];
+        assert_eq!(d.metadata(&junk).is_err(), crate::metadata(&junk).is_err());
+        assert!(d.decode(&junk).is_err());
+        assert_eq!(d.orientation(&junk), crate::orientation(&junk));
+    }
+}

core/dr-decode/src/lib.rs

@@ -11,14 +11,20 @@
 //!
 //! Fusing them would force a full decode where a header read suffices, which
 //! is exactly why Lightroom stalls ~2 s per image during culling.
+//!
+//! Callers reach these through the [`Decoder`] trait rather than by name, so a
+//! second decoder can be put behind them without changing any of them
+//! (FR-RAW-2). [`Rawler`] is the one that ships; [`default`] hands it out.
 
 pub mod base_curve;
+mod decoder;
 mod error;
 mod locate;
 mod preview;
 pub mod profile;
 
 pub use base_curve::BaseCurve;
+pub use decoder::{default, Decoder, Rawler};
 pub use error::DecodeError;
 pub use locate::{
     defects, is_complete_jpeg, jpeg_metadata, locate_preview, tiff_metadata, BadLine, BadPixel,

core/dr-gpu/src/adjust.rs

@@ -124,6 +124,9 @@ pub struct AdjustPass {
     /// Cleared by any render that does not write it, so a stale intermediate
     /// cannot survive a change of image and be handed to a later detail chain.
     colour_key: Option<(u64, u32, u32)>,
+    /// The source texels the fused pass read on an earlier frame, kept so a
+    /// slider drag at fit reads them contiguously. See [`SampleCache`].
+    sample: SampleCache,
     /// Fused dispatches actually encoded. Exposed so a test can see the reuse
     /// above happening rather than take it on trust.
     colour_dispatches: usize,
@@ -147,6 +150,205 @@ struct Target {
 const FILM_FORMAT: wgpu::TextureFormat = wgpu::TextureFormat::Rgba32Float;
 
 /// TRACES: FR-DEV-3f
+/// What a cached sample is valid for: the composer's
+/// [`ComposedShader::sample_key`], the source image, and the render size.
+type SampleKey = (u64, u64, u32, u32);
+
+/// The largest render the sample cache is kept for, in pixels.
+///
+/// 4K and a little over, which is every develop view there is. An export
+/// renders a whole sensor once and gains nothing from a cache it will not
+/// read again; without a ceiling, two exports of the same frame in a row
+/// would park a full-resolution copy of it on the device.
+const SAMPLE_CACHE_MAX_PIXELS: u64 = 3840 * 2400;
+
+/// How the fused dispatch gets its source colour this frame.
+#[derive(Clone, Copy, PartialEq, Eq, Debug)]
+enum SampleUse {
+    /// From the source, as it always did.
+    Direct,
+    /// From the source, and stored in the cache for the frames after.
+    Write,
+    /// From the cache.
+    Read,
+}
+
+/// TRACES: NFR-P5
+/// The fused pass's gather from the source, remembered across frames.
+///
+/// At fit, each output pixel of the fused pass reads one texel of a source
+/// three or four times its width, on a stride, and the memory system fetches
+/// the neighbours it skips along with it. On a 60 MP source that gather was
+/// most of the fused pass: 10.9 ms of a 2560 x 1600 frame against 3.8 ms for
+/// the same work reading contiguously (RTX 3050, clocks held down). But which
+/// texel a pixel reads depends on the framing and nothing else, and a slider
+/// drag does not move the framing. So the shader writes what it gathered to a
+/// render-sized texture on one frame and reads it back contiguously on every
+/// frame after, until the framing, the image or the size changes.
+///
+/// Bit-for-bit the same picture: the source is `rgba16float` and so is the
+/// cache, so the stored texel is the texel. Only the whole-texel sampling path
+/// takes part — [`ComposedShader::sample_key`] is `None` when the sample is
+/// interpolated.
+///
+/// **Written on the second frame with a key, not the first.** A drag of the
+/// crop or of a zoom changes the key on every frame, and a cache written then
+/// is never read — it would add a write per frame to exactly the gestures that
+/// can least afford one. Waiting for the key to repeat once costs a slider drag
+/// one uncached frame and costs a crop drag nothing.
+struct SampleCache {
+    /// The cache itself, `rgba16float`, sampled and written as storage.
+    target: Option<Target>,
+    /// What `target` holds, once a dispatch has written it.
+    holds: Option<SampleKey>,
+    /// The key the previous fused dispatch had, cached or not.
+    last: Option<SampleKey>,
+    /// What the most recent plan decided. Read by the tests, which have no
+    /// other way to tell a cached frame from an uncached one — the point being
+    /// that the pictures are identical.
+    last_use: SampleUse,
+    /// Bound at `@binding(6)` when the cache is not being read.
+    no_sampled: wgpu::TextureView,
+    /// Bound at `@binding(7)` when the cache is not being written.
+    no_sample_out: wgpu::TextureView,
+}
+
+impl SampleCache {
+    const FORMAT: wgpu::TextureFormat = DemosaicedImage::FORMAT;
+
+    fn new(ctx: &GpuContext) -> Self {
+        let placeholder = |label, usage| {
+            ctx.device
+                .create_texture(&wgpu::TextureDescriptor {
+                    label: Some(label),
+                    size: wgpu::Extent3d {
+                        width: 1,
+                        height: 1,
+                        depth_or_array_layers: 1,
+                    },
+                    mip_level_count: 1,
+                    sample_count: 1,
+                    dimension: wgpu::TextureDimension::D2,
+                    format: Self::FORMAT,
+                    usage,
+                    view_formats: &[],
+                })
+                .create_view(&Default::default())
+        };
+        Self {
+            target: None,
+            holds: None,
+            last: None,
+            last_use: SampleUse::Direct,
+            no_sampled: placeholder("adjust-no-sampled", wgpu::TextureUsages::TEXTURE_BINDING),
+            no_sample_out: placeholder(
+                "adjust-no-sample-out",
+                wgpu::TextureUsages::STORAGE_BINDING,
+            ),
+        }
+    }
+
+    /// Decide how this dispatch samples, on the assumption that it will be
+    /// submitted — call it after anything that can still fail.
+    fn plan(
+        &mut self,
+        ctx: &GpuContext,
+        source: &DemosaicedImage,
+        shader: &ComposedShader,
+        width: u32,
+        height: u32,
+    ) -> SampleUse {
+        self.last_use = self.decide(ctx, source, shader, width, height);
+        self.last_use
+    }
+
+    fn decide(
+        &mut self,
+        ctx: &GpuContext,
+        source: &DemosaicedImage,
+        shader: &ComposedShader,
+        width: u32,
+        height: u32,
+    ) -> SampleUse {
+        let key = shader
+            .sample_key
+            .filter(|_| u64::from(width) * u64::from(height) <= SAMPLE_CACHE_MAX_PIXELS)
+            .map(|k| (k, source.id(), width, height));
+        let last = std::mem::replace(&mut self.last, key);
+        let Some(key) = key else {
+            return SampleUse::Direct;
+        };
+        if self.holds == Some(key) {
+            return SampleUse::Read;
+        }
+        if last != Some(key) {
+            return SampleUse::Direct;
+        }
+
+        if !self
+            .target
+            .as_ref()
+            .is_some_and(|t| t.width == width && t.height == height)
+        {
+            let texture = ctx.device.create_texture(&wgpu::TextureDescriptor {
+                label: Some("adjust-sample-cache"),
+                size: wgpu::Extent3d {
+                    width,
+                    height,
+                    depth_or_array_layers: 1,
+                },
+                mip_level_count: 1,
+                sample_count: 1,
+                dimension: wgpu::TextureDimension::D2,
+                format: Self::FORMAT,
+                usage: wgpu::TextureUsages::STORAGE_BINDING | wgpu::TextureUsages::TEXTURE_BINDING,
+                view_formats: &[],
+            });
+            let view = texture.create_view(&Default::default());
+            self.target = Some(Target {
+                texture,
+                view,
+                width,
+                height,
+            });
+        }
+        // Marked as held now: the dispatch that writes it is submitted before
+        // any that could read it, and one queue orders the two.
+        self.holds = Some(key);
+        SampleUse::Write
+    }
+
+    /// Write the flags for `usage` into a fused uniform block.
+    fn flag(usage: SampleUse, uniforms: &mut [f32]) {
+        let o = dr_pipeline::SAMPLE_CACHE_UNIFORM_OFFSET;
+        uniforms[o] = if usage == SampleUse::Read { 1.0 } else { 0.0 };
+        uniforms[o + 1] = if usage == SampleUse::Write { 1.0 } else { 0.0 };
+    }
+
+    /// The views for `@binding(6)` and `@binding(7)`.
+    fn views(&self, usage: SampleUse) -> (wgpu::TextureView, wgpu::TextureView) {
+        let cache = || {
+            self.target
+                .as_ref()
+                .expect("planned with a target")
+                .view
+                .clone()
+        };
+        match usage {
+            SampleUse::Direct => (self.no_sampled.clone(), self.no_sample_out.clone()),
+            SampleUse::Write => (self.no_sampled.clone(), cache()),
+            SampleUse::Read => (cache(), self.no_sample_out.clone()),
+        }
+    }
+
+    /// Forget everything; the next render starts over.
+    fn release(&mut self) {
+        self.target = None;
+        self.holds = None;
+        self.last = None;
+    }
+}
+
 /// A baked film stock, resident on the GPU.
 struct FilmTextures {
     curves: wgpu::TextureView,
@@ -440,6 +642,7 @@ impl AdjustPass {
             camera_pipeline_layout,
             camera_target: None,
             colour_key: None,
+            sample: SampleCache::new(ctx),
             colour_dispatches: 0,
             detail_dispatches: 0,
         }
@@ -537,6 +740,30 @@ impl AdjustPass {
                         },
                         count: None,
                     },
+                    // The sample cache, read and written. Present in every
+                    // layout for the reason the masks are, and bound to
+                    // placeholders whenever the flags leave it alone. See
+                    // `SampleCache`.
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 6,
+                        visibility: wgpu::ShaderStages::COMPUTE,
+                        ty: wgpu::BindingType::Texture {
+                            sample_type: wgpu::TextureSampleType::Float { filterable: false },
+                            view_dimension: wgpu::TextureViewDimension::D2,
+                            multisampled: false,
+                        },
+                        count: None,
+                    },
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 7,
+                        visibility: wgpu::ShaderStages::COMPUTE,
+                        ty: wgpu::BindingType::StorageTexture {
+                            access: wgpu::StorageTextureAccess::WriteOnly,
+                            format: SampleCache::FORMAT,
+                            view_dimension: wgpu::TextureViewDimension::D2,
+                        },
+                        count: None,
+                    },
                 ],
             })
     }
@@ -716,7 +943,15 @@ impl AdjustPass {
         let (width, height) = (width.max(1), height.max(1));
         self.ensure_target(width, height);
 
-        let uniforms = Self::fused_uniforms(source, shader);
+        // Compile first: `pipeline` takes &mut self, and the sample cache's
+        // plan below assumes the dispatch it plans for is submitted, so nothing
+        // after it may fail.
+        let _ = self.pipeline(shader)?;
+
+        let mut uniforms = Self::fused_uniforms(source, shader);
+        let sampling = self.sample.plan(&self.ctx, source, shader, width, height);
+        SampleCache::flag(sampling, &mut uniforms);
+        let (sampled, sample_out) = self.sample.views(sampling);
 
         let params_buf = self
             .ctx
@@ -727,8 +962,6 @@ impl AdjustPass {
                 usage: wgpu::BufferUsages::UNIFORM,
             });
 
-        // Borrow order: compile first, since `pipeline` takes &mut self.
-        let _ = self.pipeline(shader)?;
         let pipeline = self
             .cache
             .get(&shader.structure_hash)
@@ -768,6 +1001,14 @@ impl AdjustPass {
                         binding: 5,
                         resource: wgpu::BindingResource::TextureView(self.film_lut_view()),
                     },
+                    wgpu::BindGroupEntry {
+                        binding: 6,
+                        resource: wgpu::BindingResource::TextureView(&sampled),
+                    },
+                    wgpu::BindGroupEntry {
+                        binding: 7,
+                        resource: wgpu::BindingResource::TextureView(&sample_out),
+                    },
                 ],
             });
 
@@ -885,7 +1126,12 @@ impl AdjustPass {
                 label: Some("adjust-detail-encoder"),
             });
 
+        let mut sampling = SampleUse::Direct;
         if !reuse {
+            let mut uniforms = uniforms;
+            sampling = self.sample.plan(&self.ctx, source, shader, width, height);
+            SampleCache::flag(sampling, &mut uniforms);
+            let (sampled, sample_out) = self.sample.views(sampling);
             let params_buf =
                 self.ctx
                     .device
@@ -927,6 +1173,14 @@ impl AdjustPass {
                             binding: 5,
                             resource: wgpu::BindingResource::TextureView(&film_lut),
                         },
+                        wgpu::BindGroupEntry {
+                            binding: 6,
+                            resource: wgpu::BindingResource::TextureView(&sampled),
+                        },
+                        wgpu::BindGroupEntry {
+                            binding: 7,
+                            resource: wgpu::BindingResource::TextureView(&sample_out),
+                        },
                     ],
                 });
             let pipeline = self
@@ -954,9 +1208,20 @@ impl AdjustPass {
             .expect("ensured above")
             .view
             .clone();
-        let ran = self
+        let ran = match self
             .detail
-            .encode(&mut enc, detail, &target_view, width, height)?;
+            .encode(&mut enc, detail, &target_view, width, height)
+        {
+            Ok(ran) => ran,
+            Err(e) => {
+                // Nothing is submitted, so a cache this frame was to write
+                // holds nothing, and must not be read as though it did.
+                if sampling == SampleUse::Write {
+                    self.sample.release();
+                }
+                return Err(e);
+            }
+        };
         self.ctx.queue.submit(Some(enc.finish()));
         self.detail_dispatches += ran;
         self.colour_key = Some((key, width, height));
@@ -1090,6 +1355,7 @@ impl AdjustPass {
         self.detail.release_caches();
         self.targets = [None, None];
         self.colour_key = None;
+        self.sample.release();
     }
 
     /// How many distinct pipelines are compiled. Exposed for tests asserting
@@ -1244,6 +1510,16 @@ impl AdjustPass {
                         binding: 5,
                         resource: wgpu::BindingResource::TextureView(self.film_lut_view()),
                     },
+                    // The sample cache is the display's; a camera-space tap
+                    // reads its source directly (its flags are zero).
+                    wgpu::BindGroupEntry {
+                        binding: 6,
+                        resource: wgpu::BindingResource::TextureView(&self.sample.no_sampled),
+                    },
+                    wgpu::BindGroupEntry {
+                        binding: 7,
+                        resource: wgpu::BindingResource::TextureView(&self.sample.no_sample_out),
+                    },
                 ],
             });
 
@@ -1898,6 +2174,62 @@ mod tests {
         );
     }
 
+    /// TRACES: FR-DEV-20
+    #[test]
+    fn a_keystone_reshapes_the_frame_without_exposing_a_corner() {
+        // The shader half of perspective correction, end to end. A top-bright
+        // frame with a full vertical keystone spreads its top across the
+        // output, so the bright half reaches further down than the middle;
+        // and since the frame is mapped onto a trapezoid *inside* the source,
+        // no corner is left without a pixel behind it.
+        let Some(ctx) = ctx() else { return };
+        let mut pass = AdjustPass::new(&ctx);
+        let img = split_image(&ctx, true);
+
+        let plain = EditGraph::default_chain().compose();
+        let tex = pass.render(&img, &plain, 32, 32).expect("render");
+        let below_middle = read_pixel(&ctx, tex, 16, 19)[0];
+        assert!(
+            below_middle < 90,
+            "unkeyed, row 19 is the dark half: {below_middle}"
+        );
+
+        let mut g = EditGraph::default_chain();
+        g.set_param(
+            dr_pipeline::framing::ID,
+            dr_pipeline::framing::KEYSTONE_V,
+            100.0,
+        );
+        g.set_param(
+            dr_pipeline::framing::ID,
+            dr_pipeline::framing::KEYSTONE_H,
+            100.0,
+        );
+        let shader = g.compose();
+        let tex = pass.render(&img, &shader, 32, 32).expect("render");
+
+        for (x, y) in [(0, 0), (31, 0), (0, 31), (31, 31)] {
+            assert_ne!(
+                read_pixel(&ctx, tex, x, y),
+                [0, 0, 0, 255],
+                "corner ({x},{y}) has no source pixel behind it"
+            );
+        }
+
+        let mut g = EditGraph::default_chain();
+        g.set_param(
+            dr_pipeline::framing::ID,
+            dr_pipeline::framing::KEYSTONE_V,
+            100.0,
+        );
+        let tex = pass.render(&img, &g.compose(), 32, 32).expect("render");
+        let keyed = read_pixel(&ctx, tex, 16, 19)[0];
+        assert!(
+            keyed > 128,
+            "the spread top half must reach row 19: {keyed}"
+        );
+    }
+
     #[test]
     fn dragging_the_crop_does_not_recompile() {
         // The cache contract for framing, which is what makes an interactive
@@ -2568,6 +2900,131 @@ mod tests {
 
     /// A flat RGBA8 image on the JPEG path — already gamma-encoded, as a
     /// decoded JPEG is.
+    /// A frame with a different value at every pixel, several times the size
+    /// of the renders below, so that a fit view reads it on a stride and a
+    /// texel read from the wrong place cannot go unnoticed.
+    fn busy_image(ctx: &GpuContext) -> DemosaicedImage {
+        let (w, h) = (97u32, 61u32);
+        let mut data = Vec::with_capacity((w * h * 4) as usize);
+        for y in 0..h {
+            for x in 0..w {
+                let n = (x.wrapping_mul(2_654_435_761) ^ y.wrapping_mul(1_640_531_527)) >> 7;
+                data.extend_from_slice(&[n as u8, (n >> 8) as u8, (x * 2 + y) as u8, 255]);
+            }
+        }
+        DemosaicedImage::from_rgba8(ctx, &data, w, h).expect("upload")
+    }
+
+    /// Render `g` with a pass that has never seen it, which reads the source
+    /// directly by construction: the reference a cached frame must equal.
+    fn fresh(ctx: &GpuContext, img: &DemosaicedImage, g: &EditGraph) -> Vec<u8> {
+        let mut pass = AdjustPass::new(ctx);
+        let detail = g.compose_detail(img.size(), (23, 15));
+        pass.render_detailed(img, &g.compose(), 23, 15, None, &detail, 1)
+            .expect("render");
+        assert_eq!(pass.sample.last_use, SampleUse::Direct);
+        pass.export_pixels().expect("read").0
+    }
+
+    #[test]
+    fn a_cached_sample_is_the_same_picture() {
+        // TRACES: NFR-P5
+        // The sample cache's whole claim: a frame that reads the source through
+        // it is bit-for-bit the frame that reads the source directly. Walked
+        // through each state — direct, writing, reading, reading after a
+        // slider moved, and direct again once the framing moves — against a
+        // fresh pass each time.
+        let Some(ctx) = ctx() else { return };
+        let img = busy_image(&ctx);
+        let mut pass = AdjustPass::new(&ctx);
+        let mut g = EditGraph::default_chain();
+
+        let frame = |pass: &mut AdjustPass, g: &EditGraph, key: u64| {
+            let detail = g.compose_detail(img.size(), (23, 15));
+            pass.render_detailed(&img, &g.compose(), 23, 15, None, &detail, key)
+                .expect("render");
+            (pass.sample.last_use, pass.export_pixels().expect("read").0)
+        };
+
+        for (i, (expected, value)) in [
+            (SampleUse::Direct, 0.3),
+            (SampleUse::Write, 0.4),
+            (SampleUse::Read, 0.5),
+            (SampleUse::Read, -0.7),
+        ]
+        .into_iter()
+        .enumerate()
+        {
+            g.set_param(exposure::ID, exposure::EXPOSURE, value);
+            let (used, pixels) = frame(&mut pass, &g, i as u64);
+            assert_eq!(used, expected, "frame {i}");
+            assert_eq!(pixels, fresh(&ctx, &img, &g), "frame {i} ({used:?})");
+        }
+
+        // A neighbourhood operation: the fused pass writes the linear
+        // intermediate instead, through the same sampling.
+        g.set_param(
+            dr_pipeline::ops::noise_reduction::ID,
+            dr_pipeline::ops::noise_reduction::CHROMA,
+            60.0,
+        );
+        for i in 10..13 {
+            g.set_param(exposure::ID, exposure::EXPOSURE, i as f32 * 0.01);
+            let (used, pixels) = frame(&mut pass, &g, i);
+            assert_eq!(used, SampleUse::Read, "detail frame {i}");
+            assert_eq!(pixels, fresh(&ctx, &img, &g), "detail frame {i}");
+        }
+
+        // The framing moves: what was cached is for the old framing.
+        g.set_param(dr_pipeline::framing::ID, dr_pipeline::framing::CROP_W, 0.6);
+        let (used, pixels) = frame(&mut pass, &g, 20);
+        assert_eq!(used, SampleUse::Direct, "a new framing reads directly");
+        assert_eq!(pixels, fresh(&ctx, &img, &g));
+        let (used, _) = frame(&mut pass, &g, 21);
+        assert_eq!(used, SampleUse::Write, "and caches once it holds still");
+        let (used, pixels) = frame(&mut pass, &g, 22);
+        assert_eq!(used, SampleUse::Read);
+        assert_eq!(pixels, fresh(&ctx, &img, &g));
+    }
+
+    #[test]
+    fn a_cache_is_not_read_for_another_image_or_size() {
+        // The key is the composer's half plus the two things only this side
+        // knows. A second photograph with the same edit and the same framing
+        // must not be shown the first one's texels.
+        let Some(ctx) = ctx() else { return };
+        let (a, b) = (busy_image(&ctx), grey_image(&ctx, 8000));
+        let mut pass = AdjustPass::new(&ctx);
+        let shader = EditGraph::default_chain().compose();
+        for _ in 0..3 {
+            pass.render(&a, &shader, 23, 15).expect("render");
+        }
+        assert_eq!(pass.sample.last_use, SampleUse::Read);
+
+        pass.render(&b, &shader, 23, 15).expect("render");
+        assert_eq!(pass.sample.last_use, SampleUse::Direct, "another image");
+        pass.render(&b, &shader, 23, 15).expect("render");
+        pass.render(&b, &shader, 24, 15).expect("render");
+        assert_eq!(pass.sample.last_use, SampleUse::Direct, "another size");
+    }
+
+    #[test]
+    fn a_straightened_frame_samples_directly() {
+        // Interpolated: the sample is a blend of four texels, which the cache's
+        // format could not hold exactly, so the composer offers no key.
+        let Some(ctx) = ctx() else { return };
+        let img = busy_image(&ctx);
+        let mut pass = AdjustPass::new(&ctx);
+        let mut g = EditGraph::default_chain();
+        g.set_param(dr_pipeline::framing::ID, dr_pipeline::framing::ANGLE, 3.0);
+        let shader = g.compose();
+        assert!(shader.sample_key.is_none());
+        for _ in 0..3 {
+            pass.render(&img, &shader, 23, 15).expect("render");
+            assert_eq!(pass.sample.last_use, SampleUse::Direct);
+        }
+    }
+
     fn jpeg_image(ctx: &GpuContext, rgb: [u8; 3]) -> DemosaicedImage {
         let size = 16u32;
         let mut data = Vec::with_capacity((size * size) as usize * 4);

core/dr-gpu/src/demosaic.rs

@@ -95,6 +95,15 @@ pub struct DemosaicedImage {
     base_curve: BaseCurve,
     /// Whether the texture holds gamma-encoded rather than linear values.
     non_linear: bool,
+    /// Which upload this is, unique for the life of the process. See
+    /// [`Self::id`].
+    id: u64,
+}
+
+/// The next [`DemosaicedImage::id`].
+fn next_image_id() -> u64 {
+    static NEXT: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(1);
+    NEXT.fetch_add(1, std::sync::atomic::Ordering::Relaxed)
 }
 
 impl DemosaicedImage {
@@ -112,6 +121,18 @@ impl DemosaicedImage {
         (self.width, self.height)
     }
 
+    /// Which texture this is, as a number that is never reused.
+    ///
+    /// For a cache that has to know it is still looking at the same pixels
+    /// (`AdjustPass`'s sample cache) without holding the texture alive to find
+    /// out: keeping a handle would keep half a gigabyte of a closed photograph
+    /// on the device, and comparing addresses would mistake a new upload for an
+    /// old one the moment the allocator reused the slot. A texture here is
+    /// never written after it is built, so the same id is the same pixels.
+    pub(crate) fn id(&self) -> u64 {
+        self.id
+    }
+
     /// Camera RGB → linear sRGB, row-major. Identity where the body is
     /// uncalibrated, so the image renders uncalibrated rather than black.
     pub fn color_matrix(&self) -> [f32; 9] {
@@ -246,6 +267,7 @@ impl DemosaicedImage {
             // the highlights of an image that was already finished.
             base_curve: BaseCurve::IDENTITY,
             non_linear: true,
+            id: next_image_id(),
         })
     }
 }
@@ -322,6 +344,7 @@ impl DemosaicedImage {
             as_shot_wb: [raw.wb_coeffs[0], raw.wb_coeffs[1], raw.wb_coeffs[2]],
             base_curve: raw.base_curve,
             non_linear: false,
+            id: next_image_id(),
         })
     }
 }
@@ -660,6 +683,7 @@ impl Demosaicer {
             // normalises against black and white levels and applies no
             // transfer function.
             non_linear: false,
+            id: next_image_id(),
         })
     }
 }

core/dr-gpu/src/focus.rs

@@ -470,7 +470,7 @@ impl FocusPeakPass {
             // TEXTURE_BINDING to be sampled by the compositor.
             // RENDER_ATTACHMENT is not used by anything here and is required
             // anyway: Slint rejects an imported texture without it. COPY_SRC
-            // is for `read_overlay` and its two callers.
+            // is for `read_overlay` and the tests that call it.
             usage: wgpu::TextureUsages::STORAGE_BINDING
                 | wgpu::TextureUsages::TEXTURE_BINDING
                 | wgpu::TextureUsages::RENDER_ATTACHMENT
@@ -490,18 +490,11 @@ impl FocusPeakPass {
     /// TRACES: AC-8
     /// Copy the overlay to the CPU, as RGBA8 rows with no padding.
     ///
-    /// **Two callers, and neither is the desktop display path.** The tests
-    /// below are one: an overlay is a claim about which pixels are sharp, and
-    /// there is no way to check that claim without looking at the pixels. The
-    /// other is the Android develop view, which reads the *frame* back for the
-    /// reasons `technical-debt.md` TD-1 records — wgpu's Android swapchain
-    /// tears a portrait window, so Slint is not drawing with wgpu there and no
-    /// texture can be handed over. An overlay that stayed on the device on a
-    /// platform where the picture underneath it does not would simply never be
-    /// seen.
-    ///
-    /// On desktop nothing calls this, and ARCH §6.1 holds on the path that
-    /// matters: the overlay reaches the compositor as a texture.
+    /// **The tests below are the only caller, and never the display path.** An
+    /// overlay is a claim about which pixels are sharp, and there is no way to
+    /// check that claim without looking at the pixels. On screen, on desktop
+    /// and Android alike, the overlay reaches the compositor as a texture and
+    /// ARCH §6.1 holds. (Android read it back here until TD-1 was paid off.)
     pub fn read_overlay(&self) -> Result<(Vec<u8>, u32, u32), GpuError> {
         let Some(layer) = self.layers[self.current].as_ref() else {
             return Err(GpuError::Readback("no overlay has been rendered".into()));

core/dr-gpu/src/lib.rs

@@ -6,7 +6,7 @@
 //! module doc said for eight releases that it held no pipeline and no masks.
 //! It holds both now, plus demosaic, detail, segmentation masks, two
 //! histograms and focus peaking. The zero-copy claim is still the one that
-//! matters, and TD-1 records the one platform where it does not hold.
+//! matters, and since TD-1 was paid off it holds on Android too.
 //!
 //! Deliberately free of UI dependencies (ARCH §6.5a). The texture is handed
 //! out as a `wgpu::Texture`; who composites it is not this crate's concern.

core/dr-gpu/src/mask.rs

@@ -395,6 +395,7 @@ pub struct MaskPass {
     combine_layout: wgpu::BindGroupLayout,
     combine_union: wgpu::RenderPipeline,
     combine_subtract: wgpu::RenderPipeline,
+    combine_intersect: wgpu::RenderPipeline,
     /// Where a part is drawn before it is joined.
     ///
     /// One texture for the whole stack rather than one per layer, because
@@ -651,6 +652,16 @@ impl MaskPass {
             "mask-combine-subtract",
             blend_state(wgpu::BlendFactor::Zero, wgpu::BlendFactor::OneMinusSrc),
         );
+        // TRACES: FR-DEV-19a
+        // `dst * src`: what the mask had, kept only in proportion to how much
+        // of it this part also covers. The same three vertices and the same
+        // scratch, so a third set operation is a third blend state and
+        // nothing more — which is what `Join::apply` states on the CPU and
+        // `the_joins_match_their_definition` holds this to.
+        let combine_intersect = combine(
+            "mask-combine-intersect",
+            blend_state(wgpu::BlendFactor::Zero, wgpu::BlendFactor::Src),
+        );
 
         // The same, with the deposit thrown away: coverage is only ever taken
         // off what earlier strokes on this layer put down. There is no negative
@@ -681,6 +692,7 @@ impl MaskPass {
             combine_layout,
             combine_union,
             combine_subtract,
+            combine_intersect,
             scratch: None,
             array: None,
             allocations: 0,
@@ -1376,6 +1388,7 @@ impl MaskPass {
         pass.set_pipeline(match join {
             Join::Union => &self.combine_union,
             Join::Subtract => &self.combine_subtract,
+            Join::Intersect => &self.combine_intersect,
         });
         pass.set_bind_group(0, &bind_group, &[]);
         pass.draw(0..3, 0..1);

core/dr-gpu/tests/local_adjustments.rs

@@ -790,6 +790,119 @@ fn the_order_parts_are_joined_in_is_the_mask() {
     );
 }
 
+/// TRACES: FR-DEV-19a
+/// The truth table `docs/dev/mask-editing.md` §13 asks for: one base, one
+/// part that half-covers it, joined each of the three ways. The base is the
+/// left half of the frame and the part a dab in the middle, so the four
+/// quarters of the table are four pixels.
+#[test]
+fn a_part_unioned_subtracted_and_intersected_gives_the_three_fields() {
+    let Some(ctx) = ctx() else {
+        eprintln!("no adapter; skipping");
+        return;
+    };
+
+    let field = split_field(&ctx);
+    let joined = |join: Join| {
+        let mut layer = brighten(MaskSource::Regions {
+            signature: 1,
+            level: 2,
+            ids: vec![0],
+        });
+        assert!(layer.push_part(MaskPart::painted("p2", join)));
+        paint(&mut layer, 1, false, &[(0.5, 0.5)]);
+        let mut stack = MaskStack::new();
+        stack.push(layer);
+        render(&ctx, &stack, Some(&field))
+    };
+
+    // (base, part): left outside the dab, left inside, right inside, right
+    // outside.
+    let cells = [(4, 16), (14, 16), (18, 16), (27, 16)];
+    let lit = |pixels: &[u8]| cells.map(|(x, y)| luma_at(pixels, x, y) > 200);
+
+    assert_eq!(
+        lit(&joined(Join::Union)),
+        [true, true, true, false],
+        "union: either"
+    );
+    assert_eq!(
+        lit(&joined(Join::Subtract)),
+        [true, false, false, false],
+        "subtract: the base without the dab"
+    );
+    assert_eq!(
+        lit(&joined(Join::Intersect)),
+        [false, true, false, false],
+        "intersect: only where both are"
+    );
+}
+
+/// TRACES: FR-DEV-19a
+/// Intersection on soft coverage is the product `Join::apply` defines, on
+/// either side of the join: a gradient intersected with a region it fills is
+/// the gradient there and nothing elsewhere, and a region intersected with a
+/// gradient is the gradient wherever the region is.
+#[test]
+fn the_joins_match_their_definition() {
+    let Some(ctx) = ctx() else {
+        eprintln!("no adapter; skipping");
+        return;
+    };
+
+    let field = split_field(&ctx);
+    let ramp = || MaskSource::Linear {
+        centre: (0.5, 0.5),
+        angle: 0.0,
+        width: 1.0,
+    };
+    let right_half = || MaskSource::Regions {
+        signature: 1,
+        level: 2,
+        ids: vec![1],
+    };
+    let draw = |layer: MaskLayer| {
+        let mut stack = MaskStack::new();
+        stack.push(layer);
+        render(&ctx, &stack, Some(&field))
+    };
+
+    let alone = draw(brighten(ramp()));
+
+    let mut ramp_then_region = brighten(ramp());
+    assert!(ramp_then_region.push_part(MaskPart::new("p2", Join::Intersect, right_half())));
+    let ramp_then_region = draw(ramp_then_region);
+
+    let mut region_then_ramp = brighten(whole_frame());
+    assert!(region_then_ramp.push_part(MaskPart::new("p2", Join::Intersect, ramp())));
+    let region_then_ramp = draw(region_then_ramp);
+
+    for x in 0..SIZE {
+        let y = SIZE / 2;
+        let want = luma_at(&alone, x, y);
+        // dst · 1 = dst on the right; dst · 0 = 0 on the left. Pixels
+        // within two of the seam are left out: the region's own edge
+        // is soft there, so neither side of the table is 0 or 1.
+        let got = luma_at(&ramp_then_region, x, y);
+        if x.abs_diff(SIZE / 2) <= 2 {
+            // The seam.
+        } else if x > SIZE / 2 {
+            assert!(
+                got.abs_diff(want) <= 1,
+                "x={x}: the ramp survives where the region is ({got} vs {want})"
+            );
+        } else {
+            assert_eq!(got, 128, "x={x}: and nothing survives where it is not");
+        }
+        // 1 · src = src everywhere.
+        let got = luma_at(&region_then_ramp, x, y);
+        assert!(
+            got.abs_diff(want) <= 1,
+            "x={x}: a full base intersected with the ramp is the ramp ({got} vs {want})"
+        );
+    }
+}
+
 // --- seeing the mask (FR-DEV-19c) ------------------------------------------
 
 /// A radial that covers the middle of the frame and nothing near the corners.

core/dr-pipeline/src/detail.rs

@@ -431,6 +431,24 @@ pub struct DetailPass {
     pub storage: Vec<[f32; 4]>,
 }
 
+impl DetailPass {
+    /// Whether this pass hands on exactly what it was given: a full-size pass
+    /// with nothing to bind and a body with no code in it, only comments.
+    ///
+    /// The composer drops such a pass where that is exact (see
+    /// [`compose_detail_with`]); the operation still emits it, because whether
+    /// dropping it is exact depends on what is around it in the chain.
+    pub fn is_identity(&self) -> bool {
+        self.output_scale <= 1
+            && self.storage.is_empty()
+            && self
+                .wgsl
+                .lines()
+                .map(|l| l.split("//").next().unwrap_or("").trim())
+                .all(str::is_empty)
+    }
+}
+
 /// TRACES: FR-DEV-3 | FR-DEV-8
 /// An operation that reads pixels other than the one it is writing.
 ///
@@ -646,6 +664,40 @@ pub fn compose_detail_with(
         };
     }
 
+    // TRACES: NFR-P5
+    // A pass whose body is empty changes nothing but where the pixels are: it
+    // reads the intermediate and writes the same values to the other one.
+    // Capture sharpening emits exactly that at a scale too coarse to draw its
+    // radius (`nothing_to_sharpen`), and at fit on any modern sensor that is
+    // most of the time — so an edit with sharpening *and* another kernel paid
+    // a full render-sized read and write for it on every frame, 4.4 ms of a
+    // 2560 x 1600 frame on the reference laptop with its clocks held down.
+    //
+    // Dropped here, where the chain is still a list, and only where dropping
+    // it is exact:
+    //
+    // - **Not the last pass.** The last pass performs the output transform on
+    //   what it read from an `rgba16float` intermediate. Moving that transform
+    //   onto the pass before would apply it to that pass's `f32` result
+    //   instead, which is a different rounding of the same picture.
+    // - **Not after a reduced pass.** A full-resolution pass ends the reduced
+    //   chain (see `DetailRunner::encode`), so one that follows a scaled pass
+    //   is what stops the next operation reading the last one's base. None of
+    //   today's operations leave a reduced chain open, but a declared one may.
+    //
+    // Everywhere else the pass before and the pass after exchange the same
+    // `rgba16float` texels either way, `aux` included.
+    let mut kept: Vec<(&str, &[Helper], DetailPass, usize)> = Vec::with_capacity(planned.len());
+    let total = planned.len();
+    for (position, entry) in planned.into_iter().enumerate() {
+        let after_full = kept.last().is_none_or(|(_, _, p, _)| p.output_scale <= 1);
+        let droppable = position + 1 < total && after_full && entry.2.is_identity();
+        if !droppable {
+            kept.push(entry);
+        }
+    }
+    let planned = kept;
+
     let last = planned.len().saturating_sub(1);
     let passes = planned
         .into_iter()
@@ -1211,6 +1263,56 @@ mod tests {
         assert_eq!(fused(&ops).output_mode, OutputMode::LinearWorking);
     }
 
+    #[test]
+    fn a_pass_that_changes_nothing_is_dropped_where_that_is_exact() {
+        // TRACES: NFR-P5
+        // Capture sharpening at a scale too coarse to draw its radius emits a
+        // pass with an empty body. Between two other passes it costs a
+        // render-sized read and write and changes no texel, so it goes; as the
+        // last pass it performs the output transform on the intermediate, and
+        // moving that onto the pass before would round differently, so it
+        // stays.
+        use crate::ops::{capture_sharpen, CaptureSharpen, NoiseReduction};
+        let sharpen = || -> Box<dyn Operation> {
+            let mut op = CaptureSharpen::new();
+            op.set_param(capture_sharpen::AMOUNT, 60.0);
+            Box::new(op)
+        };
+        let chroma = || -> Box<dyn Operation> { Box::new(NoiseReduction::with_amounts(0.0, 60.0)) };
+        // A 24 MP frame fitted to a panel: a one-source-pixel radius is a
+        // quarter of a render pixel.
+        let scale = RenderScale::new((1500, 1000), (6000, 4000));
+        let unresolved = sharpen().detail().expect("a detail stage").passes(scale);
+        assert!(
+            unresolved.len() == 1 && unresolved[0].is_identity(),
+            "the premise: sharpening at this scale is one pass that does nothing"
+        );
+        let labels = |ops: &[Box<dyn Operation>]| -> Vec<String> {
+            compose_detail(ops, scale, dr_types::ColourSpace::Srgb)
+                .passes
+                .iter()
+                .map(|p| p.label.clone())
+                .collect()
+        };
+
+        // First, ahead of the chroma passes: dropped.
+        let first = labels(&[sharpen(), chroma()]);
+        assert_eq!(
+            first,
+            [
+                "noise_reduction/chroma-horizontal",
+                "noise_reduction/chroma-vertical"
+            ]
+        );
+        // Last, after them: kept, and it is the pass that encodes.
+        let last = labels(&[chroma(), sharpen()]);
+        assert_eq!(last.len(), 3);
+        assert_eq!(last[2], "capture_sharpen/unresolved");
+        // Alone: kept, because the fused pass stopped short and something has
+        // to finish the frame.
+        assert_eq!(labels(&[sharpen()]), ["capture_sharpen/unresolved"]);
+    }
+
     #[test]
     fn a_pass_can_hand_a_scalar_to_the_next_one_alongside_the_colour() {
         // What makes an unsharp mask — sharpening, clarity, texture, dehaze —

core/dr-pipeline/src/framing.rs

@@ -27,6 +27,33 @@
 //! chain exactly the space it documents: normalised, centred, `r == 1` at the
 //! corner. Neither stage needs to know the other exists.
 //!
+//! # Perspective sits inside framing (FR-DEV-20)
+//!
+//! A keystone correction is composition too — straightening converging
+//! verticals reframes the photograph — so it is a step *of* framing rather
+//! than a stage beside it, and inherits framing's `Compose` attribute, its
+//! place in the sidecar and its exclusion from a default paste. Expanded, the
+//! chain reads:
+//!
+//! ```text
+//! output pixel → crop → straighten → perspective → orientation → warp (lens) → sample
+//! ```
+//!
+//! After the straightening, because the angle is a nudge applied to the
+//! corrected picture: the verticals are made parallel and *then* the whole is
+//! levelled. Before the stored orientation, because "vertical" means vertical
+//! in the photograph as it is shown — a portrait frame the camera stored on
+//! its side must converge along its displayed height, not along the sensor's
+//! rows. And before the lens warp, which still sees the whole frame it
+//! corrects, for the reason given above.
+//!
+//! The correction maps the output frame onto a trapezoid **inside** the
+//! source rather than pulling the source edges in. So a keystone on its own
+//! never exposes an empty corner, and the crop the user drew is still valid
+//! after it; only in combination with a straightening angle does the
+//! inscribed crop have anything to account for — see
+//! [`Framing::max_inscribed_crop`].
+//!
 //! # Why sampling changes with the angle
 //!
 //! At 90° steps and flips, output pixels land exactly on source pixels, so
@@ -58,6 +85,13 @@ pub const CROP_X: ParamId = ParamId("crop_x");
 pub const CROP_Y: ParamId = ParamId("crop_y");
 pub const CROP_W: ParamId = ParamId("crop_w");
 pub const CROP_H: ParamId = ParamId("crop_h");
+/// TRACES: FR-DEV-20
+/// Vertical keystone. Positive spreads the top of the frame — the correction
+/// for a building photographed looking up, whose verticals lean together.
+pub const KEYSTONE_V: ParamId = ParamId("keystone_v");
+/// TRACES: FR-DEV-20
+/// Horizontal keystone. Positive spreads the right-hand side of the frame.
+pub const KEYSTONE_H: ParamId = ParamId("keystone_h");
 
 /// Widest straightening the control offers, in degrees either way.
 ///
@@ -66,12 +100,28 @@ pub const CROP_H: ParamId = ParamId("crop_h");
 /// the edits actually are.
 pub const MAX_STRAIGHTEN: f32 = 45.0;
 
+/// The keystone sliders' travel either way.
+///
+/// A plain amount rather than degrees of tilt: the angle a camera was tilted
+/// by depends on a focal length the correction does not know, and a number
+/// that claimed to be one would be wrong for every lens but one.
+pub const MAX_KEYSTONE: f32 = 100.0;
+
+/// How far a full keystone narrows the far edge of the frame, as a fraction
+/// of its width: at `MAX_KEYSTONE` the source trapezoid's short side is half
+/// its long one.
+///
+/// Enough for a tall building from its own pavement, and short of the point
+/// where the stretched edge is so magnified that the correction reads as a
+/// fault of its own.
+const KEYSTONE_REACH: f64 = 0.5;
+
 /// The parameters the framing widget owns — every one of them.
 ///
 /// In the order the widget expects: the rect first, then the angle it is
 /// straightened by, then the exact reorientations.
-static FRAMING_PARAMS: [ParamId; 8] = [
-    CROP_X, CROP_Y, CROP_W, CROP_H, ANGLE, ROTATION, FLIP_H, FLIP_V,
+static FRAMING_PARAMS: [ParamId; 10] = [
+    CROP_X, CROP_Y, CROP_W, CROP_H, ANGLE, KEYSTONE_V, KEYSTONE_H, ROTATION, FLIP_H, FLIP_V,
 ];
 
 static DESCRIPTOR: LazyLock<Arc<OpDescriptor>> = LazyLock::new(|| {
@@ -117,6 +167,30 @@ static DESCRIPTOR: LazyLock<Arc<OpDescriptor>> = LazyLock::new(|| {
             ParamDescriptor::fraction("crop_y", "param.crop_y", 0.0),
             ParamDescriptor::fraction("crop_w", "param.crop_w", 1.0),
             ParamDescriptor::fraction("crop_h", "param.crop_h", 1.0),
+            // TRACES: FR-DEV-20
+            // Perspective. Last so every sidecar written before these existed
+            // reads exactly as it did: a missing parameter is its default, and
+            // the default is no correction.
+            ParamDescriptor::scalar(
+                "keystone_v",
+                "param.keystone_v",
+                -MAX_KEYSTONE,
+                MAX_KEYSTONE,
+                0.0,
+                Unit::None,
+                Scale::Linear,
+                0,
+            ),
+            ParamDescriptor::scalar(
+                "keystone_h",
+                "param.keystone_h",
+                -MAX_KEYSTONE,
+                MAX_KEYSTONE,
+                0.0,
+                Unit::None,
+                Scale::Linear,
+                0,
+            ),
         ],
     })
 });
@@ -311,6 +385,86 @@ fn finite(v: f32, fallback: f32) -> f32 {
     }
 }
 
+/// TRACES: FR-DEV-20
+/// A plane projective map, row-major, acting on `(x, y, 1)`.
+///
+/// Held in `f64` because it is built by solving for four corners and then
+/// inverted for [`Framing::output_at`]; the shader gets `f32` copies of the
+/// forward map only.
+#[derive(Debug, Clone, Copy, PartialEq)]
+struct Homography([[f64; 3]; 3]);
+
+impl Homography {
+    /// The map taking the square `[-0.5, 0.5]²` onto the quadrilateral whose
+    /// corners are `q`, listed top-left, top-right, bottom-right, bottom-left.
+    ///
+    /// Heckbert's closed form for the unit square, composed with the shift
+    /// from the centred square onto it.
+    fn square_to_quad(q: [(f64, f64); 4]) -> Self {
+        let [(x0, y0), (x1, y1), (x2, y2), (x3, y3)] = q;
+        let (dx1, dx2, dx3) = (x1 - x2, x3 - x2, x0 - x1 + x2 - x3);
+        let (dy1, dy2, dy3) = (y1 - y2, y3 - y2, y0 - y1 + y2 - y3);
+        let den = dx1 * dy2 - dx2 * dy1;
+        let (g, h) = if den.abs() < 1e-12 {
+            (0.0, 0.0)
+        } else {
+            ((dx3 * dy2 - dx2 * dy3) / den, (dx1 * dy3 - dx3 * dy1) / den)
+        };
+        // Unit square (u, v) -> quad.
+        let unit = [
+            [x1 - x0 + g * x1, x3 - x0 + h * x3, x0],
+            [y1 - y0 + g * y1, y3 - y0 + h * y3, y0],
+            [g, h, 1.0],
+        ];
+        // Centred square -> unit square is `u = x + 0.5`, so fold the shift
+        // into the constant column.
+        let mut m = unit;
+        for row in &mut m {
+            row[2] += 0.5 * (row[0] + row[1]);
+        }
+        Self(m)
+    }
+
+    /// Where `(x, y)` lands, or `None` past the line the map sends to
+    /// infinity — a point with no image, which the caller treats as outside
+    /// the source.
+    fn apply(&self, (x, y): (f64, f64)) -> Option<(f64, f64)> {
+        let m = &self.0;
+        let w = m[2][0] * x + m[2][1] * y + m[2][2];
+        if w <= 1e-9 {
+            return None;
+        }
+        Some((
+            (m[0][0] * x + m[0][1] * y + m[0][2]) / w,
+            (m[1][0] * x + m[1][1] * y + m[1][2]) / w,
+        ))
+    }
+
+    /// The inverse map, by the adjugate. Scale is irrelevant to a projective
+    /// map, so the determinant is only divided out to keep `w` positive and
+    /// near one — which is what [`Self::apply`]'s horizon test relies on.
+    fn inverse(&self) -> Self {
+        let m = &self.0;
+        let c = |r0: usize, c0: usize, r1: usize, c1: usize| {
+            m[r0][c0] * m[r1][c1] - m[r0][c1] * m[r1][c0]
+        };
+        let adj = [
+            [c(1, 1, 2, 2), -c(0, 1, 2, 2), c(0, 1, 1, 2)],
+            [-c(1, 0, 2, 2), c(0, 0, 2, 2), -c(0, 0, 1, 2)],
+            [c(1, 0, 2, 1), -c(0, 0, 2, 1), c(0, 0, 1, 1)],
+        ];
+        let det = m[0][0] * adj[0][0] + m[0][1] * adj[1][0] + m[0][2] * adj[2][0];
+        let det = if det.abs() < 1e-12 { 1.0 } else { det };
+        let mut inv = adj;
+        for row in &mut inv {
+            for v in row.iter_mut() {
+                *v /= det;
+            }
+        }
+        Self(inv)
+    }
+}
+
 /// TRACES: FR-DEV-3 | FR-DEV-3d
 /// Crop, straighten, rotation and flips for one image.
 ///
@@ -325,6 +479,10 @@ pub struct Framing {
     quarter_turns: u8,
     flip_h: bool,
     flip_v: bool,
+    /// TRACES: FR-DEV-20
+    /// Vertical and horizontal keystone, each `-MAX_KEYSTONE..=MAX_KEYSTONE`.
+    keystone_v: f32,
+    keystone_h: f32,
     /// TRACES: FR-DEV-3h
     /// How the file's pixels were stored, from its EXIF orientation.
     ///
@@ -376,6 +534,8 @@ impl Default for Framing {
             quarter_turns: 0,
             flip_h: false,
             flip_v: false,
+            keystone_v: 0.0,
+            keystone_h: 0.0,
             baseline: dr_types::Orientation::NORMAL,
             crop: CropRect::default(),
             view: CropRect::default(),
@@ -396,7 +556,7 @@ impl Framing {
     /// How framing would like to be presented.
     ///
     /// **This is what stops a frontend having to name this stage.** Rendered
-    /// generically these eight parameters are eight bad controls: four crop
+    /// generically these ten parameters are ten bad controls: four crop
     /// edges the photographer would have to type coordinates into, a "rotate"
     /// slider running 0..3, and two switches. Every one of them is a worse
     /// control than the gesture it stands for — a crop is dragged on the
@@ -407,10 +567,10 @@ impl Framing {
     /// a second frontend would have had to learn the same special case, and
     /// nothing in the capability output said why. Now the preference is
     /// declared, the demand says what the widget needs, and a frontend that
-    /// cannot meet it falls back to the eight sliders — tedious, but complete,
+    /// cannot meet it falls back to the ten sliders — tedious, but complete,
     /// which is the guarantee the whole hint mechanism rests on.
     ///
-    /// The widget owns **all eight** parameters rather than only the rect: a
+    /// The widget owns **all ten** parameters rather than only the rect: a
     /// frontend that takes this on is taking on the whole framing control
     /// surface, and leaving rotation and the flips behind would scatter them
     /// into the generated panel underneath a crop control that already exists.
@@ -476,6 +636,52 @@ impl Framing {
         (self.flip_h, self.flip_v)
     }
 
+    /// TRACES: FR-DEV-20
+    /// The vertical and horizontal keystone, as the sliders show them.
+    pub fn keystone(&self) -> (f32, f32) {
+        (self.keystone_v, self.keystone_h)
+    }
+
+    /// Whether a perspective correction is applied at all.
+    pub fn has_keystone(&self) -> bool {
+        self.keystone_v != 0.0 || self.keystone_h != 0.0
+    }
+
+    /// TRACES: FR-DEV-20
+    /// The perspective map, from the straightened output frame to the upright
+    /// source frame, both measured as the centred square `[-0.5, 0.5]²`.
+    ///
+    /// **Measured in fractions of the frame, not in the aspect-scaled space
+    /// the rest of the prologue works in**, so the map is the same for every
+    /// frame shape and the uniforms need no image size. The prologue divides
+    /// `p.x` by the frame's aspect on the way in and multiplies it back on
+    /// the way out.
+    ///
+    /// The output frame's corners go to a trapezoid inside the source: a
+    /// positive vertical keystone brings the top corners in, so the top of
+    /// the source is spread across the full width of the output and lines
+    /// that converged upward come out parallel. Nothing is ever mapped from
+    /// outside the source, which is why a keystone alone needs no crop.
+    fn keystone_map(&self) -> Option<Homography> {
+        if !self.has_keystone() {
+            return None;
+        }
+        let amount = |v: f32| f64::from(v / MAX_KEYSTONE).clamp(-1.0, 1.0) * KEYSTONE_REACH;
+        let (tv, th) = (amount(self.keystone_v), amount(self.keystone_h));
+        // How much of each edge survives: the top and bottom rows' widths,
+        // the left and right columns' heights.
+        let top = 1.0 - tv.max(0.0);
+        let bottom = 1.0 + tv.min(0.0);
+        let right = 1.0 - th.max(0.0);
+        let left = 1.0 + th.min(0.0);
+        Some(Homography::square_to_quad([
+            (-0.5 * top, -0.5 * left),
+            (0.5 * top, -0.5 * right),
+            (0.5 * bottom, 0.5 * right),
+            (-0.5 * bottom, 0.5 * left),
+        ]))
+    }
+
     /// Add quarter turns, wrapping. The rotate-left/right buttons.
     pub fn rotate_quarters(&mut self, turns: i32) {
         self.quarter_turns = (i32::from(self.quarter_turns) + turns).rem_euclid(4) as u8;
@@ -544,6 +750,7 @@ impl Framing {
     pub fn is_active(&self) -> bool {
         let (turns, flip_h, flip_v) = self.effective();
         self.angle != 0.0
+            || self.has_keystone()
             // Effective, not the user's: a file stored sideways needs the
             // prologue emitted even on an untouched image, or it renders
             // through the identity map and lies on its side.
@@ -572,6 +779,7 @@ impl Framing {
     /// edited, the file was merely read correctly.
     pub fn edits_image(&self) -> bool {
         self.angle != 0.0
+            || self.has_keystone()
             || self.quarter_turns != 0
             || self.flip_h
             || self.flip_v
@@ -591,7 +799,9 @@ impl Framing {
     /// warp being active forces interpolation regardless, which is the
     /// composer's call to make rather than this stage's.
     pub fn needs_interpolation(&self) -> bool {
-        self.angle != 0.0
+        // A keystone stretches the frame by a different amount at every row,
+        // so it lands between pixels everywhere but on its centre line.
+        self.angle != 0.0 || self.has_keystone()
     }
 
     pub fn set_param(&mut self, id: ParamId, value: f32) {
@@ -629,6 +839,8 @@ impl Framing {
                 }
                 .normalised()
             }
+            KEYSTONE_V => self.keystone_v = finite(value, 0.0).clamp(-MAX_KEYSTONE, MAX_KEYSTONE),
+            KEYSTONE_H => self.keystone_h = finite(value, 0.0).clamp(-MAX_KEYSTONE, MAX_KEYSTONE),
             _ => log::warn!("framing: unknown parameter {id}"),
         }
     }
@@ -643,6 +855,8 @@ impl Framing {
             CROP_Y => self.crop.y,
             CROP_W => self.crop.width,
             CROP_H => self.crop.height,
+            KEYSTONE_V => self.keystone_v,
+            KEYSTONE_H => self.keystone_h,
             _ => 0.0,
         }
     }
@@ -707,8 +921,22 @@ impl Framing {
     ///
     /// The standard largest-inscribed-rectangle result for a rotated
     /// rectangle of the same aspect ratio.
+    ///
+    /// TRACES: FR-DEV-20
+    /// **With a keystone the closed form no longer applies**: the area with a
+    /// source pixel behind it is the source rectangle pulled back through the
+    /// perspective map and then turned, a quadrilateral no textbook result
+    /// describes. That case is searched instead — see
+    /// [`Self::inscribed_by_search`]. A keystone alone never needs a crop, so
+    /// the search returns the whole frame for it, exactly.
     pub fn max_inscribed_crop(&self, width: u32, height: u32) -> CropRect {
-        if self.angle == 0.0 || width == 0 || height == 0 {
+        if width == 0 || height == 0 {
+            return CropRect::default();
+        }
+        if self.has_keystone() {
+            return self.inscribed_by_search(width, height);
+        }
+        if self.angle == 0.0 {
             return CropRect::default();
         }
 
@@ -750,6 +978,123 @@ impl Framing {
         .normalised()
     }
 
+    /// TRACES: FR-DEV-20
+    /// The largest centred crop with a source pixel behind every point, found
+    /// by search rather than by formula.
+    ///
+    /// The area that has a source pixel behind it is convex — the source
+    /// rectangle pulled back through a projective map whose horizon lies
+    /// outside it, then turned — and a rectangle lies inside a convex region
+    /// exactly when its four corners do. For a given width the tallest
+    /// rectangle that fits is therefore found by bisection, and the area
+    /// `width × tallest(width)` is unimodal in the width (a positive concave
+    /// function times a line), so a golden-section search finds the best
+    /// width. Every rectangle returned has been tested corner by corner, so
+    /// the answer errs inside, never outside.
+    ///
+    /// A few hundred corner tests, on a gesture's release, is nothing next to
+    /// the render that follows it.
+    fn inscribed_by_search(&self, width: u32, height: u32) -> CropRect {
+        let (w, h) = if self.swaps_axes() {
+            (f64::from(height), f64::from(width))
+        } else {
+            (f64::from(width), f64::from(height))
+        };
+        let fa = w / h;
+        let rad = f64::from(self.angle).to_radians();
+        let (sn, cs) = (rad.sin(), rad.cos());
+        let map = self.keystone_map();
+
+        // Whether the output point `(fx, fy)`, in fractions of the frame from
+        // its centre, has a source pixel behind it. The prologue's steps, in
+        // its order, stopping short of the turns: those are a permutation of
+        // the frame and cannot move a point across its edge.
+        let defined = |fx: f64, fy: f64| {
+            let p = (fx * fa, fy);
+            let q = (p.0 * cs - p.1 * sn, p.0 * sn + p.1 * cs);
+            let n = (q.0 / fa, q.1);
+            let src = match &map {
+                Some(m) => m.apply(n),
+                None => Some(n),
+            };
+            const EDGE: f64 = 0.5 + 1e-9;
+            src.is_some_and(|(x, y)| x.abs() <= EDGE && y.abs() <= EDGE)
+        };
+        let fits = |hw: f64, hh: f64| {
+            [(-1.0, -1.0), (1.0, -1.0), (1.0, 1.0), (-1.0, 1.0)]
+                .iter()
+                .all(|(sx, sy)| defined(sx * hw, sy * hh))
+        };
+
+        if fits(0.5, 0.5) {
+            return CropRect::default();
+        }
+
+        // Half the tallest height that fits at half-width `hw`.
+        let tallest = |hw: f64| {
+            if fits(hw, 0.5) {
+                return 0.5;
+            }
+            if !fits(hw, 0.0) {
+                return 0.0;
+            }
+            let (mut lo, mut hi) = (0.0, 0.5);
+            for _ in 0..40 {
+                let mid = 0.5 * (lo + hi);
+                if fits(hw, mid) {
+                    lo = mid;
+                } else {
+                    hi = mid;
+                }
+            }
+            lo
+        };
+        let area = |hw: f64| hw * tallest(hw);
+
+        let ratio = (5.0_f64.sqrt() - 1.0) * 0.5;
+        let (mut a, mut b) = (0.0, 0.5);
+        let mut c = b - ratio * (b - a);
+        let mut d = a + ratio * (b - a);
+        let (mut fc, mut fd) = (area(c), area(d));
+        for _ in 0..48 {
+            if fc < fd {
+                a = c;
+                c = d;
+                fc = fd;
+                d = a + ratio * (b - a);
+                fd = area(d);
+            } else {
+                b = d;
+                d = c;
+                fd = fc;
+                c = b - ratio * (b - a);
+                fc = area(c);
+            }
+        }
+        let hw = 0.5 * (a + b);
+        let hh = tallest(hw);
+        // Tested at `hw` as returned, so a width that the search's last step
+        // nudged past the boundary cannot come back with a height that no
+        // longer fits it.
+        let (hw, hh) = if hh > 0.0 && fits(hw, hh) {
+            (hw, hh)
+        } else {
+            (c.min(d), tallest(c.min(d)))
+        };
+
+        let fw = (2.0 * hw) as f32;
+        let fh = (2.0 * hh) as f32;
+        let fw = fw.clamp(CropRect::MIN_EXTENT, 1.0);
+        let fh = fh.clamp(CropRect::MIN_EXTENT, 1.0);
+        CropRect {
+            x: (1.0 - fw) * 0.5,
+            y: (1.0 - fh) * 0.5,
+            width: fw,
+            height: fh,
+        }
+        .normalised()
+    }
+
     /// TRACES: FR-DEV-3
     /// Where an output point comes from in the source, both in normalised
     /// `0..1` coordinates.
@@ -782,6 +1127,15 @@ impl Framing {
             p = (p.0 * c - p.1 * s, p.0 * s + p.1 * c);
         }
 
+        if let Some(m) = self.keystone_map() {
+            p = match m.apply((f64::from(p.0 / fx), f64::from(p.1))) {
+                Some((x, y)) => (x as f32 * fx, y as f32),
+                // Beyond the map's horizon: no source point at all, reported
+                // as one far outside the frame rather than as a NaN.
+                None => (1e6, 1e6),
+            };
+        }
+
         let (turns, flip_h, flip_v) = self.effective();
         p = match turns {
             1 => (p.1 * ax, -p.0 / fx),
@@ -824,6 +1178,13 @@ impl Framing {
             _ => p,
         };
 
+        if let Some(m) = self.keystone_map() {
+            p = match m.inverse().apply((f64::from(p.0 / fx), f64::from(p.1))) {
+                Some((x, y)) => (x as f32 * fx, y as f32),
+                None => (1e6, 1e6),
+            };
+        }
+
         if self.angle != 0.0 {
             let rad = -self.angle * PI / 180.0;
             let (s, c) = (rad.sin(), rad.cos());
@@ -865,6 +1226,19 @@ impl Framing {
         // identical whether or not the user is zoomed in, and costs no extra
         // uniform slot.
         let rect = self.visible_rect();
+        // The perspective map by columns, so the prologue can apply it as
+        // three multiply-adds. The identity when there is none: the slots
+        // exist either way and the prologue does not read them.
+        let m = self
+            .keystone_map()
+            .unwrap_or(Homography([
+                [1.0, 0.0, 0.0],
+                [0.0, 1.0, 0.0],
+                [0.0, 0.0, 1.0],
+            ]))
+            .0;
+        let col = |c: usize| [m[0][c] as f32, m[1][c] as f32, m[2][c] as f32, 0.0];
+        let [c0, c1, c2] = [col(0), col(1), col(2)];
         [
             rect.x,
             rect.y,
@@ -874,6 +1248,18 @@ impl Framing {
             rad.cos(),
             0.0,
             0.0,
+            c0[0],
+            c0[1],
+            c0[2],
+            c0[3],
+            c1[0],
+            c1[1],
+            c1[2],
+            c1[3],
+            c2[0],
+            c2[1],
+            c2[2],
+            c2[3],
         ]
     }
 
@@ -972,6 +1358,28 @@ impl Framing {
             );
         }
 
+        if self.has_keystone() {
+            // TRACES: FR-DEV-20
+            // After the straightening and before the turns, so the keystone
+            // acts on the photograph as it is shown. The map is measured in
+            // fractions of the frame (see `Framing::keystone_map`), hence the
+            // aspect divided out and put back. A point past the map's horizon
+            // has no source at all and is sent far outside it, where the
+            // sampler's bounds test renders it void.
+            s.push_str(
+                "
+    // Perspective: the straightened frame onto a trapezoid of the source.
+    let key_n = vec2<f32>(p.x / frame_aspect.x, p.y);
+    let key_h = u.keystone_c0.xyz * key_n.x + u.keystone_c1.xyz * key_n.y + u.keystone_c2.xyz;
+    p = select(
+        vec2<f32>(1.0e6),
+        vec2<f32>(key_h.x / key_h.z * frame_aspect.x, key_h.y / key_h.z),
+        key_h.z > 1.0e-6,
+    );
+",
+            );
+        }
+
         // The user's turns and mirrors composed with the file's stored
         // orientation. One permutation covers both, so honouring the EXIF tag
         // adds no per-pixel work over an untagged file.
@@ -1040,13 +1448,15 @@ impl Framing {
             | u64::from(flip_v) << 3
             | u64::from(turns) << 4
             | u64::from(self.is_active()) << 6
+            | u64::from(self.has_keystone()) << 7
     }
 }
 
 /// Floats the framing block occupies in the generated uniform struct.
 ///
-/// Two `vec4`s: the crop rect, and the angle's sin/cos with padding.
-pub const FRAMING_UNIFORM_FIELDS: usize = 8;
+/// Five `vec4`s: the crop rect, the angle's sin/cos with padding, and the
+/// perspective map's three columns, each padded.
+pub const FRAMING_UNIFORM_FIELDS: usize = 20;
 
 #[cfg(test)]
 mod tests {
@@ -2025,6 +2435,8 @@ mod tests {
             (CROP_Y, 0.2),
             (CROP_W, 0.5),
             (CROP_H, 0.4),
+            (KEYSTONE_V, 35.0),
+            (KEYSTONE_H, -20.0),
         ] {
             f.set_param(id, v);
             assert_eq!(f.param(id), v, "{id} did not round-trip");
@@ -2241,6 +2653,8 @@ mod tests {
             f.rotate_quarters(turns);
             f.set_param(FLIP_H, 1.0);
             f.set_param(FLIP_V, 1.0);
+            f.set_param(KEYSTONE_V, 60.0);
+            f.set_param(KEYSTONE_H, -25.0);
 
             for out in [(0.0, 0.0), (0.5, 0.5), (0.2, 0.9), (0.95, 0.05)] {
                 let src = f.source_at(out, SRC.0, SRC.1);
@@ -2316,6 +2730,27 @@ mod tests {
             "the three-turn permutation moved; `source_at` must move with it"
         );
 
+        // The perspective step: the map applied in fractions of the frame,
+        // which is what `source_at` divides the aspect out for.
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_V, 40.0);
+        let prologue = f.wgsl_prologue();
+        assert!(
+            prologue.contains("let key_n = vec2<f32>(p.x / frame_aspect.x, p.y);")
+                && prologue.contains("key_h.x / key_h.z * frame_aspect.x"),
+            "the perspective step moved; `source_at` must move with it"
+        );
+        // After the straightening and before the turns, as `source_at` has it.
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_V, 40.0);
+        f.set_param(ANGLE, 3.0);
+        f.rotate_quarters(1);
+        let prologue = f.wgsl_prologue();
+        let straighten = prologue.find("// Straighten").unwrap();
+        let keystone = prologue.find("// Perspective").unwrap();
+        let turn = prologue.find("90° clockwise").unwrap();
+        assert!(straighten < keystone && keystone < turn, "{prologue}");
+
         // And the sampler's last step, which lives in `operation.rs` and is
         // the half of the map this file does not emit.
         assert!(
@@ -2323,4 +2758,241 @@ mod tests {
             "the sampler's return to texture coordinates moved"
         );
     }
+
+    // ---- perspective (FR-DEV-20) -----------------------------------------
+
+    /// A grid over the whole output frame, edges included.
+    fn grid() -> impl Iterator<Item = (f32, f32)> {
+        (0..=10).flat_map(|j| (0..=10).map(move |i| (i as f32 / 10.0, j as f32 / 10.0)))
+    }
+
+    fn inside(p: (f32, f32)) -> bool {
+        (-1e-4..=1.0 + 1e-4).contains(&p.0) && (-1e-4..=1.0 + 1e-4).contains(&p.1)
+    }
+
+    #[test]
+    fn a_keystone_is_an_edit_and_a_resample() {
+        let mut f = Framing::new();
+        let neutral = f.structure_key();
+        f.set_param(KEYSTONE_V, 30.0);
+        assert!(f.is_active());
+        assert!(f.edits_image(), "a keystone must light the modified dot");
+        assert!(f.needs_interpolation());
+        assert_ne!(f.structure_key(), neutral);
+        assert!(f.wgsl_prologue().contains("u.keystone_c0"));
+        // Neither the output size nor the crop moves: the frame is reshaped
+        // inside itself, so what the user cropped stays cropped.
+        assert_eq!(f.output_size(6000, 4000), (6000, 4000));
+        assert!(f.crop().is_full());
+    }
+
+    #[test]
+    fn the_keystone_magnitude_does_not_reach_the_structure_key() {
+        // Dragging the slider is a uniform upload, never a shader build.
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_V, 10.0);
+        let key = f.structure_key();
+        for (v, h) in [(80.0, 0.0), (-45.0, 30.0), (1.0, -100.0)] {
+            f.set_param(KEYSTONE_V, v);
+            f.set_param(KEYSTONE_H, h);
+            assert_eq!(f.structure_key(), key, "{v}/{h} forced a recompile");
+        }
+    }
+
+    #[test]
+    fn the_keystone_is_clamped_to_its_travel() {
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_V, 1e9);
+        f.set_param(KEYSTONE_H, f32::NAN);
+        assert_eq!(f.keystone(), (MAX_KEYSTONE, 0.0));
+        assert!(f.uniforms().iter().all(|v| v.is_finite()));
+    }
+
+    #[test]
+    fn a_keystone_alone_never_reaches_outside_the_source() {
+        // The design decision the crop relies on: the output frame is mapped
+        // onto a trapezoid *inside* the source, so no corner goes empty and
+        // a crop drawn before the keystone is still a crop of the picture.
+        for (v, h) in [
+            (100.0, 0.0),
+            (-100.0, 0.0),
+            (0.0, 100.0),
+            (0.0, -100.0),
+            (100.0, 100.0),
+            (-100.0, 100.0),
+            (37.0, -64.0),
+        ] {
+            for turns in 0..4 {
+                let mut f = Framing::new();
+                f.rotate_quarters(turns);
+                f.set_param(KEYSTONE_V, v);
+                f.set_param(KEYSTONE_H, h);
+                for out in grid() {
+                    let src = f.source_at(out, SRC.0, SRC.1);
+                    assert!(inside(src), "{v}/{h}, {turns} turn(s): {out:?} -> {src:?}");
+                }
+                assert!(f.max_inscribed_crop(SRC.0, SRC.1).is_full());
+            }
+        }
+    }
+
+    #[test]
+    fn a_vertical_keystone_makes_upward_converging_lines_parallel() {
+        // What the control is for. Output columns are straight verticals;
+        // with a positive keystone each must come from a straight source line
+        // that leans in toward the centre as it rises — the shape a building
+        // has when photographed looking up.
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_V, 60.0);
+
+        for x in [0.1f32, 0.3, 0.7, 0.9] {
+            let bottom = f.source_at((x, 1.0), SRC.0, SRC.1);
+            let middle = f.source_at((x, 0.5), SRC.0, SRC.1);
+            let top = f.source_at((x, 0.0), SRC.0, SRC.1);
+
+            // Straight: the middle sits on the line through the two ends.
+            let cross = (top.0 - bottom.0) * (middle.1 - bottom.1)
+                - (top.1 - bottom.1) * (middle.0 - bottom.0);
+            assert!(cross.abs() < 1e-4, "column {x} is not a straight line");
+            // Leaning in: the top is nearer the centre than the bottom.
+            assert!(
+                (top.0 - 0.5).abs() < (bottom.0 - 0.5).abs(),
+                "column {x}: top {top:?} is not inside bottom {bottom:?}"
+            );
+        }
+
+        // The bottom row is left where it was; the top row is the one spread.
+        close(
+            f.source_at((0.0, 1.0), SRC.0, SRC.1),
+            (0.0, 1.0),
+            "bottom-left",
+        );
+        close(
+            f.source_at((0.0, 0.0), SRC.0, SRC.1),
+            (0.15, 0.0),
+            "top-left",
+        );
+    }
+
+    #[test]
+    fn a_horizontal_keystone_spreads_the_right_hand_side() {
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_H, 100.0);
+        // The right-hand column comes from half the source's height.
+        close(
+            f.source_at((1.0, 0.0), SRC.0, SRC.1),
+            (1.0, 0.25),
+            "top-right",
+        );
+        close(
+            f.source_at((1.0, 1.0), SRC.0, SRC.1),
+            (1.0, 0.75),
+            "bottom-right",
+        );
+        close(
+            f.source_at((0.0, 0.0), SRC.0, SRC.1),
+            (0.0, 0.0),
+            "top-left",
+        );
+    }
+
+    #[test]
+    fn the_keystone_acts_on_the_frame_as_shown() {
+        // A portrait frame the camera stored on its side: "vertical" is the
+        // frame's displayed height, so the same keystone must move the same
+        // *displayed* points whatever the file's stored orientation.
+        let mut upright = Framing::new();
+        upright.set_param(KEYSTONE_V, 50.0);
+        let mut sideways = Framing::new();
+        sideways.set_baseline(dr_types::Orientation::from_exif(6));
+        sideways.set_param(KEYSTONE_V, 50.0);
+
+        // Compare in the displayed frame: map the sideways result back
+        // through the orientation alone.
+        let mut turn_only = Framing::new();
+        turn_only.set_baseline(dr_types::Orientation::from_exif(6));
+        for out in grid() {
+            let a = upright.source_at(out, SRC.1, SRC.0);
+            let b = turn_only.output_at(sideways.source_at(out, SRC.0, SRC.1), SRC.0, SRC.1);
+            close(a, b, "the keystone turned with the file");
+        }
+    }
+
+    #[test]
+    fn the_inscribed_crop_accounts_for_the_keystone() {
+        // Straightening a keystoned frame: the empty area is no longer the
+        // rotated rectangle's, and the crop must avoid the area that is.
+        for (angle, v, h) in [
+            (5.0f32, 50.0f32, 0.0f32),
+            (-8.0, -70.0, 20.0),
+            (12.0, 100.0, 100.0),
+            (2.0, 0.0, -40.0),
+        ] {
+            for turns in [0, 1] {
+                let mut f = Framing::new();
+                f.rotate_quarters(turns);
+                f.set_param(ANGLE, angle);
+                f.set_param(KEYSTONE_V, v);
+                f.set_param(KEYSTONE_H, h);
+                let c = f.max_inscribed_crop(SRC.0, SRC.1);
+                assert!(
+                    c.width > 0.3 && c.height > 0.3 && !c.is_full(),
+                    "{angle}°/{v}/{h}: {c:?}"
+                );
+                assert!(
+                    ((c.x + c.width * 0.5) - 0.5).abs() < 1e-4
+                        && ((c.y + c.height * 0.5) - 0.5).abs() < 1e-4,
+                    "{c:?} is not centred"
+                );
+
+                // Every point of it, edges included, has a source pixel.
+                f.set_crop(c);
+                for out in grid() {
+                    let src = f.source_at(out, SRC.0, SRC.1);
+                    assert!(inside(src), "{angle}°/{v}/{h}: {out:?} -> {src:?}");
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn the_inscribed_crop_with_a_keystone_is_not_needlessly_small() {
+        // The search must find the best rectangle, not merely a safe one. A
+        // tenth larger in either direction has to reach outside the source.
+        let mut f = Framing::new();
+        f.set_param(ANGLE, 6.0);
+        f.set_param(KEYSTONE_V, 60.0);
+        let c = f.max_inscribed_crop(SRC.0, SRC.1);
+        for (gw, gh) in [(1.1, 1.0), (1.0, 1.1)] {
+            let mut g = f;
+            let (w, h) = (c.width * gw, c.height * gh);
+            g.set_crop(CropRect {
+                x: 0.5 - w * 0.5,
+                y: 0.5 - h * 0.5,
+                width: w,
+                height: h,
+            });
+            let spills = [(0.0, 0.0), (1.0, 0.0), (1.0, 1.0), (0.0, 1.0)]
+                .into_iter()
+                .any(|out| !inside(g.source_at(out, SRC.0, SRC.1)));
+            // Either the grown rect spills, or it could not grow at all
+            // because the crop was already at the frame's edge on that axis.
+            assert!(
+                spills
+                    || (gw > 1.0 && c.width >= 1.0 - 1e-4)
+                    || (gh > 1.0 && c.height >= 1.0 - 1e-4),
+                "{c:?} grown by {gw}x{gh} still fits"
+            );
+        }
+    }
+
+    #[test]
+    fn reset_clears_the_keystone() {
+        let mut f = Framing::new();
+        f.set_param(KEYSTONE_V, 30.0);
+        f.set_param(KEYSTONE_H, -30.0);
+        f.reset();
+        assert_eq!(f.keystone(), (0.0, 0.0));
+        assert!(!f.is_active());
+    }
 }

core/dr-pipeline/src/lib.rs

@@ -44,6 +44,7 @@ pub mod mask;
 pub mod neutral;
 pub mod operation;
 pub mod ops;
+pub mod orphan;
 pub mod preset;
 pub mod sidecar;
 pub mod spot;
@@ -66,7 +67,7 @@ pub use lens::{compose_warps, ComposedWarp, LensProfile, Tca, Warp};
 pub use operation::{
     compose, compose_with_framing, Affects, ComposedShader, Helper, Invalidation, Operation,
     OutputMode, Uniform, BASE_CURVE_POINTS, BASE_CURVE_UNIFORM_OFFSET, CLIP_ONSET,
-    RESERVED_UNIFORM_FIELDS,
+    RESERVED_UNIFORM_FIELDS, SAMPLE_CACHE_UNIFORM_OFFSET,
 };
 pub use preset::{LibraryParseError, NameError, Preset, PresetLibrary, Scope};
 pub use sidecar::{Sidecar, Version};

core/dr-pipeline/src/mask.rs

@@ -926,6 +926,21 @@ pub enum Join {
     /// erase stroke is a hole in the part it was painted into and reads as
     /// nothing at all.
     Subtract,
+    /// TRACES: FR-DEV-19a
+    /// Only where both agree. "Keep the part of this mask that is also that."
+    ///
+    /// The join that makes cheap criteria precise: a sky is a category *and*
+    /// a luminance band, skin is a subject *and* a hue. Neither alone is the
+    /// selection, and no feather on either makes it one.
+    ///
+    /// The product of the two coverages rather than their minimum, because
+    /// that is what one fixed-function blend gives on the device
+    /// (`dst · src`, `docs/dev/mask-editing.md` §5.2) and it agrees with the
+    /// minimum wherever either side is fully in or fully out. Between two soft
+    /// edges it is the softer of the two readings, which is the right way to
+    /// be wrong: an overlap of two partial selections is less certainly
+    /// selected than either.
+    Intersect,
 }
 
 impl Join {
@@ -933,6 +948,7 @@ impl Join {
         match self {
             Self::Union => "union",
             Self::Subtract => "subtract",
+            Self::Intersect => "intersect",
         }
     }
 
@@ -940,12 +956,30 @@ impl Join {
         Some(match name {
             "union" => Self::Union,
             "subtract" => Self::Subtract,
+            "intersect" => Self::Intersect,
             _ => return None,
         })
     }
 
-    /// Every variant, for a UI building a choice control.
-    pub const ALL: [Join; 2] = [Join::Union, Join::Subtract];
+    /// TRACES: FR-DEV-19a
+    /// The coverage this join leaves at one point, given what the mask had
+    /// there (`dst`) and what the part covers (`src`), both in `0..=1`.
+    ///
+    /// The definition the device's blend states implement, spelled out once
+    /// on the CPU so a test can hold the GPU to it and a reader can see the
+    /// three set operations side by side without reading `wgpu` enums.
+    pub fn apply(self, dst: f32, src: f32) -> f32 {
+        match self {
+            Self::Union => dst.max(src),
+            Self::Subtract => dst * (1.0 - src),
+            Self::Intersect => dst * src,
+        }
+    }
+
+    /// Every variant, for a UI building a choice control. The order is the
+    /// panel's: a chip cycles through it and a stored index names a place in
+    /// it, so a new join goes on the end.
+    pub const ALL: [Join; 3] = [Join::Union, Join::Subtract, Join::Intersect];
 }
 
 /// One selection inside a layer's mask.
@@ -1580,9 +1614,20 @@ impl MaskLayer {
         // A hidden part is not in the build, whichever way it joins — and a
         // hidden base hands its role to the first part that is shown, which
         // is why "adds" is asked of the shown parts rather than of index 0.
+        //
+        // Folded rather than asked with `any`, because an intersection can
+        // take away everything the parts before it added: a subject
+        // intersected with an unpainted brush covers nothing. An inverted
+        // part is taken to cover, whatever its source — an inverted empty
+        // brush is the whole frame, and saying "covers nothing" of a layer
+        // that does would hide an adjustment the photographer made.
         self.shown_parts()
             .enumerate()
-            .any(|(i, p)| (i == 0 || p.join == Join::Union) && p.covers())
+            .fold(false, |acc, (i, p)| match (i, p.join) {
+                (0, _) | (_, Join::Union) => acc || p.covers(),
+                (_, Join::Subtract) => acc,
+                (_, Join::Intersect) => acc && (p.covers() || p.invert),
+            })
     }
 
     /// TRACES: FR-DEV-19a
@@ -3037,6 +3082,69 @@ mod tests {
         );
     }
 
+    /// TRACES: FR-DEV-19a
+    /// The three joins, pointwise, on every pair of coverages a part and a
+    /// mask can meet at: fully in, fully out, and each soft edge. Intersection
+    /// is the product, which agrees with the minimum wherever either side is
+    /// decided and is the softer reading where both are not.
+    #[test]
+    fn the_joins_are_max_cut_and_product() {
+        let levels = [0.0f32, 0.25, 0.5, 0.75, 1.0];
+        for &dst in &levels {
+            for &src in &levels {
+                assert_eq!(Join::Union.apply(dst, src), dst.max(src));
+                assert_eq!(Join::Subtract.apply(dst, src), dst * (1.0 - src));
+                let meet = Join::Intersect.apply(dst, src);
+                assert_eq!(meet, dst * src);
+                assert!(meet <= dst.min(src), "never more than either side");
+                if dst == 0.0 || dst == 1.0 || src == 0.0 || src == 1.0 {
+                    assert_eq!(meet, dst.min(src), "the minimum where either is decided");
+                }
+            }
+        }
+    }
+
+    /// A stored index names a place in [`Join::ALL`], and a sidecar names a
+    /// join by word: both have to survive a third join arriving, which means
+    /// the first two keep their places and every name reads back as itself.
+    #[test]
+    fn every_join_reads_back_by_name_and_keeps_its_place() {
+        assert_eq!(Join::ALL[0], Join::Union);
+        assert_eq!(Join::ALL[1], Join::Subtract);
+        for join in Join::ALL {
+            assert_eq!(Join::from_name(join.name()), Some(join));
+        }
+        assert_eq!(Join::from_name("intersect"), Some(Join::Intersect));
+    }
+
+    /// TRACES: FR-DEV-19a
+    /// An intersection can empty a mask the parts before it filled: a range
+    /// meeting an unpainted brush selects nothing, and rasterising it would
+    /// spend a slice to draw an empty field. Painting the brush, or inverting
+    /// it, gives the intersection something to keep.
+    #[test]
+    fn an_intersection_with_nothing_covers_nothing() {
+        let mut layer = MaskLayer::new("m1", MaskSource::highlights());
+        layer.set_param("exposure", ParamId("exposure"), 1.0);
+        layer.push_part(MaskPart::painted("p2", Join::Intersect));
+        assert!(
+            !layer.is_active(),
+            "a range intersected with an unpainted brush selects nothing"
+        );
+
+        layer.part_mut(1).expect("p2").invert = true;
+        assert!(layer.is_active(), "inverted, the empty brush is everywhere");
+        layer.part_mut(1).expect("p2").invert = false;
+
+        layer.begin_stroke(1, false, 0.1, 0.5, 1.0);
+        layer.extend_stroke(1, 0.5, 0.5);
+        layer.end_stroke(1);
+        assert!(layer.is_active(), "and painted, it keeps what it covers");
+
+        layer.part_mut(1).expect("p2").hidden = true;
+        assert!(layer.is_active(), "hidden, it is out of the build");
+    }
+
     /// One stale part is a stale layer: the mask is the fold over all of them,
     /// so a part that would draw a confidently wrong shape makes the result
     /// wrong whichever way it joins.

core/dr-pipeline/src/operation.rs

@@ -456,6 +456,25 @@ pub struct ComposedShader {
     pub structure_hash: u64,
     /// What this shader writes. See [`OutputMode`].
     pub output_mode: OutputMode,
+    /// What decides which source texel each output pixel reads, when that
+    /// texel is read whole — `None` when it is interpolated.
+    ///
+    /// A fit view reads one texel in every three or four of a 60 MP source,
+    /// on a stride, and that gather is most of what the fused pass costs
+    /// there: the texel it wants shares a cache line with neighbours nobody
+    /// reads. But the gather depends on the framing and nothing else, so it
+    /// is the same on every frame of a slider drag. The shader can therefore
+    /// write what it gathered to a viewport-sized texture once and read it
+    /// back contiguously thereafter; the flags in the uniform block at
+    /// [`SAMPLE_CACHE_UNIFORM_OFFSET`] say which, and `dr-gpu` decides.
+    ///
+    /// This key is the half of that decision only the composer can make: a
+    /// hash of the generated prologue and the framing and warp uniforms, which
+    /// together are everything that maps an output pixel to a source texel.
+    /// The caller mixes in the source image and the render size. `None` for
+    /// the interpolating paths, whose sample is a blend of four texels and
+    /// not representable exactly in the source's own format.
+    pub sample_key: Option<u64>,
 }
 
 /// Fields the generated uniform struct always carries, before op uniforms.
@@ -466,7 +485,19 @@ pub struct ComposedShader {
 /// Twelve of the twenty-eight are the camera profile's base curve
 /// ([`BASE_CURVE_UNIFORM_FIELDS`]); the rest are the matrix, the as-shot
 /// balance and framing's own block.
-const BASE_UNIFORM_FIELDS: usize = 16 + BASE_CURVE_UNIFORM_FIELDS;
+const BASE_UNIFORM_FIELDS: usize = 16 + SAMPLE_CACHE_UNIFORM_FIELDS + BASE_CURVE_UNIFORM_FIELDS;
+
+/// Slots the sample cache's two flags occupy: read, write, and two spare to
+/// keep the block a whole `vec4`. See [`ComposedShader::sample_key`].
+const SAMPLE_CACHE_UNIFORM_FIELDS: usize = 4;
+
+/// Where the sample cache's flags sit in the generated uniform block: `x` says
+/// read the source colour from the cache, `y` says write it there.
+///
+/// Exported for the reason [`BASE_CURVE_UNIFORM_OFFSET`] is — `dr-gpu` writes
+/// these by index — and zero in every block the composer hands out, so a
+/// caller that never heard of the cache gets the direct read it always had.
+pub const SAMPLE_CACHE_UNIFORM_OFFSET: usize = 16;
 
 /// TRACES: FR-DEV-3e
 /// Slots the base curve occupies: five `(x, y)` points and an active flag.
@@ -484,7 +515,8 @@ const BASE_CURVE_UNIFORM_FIELDS: usize = 12;
 /// Exported for the same reason [`RESERVED_UNIFORM_FIELDS`] is: `dr-gpu`
 /// writes these by index, and an offset computed independently at both ends is
 /// an offset that will eventually disagree with itself.
-pub const BASE_CURVE_UNIFORM_OFFSET: usize = 16;
+pub const BASE_CURVE_UNIFORM_OFFSET: usize =
+    SAMPLE_CACHE_UNIFORM_OFFSET + SAMPLE_CACHE_UNIFORM_FIELDS;
 
 /// How many control points a base curve carries.
 ///
@@ -756,6 +788,10 @@ fn compose_inner(
          \x20   // gamma-encoded JPEG, 0.0 for demosaiced sensor data), which the\n\
          \x20   // prologue reads to decide whether to linearise.\n\
          \x20   as_shot_wb: vec4<f32>,\n\
+         \x20   // The sample cache (see `ComposedShader::sample_key`): `.x` reads\n\
+         \x20   // the source colour from `sampled`, `.y` writes it to\n\
+         \x20   // `sample_out`. Zero for both is the direct read.\n\
+         \x20   sample_cache: vec4<f32>,\n\
          \x20   // The camera profile's base curve (FR-DEV-3e): five points on a\n\
          \x20   // monotone spline, packed as x0..x3, y0..y3, then (x4, y4, on).\n\
          \x20   // `.z` of the last is the flag, not padding — it is 0 for a\n\
@@ -801,7 +837,11 @@ fn compose_inner(
          \x20   // angle as sin/cos — a trig call per pixel would recompute a\n\
          \x20   // value that is constant across the dispatch.\n\
          \x20   crop_rect: vec4<f32>,\n\
-         \x20   framing_angle: vec4<f32>,\n",
+         \x20   framing_angle: vec4<f32>,\n\
+         \x20   // The perspective map (FR-DEV-20) by columns, `.w` unused.\n\
+         \x20   keystone_c0: vec4<f32>,\n\
+         \x20   keystone_c1: vec4<f32>,\n\
+         \x20   keystone_c2: vec4<f32>,\n",
     );
     uniform_values.extend_from_slice(&framing.uniforms());
 
@@ -932,6 +972,19 @@ fn compose_inner(
     );
     let sampler_helper = if interpolate { BILINEAR_HELPER } else { "" };
 
+    // Everything that decides which texel an output pixel reads: the code that
+    // computes `coord`, and the uniforms that code reads. Only on the path that
+    // reads a texel whole — see `ComposedShader::sample_key`.
+    let sample_key = (!interpolate && !warp.splits_channels).then(|| {
+        framing
+            .uniforms()
+            .iter()
+            .chain(&warp.uniforms)
+            .fold(hash_source(&prologue), |h, v| {
+                mix(h, u64::from(v.to_bits()))
+            })
+    });
+
     // The tail, and it is the whole of the difference between the two output
     // modes. Everything above — the prologue, the fragments, the mask layers,
     // the camera matrix — is emitted identically either way, so an operation
@@ -1083,6 +1136,12 @@ struct Params {{
 // stock is loaded, which costs eight bytes and no branch.
 @group(0) @binding(4) var film_curves: texture_2d<f32>;
 @group(0) @binding(5) var film_lut_texture: texture_3d<f32>;
+// The sample cache: the source texel each output pixel read on an earlier
+// frame with this framing, and where this frame writes it when asked. See
+// `ComposedShader::sample_key`. Declared unconditionally, like the masks, and
+// bound to 1x1 placeholders whenever the flags say not to touch them.
+@group(0) @binding(6) var sampled: texture_2d<f32>;
+@group(0) @binding(7) var sample_out: texture_storage_2d<rgba16float, write>;
 
 {sampler_helper}{helper_src}{encode_output}
 // Display-encoded sRGB back to linear, for sources that arrive that way.
@@ -1192,6 +1251,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {{
         uniforms: uniform_values,
         structure_hash,
         output_mode,
+        sample_key,
     }
 }
 
@@ -1411,7 +1471,19 @@ pub(crate) fn sample_source(interpolate: bool, splits_channels: bool) -> &'stati
     // amount that changes with the aspect ratio. It reads as a correction that
     // is simply too weak, which is indistinguishable from a bad profile.
     let radius = length(p) / (0.5 * length(aspect));
-    var c = textureLoad(source, coord, 0).rgb;
+    // The texel itself, from the source or from the cache of it an earlier
+    // frame wrote (see `ComposedShader::sample_key`). Both branches yield the
+    // same bits: the source is `rgba16float` and so is the cache. The flags
+    // are uniforms, so the whole dispatch takes one branch.
+    var c: vec3<f32>;
+    if (u.sample_cache.x > 0.5) {
+        c = textureLoad(sampled, vec2<i32>(gid.xy), 0).rgb;
+    } else {
+        c = textureLoad(source, coord, 0).rgb;
+        if (u.sample_cache.y > 0.5) {
+            textureStore(sample_out, vec2<i32>(gid.xy), vec4<f32>(c, 1.0));
+        }
+    }
 "
     }
 }

core/dr-pipeline/src/orphan.rs

@@ -0,0 +1,645 @@
+//! TRACES: FR-DEV-17
+//! Whether a crop leaves a mask layer's work outside the frame.
+//!
+//! # Why this is a question worth asking
+//!
+//! Mask geometry is stored in normalised *source* coordinates (see
+//! [`MaskSource::Linear`] and [`Stroke::points`]), so a tighter crop never
+//! destroys a layer. It makes it invisible — which is worse, because nothing
+//! announces it. The layer is still in the panel, still in the sidecar, still
+//! costing a rasterisation, and its adjustment lands on pixels nobody will
+//! ever see. The crop that did that is exactly the kind of edit made early
+//! and quickly, and the loss is found, if at all, much later.
+//!
+//! This module answers one question on the CPU, cheaply enough to ask once
+//! per committed crop: *which layers did this change of crop take out of the
+//! picture?* The interface turns the answer into a notice. It never refuses
+//! the crop — the photographer may well mean it.
+//!
+//! # How it is measured
+//!
+//! Each layer's mask is sampled on a [`GRID`]×[`GRID`] lattice over the
+//! source, with the same geometry the mask shader uses (`mask.wgsl`), folded
+//! part by part with the layer's joins and inversions. The sample points are
+//! then mapped through the framing — crop, straighten, turns and flips — into
+//! the frame, and the layer's *share inside* is the coverage that lands in
+//! the crop over the coverage there is. A layer is hidden by a crop when that
+//! share falls below [`HIDDEN_SHARE`] and was not already below it.
+//!
+//! The lattice is coarse on purpose. The question is "is this layer mostly
+//! gone", not "which pixels are": the edge treatment — feather, morphology —
+//! moves a boundary by a few hundredths of the frame and cannot turn a layer
+//! that is mostly inside into one that is mostly outside, so it is left out.
+//!
+//! # What cannot be orphaned
+//!
+//! A range ([`MaskSource::Luminance`], [`MaskSource::Colour`]) selects by a
+//! property of the picture, so wherever the crop falls it selects whatever
+//! part of the picture remains — there is nothing for a crop to strand. A
+//! region selection needs the segmentation's label map, which this crate does
+//! not hold, and a model's selection with no raster to hand has nothing to
+//! measure. A layer that *adds* any of these is therefore never reported: a
+//! false alarm on the common path would teach the photographer to dismiss the
+//! notice unread, which is the failure it exists to prevent. Subtracting one
+//! is ignored, which can only make the layer look larger — the safe side.
+
+use std::borrow::Cow;
+
+use crate::framing::{CropRect, Framing};
+use crate::mask::{Join, MaskLayer, MaskPart, MaskSource, MaskStack, Stroke};
+
+/// Samples per axis over the source.
+///
+/// 4096 points: enough that a brush dab a twentieth of the frame across is
+/// several samples wide, and few enough that the whole stack is measured in
+/// well under a frame when a crop is let go.
+pub const GRID: usize = 64;
+
+/// The share of a layer's coverage below which it counts as cropped away.
+///
+/// "Mostly outside" rather than "entirely": a gradient reduced to a sliver
+/// along one edge, or a subject of which one elbow survives, has lost the
+/// work as surely as one that is wholly gone. Low enough that an ordinary
+/// recomposition which trims part of a subject is not reported.
+pub const HIDDEN_SHARE: f32 = 0.1;
+
+/// A model's selection, as bytes over the source at some proxy size.
+///
+/// Supplied by the caller for a part whose own [`MaskPart::coverage`] is not
+/// set — the live session holds its model output outside the graph and folds
+/// it into the parts only when saving.
+pub struct Raster<'a> {
+    pub values: Cow<'a, [u8]>,
+    pub width: usize,
+    pub height: usize,
+}
+
+/// One layer's coverage, sampled over the source.
+#[derive(Debug, Clone, PartialEq)]
+pub struct Footprint {
+    /// Row-major over the lattice, `0.0..=1.0`, one per cell centre.
+    weights: Vec<f32>,
+}
+
+impl Footprint {
+    /// The layer's coverage on the lattice, or `None` when it has none to
+    /// strand — see the module header for which layers those are.
+    ///
+    /// `model` is asked for the raster behind a subject or category part that
+    /// carries none of its own.
+    pub fn of<'r>(
+        layer: &MaskLayer,
+        source: (u32, u32),
+        model: &dyn Fn(&MaskPart) -> Option<Raster<'r>>,
+    ) -> Option<Self> {
+        let mut acc: Option<Vec<f32>> = None;
+        for (i, part) in layer.shown_parts().enumerate() {
+            let adds = i == 0 || part.join == Join::Union;
+            let Some(values) = part_values(part, source, model) else {
+                if adds {
+                    // It follows the picture, or cannot be measured: either
+                    // way it is not a shape a crop can leave behind.
+                    return None;
+                }
+                continue;
+            };
+            acc = Some(match acc {
+                None => values,
+                Some(mut a) => {
+                    for (d, s) in a.iter_mut().zip(values) {
+                        *d = part.join.apply(*d, s);
+                    }
+                    a
+                }
+            });
+        }
+        let mut weights = acc?;
+        if layer.invert {
+            weights.iter_mut().for_each(|w| *w = 1.0 - *w);
+        }
+        Some(Self { weights })
+    }
+
+    /// Of this layer's coverage, the share `inside` marks as in frame.
+    /// `None` when there is no coverage at all.
+    fn share(&self, inside: &[bool]) -> Option<f32> {
+        let (mut total, mut kept) = (0.0f32, 0.0f32);
+        for (w, &i) in self.weights.iter().zip(inside) {
+            total += w;
+            if i {
+                kept += w;
+            }
+        }
+        (total > 1e-3).then(|| kept / total)
+    }
+
+    /// Of this layer's coverage, the share `framing`'s crop keeps.
+    pub fn share_inside(&self, framing: &Framing, source: (u32, u32)) -> Option<f32> {
+        self.share(&in_frame(framing, source))
+    }
+}
+
+/// TRACES: FR-DEV-17
+/// The layers that changing the framing from `before` to `after` took out of
+/// the picture, in stack order.
+///
+/// Only those it *newly* hid: a layer already cropped away by `before` is not
+/// reported again, or every later adjustment of the crop would repeat a notice
+/// the photographer has already answered.
+///
+/// The view (zoom and pan) of either framing is ignored. It is a way of
+/// looking, not the frame.
+pub fn hidden_by_crop<'m, 'r>(
+    masks: &'m MaskStack,
+    before: &Framing,
+    after: &Framing,
+    source: (u32, u32),
+    model: &dyn Fn(&MaskPart) -> Option<Raster<'r>>,
+) -> Vec<&'m MaskLayer> {
+    if masks.is_empty() || framed_alike(before, after) {
+        return Vec::new();
+    }
+    let was = in_frame(before, source);
+    let now = in_frame(after, source);
+    masks
+        .layers()
+        .iter()
+        .filter(|layer| {
+            let Some(print) = Footprint::of(layer, source, model) else {
+                return false;
+            };
+            let (Some(was), Some(now)) = (print.share(&was), print.share(&now)) else {
+                return false;
+            };
+            was >= HIDDEN_SHARE && now < HIDDEN_SHARE
+        })
+        .collect()
+}
+
+/// Whether the two frame the same part of the source, zoom aside.
+fn framed_alike(a: &Framing, b: &Framing) -> bool {
+    let mut a = *a;
+    let mut b = *b;
+    a.set_view(CropRect::default());
+    b.set_view(CropRect::default());
+    a == b
+}
+
+/// Which lattice cells the framing's crop keeps.
+fn in_frame(framing: &Framing, source: (u32, u32)) -> Vec<bool> {
+    let mut f = *framing;
+    f.set_view(CropRect::default());
+    lattice()
+        .map(|uv| {
+            let (x, y) = f.output_at(uv, source.0, source.1);
+            (0.0..=1.0).contains(&x) && (0.0..=1.0).contains(&y)
+        })
+        .collect()
+}
+
+/// Cell centres, row-major, in normalised source coordinates.
+fn lattice() -> impl Iterator<Item = (f32, f32)> {
+    let step = 1.0 / GRID as f32;
+    (0..GRID).flat_map(move |y| {
+        (0..GRID).map(move |x| ((x as f32 + 0.5) * step, (y as f32 + 0.5) * step))
+    })
+}
+
+/// One part's coverage on the lattice, its own inversion applied, or `None`
+/// where it cannot be measured as a shape.
+fn part_values<'r>(
+    part: &MaskPart,
+    source: (u32, u32),
+    model: &dyn Fn(&MaskPart) -> Option<Raster<'r>>,
+) -> Option<Vec<f32>> {
+    let aspect = source.0.max(1) as f32 / source.1.max(1) as f32;
+    let mut values: Vec<f32> = match &part.source {
+        MaskSource::Linear {
+            centre,
+            angle,
+            width,
+        } => {
+            let axis = (angle.cos(), angle.sin());
+            lattice()
+                .map(|(u, v)| {
+                    let d = (u - centre.0) * aspect * axis.0 + (v - centre.1) * axis.1;
+                    if *width <= 0.0 {
+                        if d >= 0.0 {
+                            1.0
+                        } else {
+                            0.0
+                        }
+                    } else {
+                        smoothstep(-width * 0.5, width * 0.5, d)
+                    }
+                })
+                .collect()
+        }
+        MaskSource::Radial {
+            centre,
+            radii,
+            angle,
+            feather,
+        } => {
+            let (sa, ca) = (-angle).sin_cos();
+            let radii = (radii.0.max(1e-6), radii.1.max(1e-6));
+            let edge = feather.clamp(0.0, 1.0);
+            lattice()
+                .map(|(u, v)| {
+                    let d = ((u - centre.0) * aspect, v - centre.1);
+                    let local = (d.0 * ca - d.1 * sa, d.0 * sa + d.1 * ca);
+                    let r = (local.0 / radii.0).hypot(local.1 / radii.1);
+                    if edge <= 0.0 {
+                        if r <= 1.0 {
+                            1.0
+                        } else {
+                            0.0
+                        }
+                    } else {
+                        1.0 - smoothstep(1.0 - edge, 1.0, r)
+                    }
+                })
+                .collect()
+        }
+        MaskSource::Brush { strokes } => brush_values(strokes, source),
+        MaskSource::Subject { .. } | MaskSource::Category { .. } => {
+            if let Some(coverage) = &part.coverage {
+                // The lattice is a regular grid over the source, which is
+                // exactly the resample `decode_at` does.
+                coverage
+                    .decode_at(GRID, GRID)
+                    .into_iter()
+                    .map(|b| f32::from(b) / 255.0)
+                    .collect()
+            } else {
+                let raster = model(part)?;
+                if raster.width == 0
+                    || raster.height == 0
+                    || raster.values.len() < raster.width * raster.height
+                {
+                    return None;
+                }
+                lattice()
+                    .map(|(u, v)| {
+                        let x = ((u * raster.width as f32) as usize).min(raster.width - 1);
+                        let y = ((v * raster.height as f32) as usize).min(raster.height - 1);
+                        f32::from(raster.values[y * raster.width + x]) / 255.0
+                    })
+                    .collect()
+            }
+        }
+        MaskSource::Regions { .. } | MaskSource::Luminance { .. } | MaskSource::Colour { .. } => {
+            return None
+        }
+    };
+    if part.invert {
+        values.iter_mut().for_each(|w| *w = 1.0 - *w);
+    }
+    Some(values)
+}
+
+/// Strokes composited in order, the way `fs_brush` and its blend states do.
+fn brush_values(strokes: &[Stroke], source: (u32, u32)) -> Vec<f32> {
+    // Into units of the shorter edge, as `to_square` does, so a dab is round.
+    let short = source.0.min(source.1).max(1) as f32;
+    let scale = (
+        source.0.max(1) as f32 / short,
+        source.1.max(1) as f32 / short,
+    );
+    let square = |p: (f32, f32)| (p.0 * scale.0, p.1 * scale.1);
+
+    let cells: Vec<(f32, f32)> = lattice().map(square).collect();
+    let mut out = vec![0.0f32; cells.len()];
+    for stroke in strokes.iter().filter(|s| !s.is_empty()) {
+        let points: Vec<(f32, f32)> = stroke.points.iter().copied().map(square).collect();
+        let r = stroke.radius;
+        let inner = r * stroke.hardness.clamp(0.0, 1.0);
+        // Only cells inside the stroke's bounding box can be reached.
+        let (lo, hi) = points.iter().fold(
+            ((f32::MAX, f32::MAX), (f32::MIN, f32::MIN)),
+            |(lo, hi), p| {
+                (
+                    (lo.0.min(p.0), lo.1.min(p.1)),
+                    (hi.0.max(p.0), hi.1.max(p.1)),
+                )
+            },
+        );
+        for (cell, dst) in cells.iter().zip(out.iter_mut()) {
+            if cell.0 < lo.0 - r || cell.0 > hi.0 + r || cell.1 < lo.1 - r || cell.1 > hi.1 + r {
+                continue;
+            }
+            let d = if points.len() == 1 {
+                dist(*cell, points[0])
+            } else {
+                points
+                    .windows(2)
+                    .map(|s| segment_distance(*cell, s[0], s[1]))
+                    .fold(f32::MAX, f32::min)
+            };
+            let c = ((1.0 - smoothstep(inner, r, d)) * stroke.flow).clamp(0.0, 1.0);
+            *dst = if stroke.erase {
+                *dst * (1.0 - c)
+            } else {
+                *dst + c - *dst * c
+            };
+        }
+    }
+    out
+}
+
+fn dist(a: (f32, f32), b: (f32, f32)) -> f32 {
+    (a.0 - b.0).hypot(a.1 - b.1)
+}
+
+fn segment_distance(q: (f32, f32), a: (f32, f32), b: (f32, f32)) -> f32 {
+    let ab = (b.0 - a.0, b.1 - a.1);
+    let len2 = ab.0 * ab.0 + ab.1 * ab.1;
+    if len2 <= 1e-12 {
+        return dist(q, a);
+    }
+    let t = (((q.0 - a.0) * ab.0 + (q.1 - a.1) * ab.1) / len2).clamp(0.0, 1.0);
+    dist(q, (a.0 + ab.0 * t, a.1 + ab.1 * t))
+}
+
+/// WGSL's `smoothstep`, including its behaviour when the edges meet.
+fn smoothstep(e0: f32, e1: f32, x: f32) -> f32 {
+    if e1 <= e0 {
+        return if x < e0 { 0.0 } else { 1.0 };
+    }
+    let t = ((x - e0) / (e1 - e0)).clamp(0.0, 1.0);
+    t * t * (3.0 - 2.0 * t)
+}
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use super::*;
+    use crate::coverage::Coverage;
+
+    const SOURCE: (u32, u32) = (3000, 2000);
+
+    fn no_model(_: &MaskPart) -> Option<Raster<'static>> {
+        None
+    }
+
+    fn cropped(x: f32, y: f32, w: f32, h: f32) -> Framing {
+        let mut f = Framing::new();
+        f.set_crop(CropRect {
+            x,
+            y,
+            width: w,
+            height: h,
+        });
+        f
+    }
+
+    /// A small circle near the source's top-left corner.
+    fn top_left_circle() -> MaskLayer {
+        MaskLayer::new(
+            "m1",
+            MaskSource::Radial {
+                centre: (0.15, 0.15),
+                radii: (0.08, 0.08),
+                angle: 0.0,
+                feather: 0.2,
+            },
+        )
+    }
+
+    fn stack(layers: Vec<MaskLayer>) -> MaskStack {
+        let mut s = MaskStack::new();
+        for l in layers {
+            assert!(s.push(l));
+        }
+        s
+    }
+
+    fn hidden(masks: &MaskStack, before: &Framing, after: &Framing) -> Vec<String> {
+        hidden_by_crop(masks, before, after, SOURCE, &no_model)
+            .into_iter()
+            .map(|l| l.id.clone())
+            .collect()
+    }
+
+    #[test]
+    fn a_crop_away_from_a_shape_hides_it() {
+        let masks = stack(vec![top_left_circle()]);
+        let after = cropped(0.5, 0.5, 0.5, 0.5);
+        assert_eq!(hidden(&masks, &Framing::new(), &after), vec!["m1"]);
+    }
+
+    #[test]
+    fn a_crop_that_keeps_the_shape_is_silent() {
+        let masks = stack(vec![top_left_circle()]);
+        let after = cropped(0.0, 0.0, 0.6, 0.6);
+        assert!(hidden(&masks, &Framing::new(), &after).is_empty());
+    }
+
+    #[test]
+    fn a_crop_that_trims_part_of_a_shape_is_silent() {
+        // Half the circle survives, which is a recomposition, not a loss.
+        let masks = stack(vec![top_left_circle()]);
+        let after = cropped(0.15, 0.0, 0.85, 1.0);
+        assert!(hidden(&masks, &Framing::new(), &after).is_empty());
+    }
+
+    #[test]
+    fn a_layer_already_cropped_away_is_not_reported_again() {
+        let masks = stack(vec![top_left_circle()]);
+        let before = cropped(0.5, 0.5, 0.5, 0.5);
+        let after = cropped(0.6, 0.6, 0.4, 0.4);
+        assert!(hidden(&masks, &before, &after).is_empty());
+    }
+
+    #[test]
+    fn uncropping_brings_a_layer_back_and_says_nothing() {
+        let masks = stack(vec![top_left_circle()]);
+        let before = cropped(0.5, 0.5, 0.5, 0.5);
+        assert!(hidden(&masks, &before, &Framing::new()).is_empty());
+    }
+
+    #[test]
+    fn an_unchanged_frame_is_silent_whatever_the_zoom() {
+        let masks = stack(vec![top_left_circle()]);
+        let before = cropped(0.5, 0.5, 0.5, 0.5);
+        let mut after = before;
+        after.set_view(CropRect {
+            x: 0.5,
+            y: 0.5,
+            width: 0.5,
+            height: 0.5,
+        });
+        assert!(hidden(&masks, &Framing::new(), &after).len() == 1);
+        assert!(hidden(&masks, &before, &after).is_empty());
+    }
+
+    #[test]
+    fn a_range_follows_the_picture_and_is_never_stranded() {
+        let masks = stack(vec![
+            MaskLayer::new("lum", MaskSource::highlights()),
+            MaskLayer::new("skin", MaskSource::skin_tones()),
+        ]);
+        let after = cropped(0.9, 0.9, 0.1, 0.1);
+        assert!(hidden(&masks, &Framing::new(), &after).is_empty());
+    }
+
+    #[test]
+    fn a_linear_gradient_over_the_bottom_is_hidden_by_keeping_the_top() {
+        let masks = stack(vec![MaskLayer::new(
+            "grad",
+            MaskSource::Linear {
+                centre: (0.5, 0.8),
+                angle: std::f32::consts::FRAC_PI_2,
+                width: 0.05,
+            },
+        )]);
+        assert_eq!(
+            hidden(&masks, &Framing::new(), &cropped(0.0, 0.0, 1.0, 0.5)),
+            vec!["grad"]
+        );
+        assert!(hidden(&masks, &Framing::new(), &cropped(0.0, 0.5, 1.0, 0.5)).is_empty());
+    }
+
+    #[test]
+    fn a_painted_stroke_is_measured_where_it_was_painted() {
+        let mut layer = MaskLayer::new("paint", MaskSource::brush());
+        layer.begin_stroke(0, false, 0.05, 0.8, 1.0);
+        for i in 0..10 {
+            layer.extend_stroke(0, 0.8 + i as f32 * 0.01, 0.8);
+        }
+        layer.end_stroke(0);
+        let masks = stack(vec![layer]);
+        assert_eq!(
+            hidden(&masks, &Framing::new(), &cropped(0.0, 0.0, 0.5, 0.5)),
+            vec!["paint"]
+        );
+        assert!(hidden(&masks, &Framing::new(), &cropped(0.5, 0.5, 0.5, 0.5)).is_empty());
+    }
+
+    #[test]
+    fn an_unpainted_brush_has_nothing_to_lose() {
+        let masks = stack(vec![MaskLayer::new("empty", MaskSource::brush())]);
+        assert!(hidden(&masks, &Framing::new(), &cropped(0.0, 0.0, 0.3, 0.3)).is_empty());
+    }
+
+    #[test]
+    fn an_inverted_layer_is_measured_as_what_it_selects() {
+        // Everything *but* a corner circle: a crop into the other corner
+        // keeps most of it.
+        let mut layer = top_left_circle();
+        layer.invert = true;
+        let masks = stack(vec![layer]);
+        assert!(hidden(&masks, &Framing::new(), &cropped(0.5, 0.5, 0.5, 0.5)).is_empty());
+    }
+
+    #[test]
+    fn a_subject_is_measured_from_its_stored_coverage_or_the_model() {
+        // A subject occupying the right-hand quarter of a 40x20 proxy.
+        let (w, h) = (40usize, 20usize);
+        let bytes: Vec<u8> = (0..w * h)
+            .map(|i| if i % w >= 30 { 255 } else { 0 })
+            .collect();
+        let subject = MaskSource::Subject {
+            signature: 1,
+            index: 0,
+            class: "dog".into(),
+            score: 0.9,
+        };
+        let left = cropped(0.0, 0.0, 0.5, 1.0);
+
+        let mut stored = MaskLayer::new("stored", subject.clone());
+        stored.base_mut().coverage = Coverage::encode(&bytes, w, h, 2).map(Arc::new);
+        let live = MaskLayer::new("live", subject);
+        let masks = stack(vec![stored, live]);
+
+        // No model to hand: only the layer carrying its raster is measured.
+        assert_eq!(hidden(&masks, &Framing::new(), &left), vec!["stored"]);
+
+        let model = |_: &MaskPart| {
+            Some(Raster {
+                values: Cow::Borrowed(bytes.as_slice()),
+                width: w,
+                height: h,
+            })
+        };
+        let ids: Vec<_> = hidden_by_crop(&masks, &Framing::new(), &left, SOURCE, &model)
+            .into_iter()
+            .map(|l| l.id.as_str())
+            .collect();
+        assert_eq!(ids, vec!["stored", "live"]);
+    }
+
+    #[test]
+    fn a_crop_is_read_in_the_turned_frame() {
+        // Turned a quarter clockwise, the source's top-left lands at the
+        // frame's top-right, so keeping the right half of the frame keeps it.
+        let masks = stack(vec![top_left_circle()]);
+        let mut before = Framing::new();
+        before.rotate_quarters(1);
+        let mut right = before;
+        right.set_crop(CropRect {
+            x: 0.5,
+            y: 0.0,
+            width: 0.5,
+            height: 1.0,
+        });
+        let mut left = before;
+        left.set_crop(CropRect {
+            x: 0.0,
+            y: 0.0,
+            width: 0.5,
+            height: 1.0,
+        });
+        assert!(hidden(&masks, &before, &right).is_empty());
+        assert_eq!(hidden(&masks, &before, &left), vec!["m1"]);
+    }
+
+    #[test]
+    fn a_subtracted_range_does_not_hide_the_shape_it_cuts() {
+        let mut layer = top_left_circle();
+        assert!(layer.push_part(MaskPart::new(
+            "p2",
+            Join::Subtract,
+            MaskSource::highlights()
+        )));
+        let masks = stack(vec![layer.clone()]);
+        assert_eq!(
+            hidden(&masks, &Framing::new(), &cropped(0.5, 0.5, 0.5, 0.5)),
+            vec!["m1"]
+        );
+        // Added instead, the range reaches everywhere and nothing is lost.
+        let mut layer = top_left_circle();
+        assert!(layer.push_part(MaskPart::new("p2", Join::Union, MaskSource::highlights())));
+        let masks = stack(vec![layer]);
+        assert!(hidden(&masks, &Framing::new(), &cropped(0.5, 0.5, 0.5, 0.5)).is_empty());
+    }
+
+    #[test]
+    fn an_intersection_keeps_only_what_both_parts_cover() {
+        let circle = |id: &str, join, centre, r| {
+            MaskPart::new(
+                id,
+                join,
+                MaskSource::Radial {
+                    centre,
+                    radii: (r, r),
+                    angle: 0.0,
+                    feather: 0.2,
+                },
+            )
+        };
+        // Two circles in opposite corners: keeping the bottom-right quarter
+        // keeps one of them, and nothing is lost.
+        let mut layer = top_left_circle();
+        assert!(layer.push_part(circle("p2", Join::Union, (0.85, 0.85), 0.08)));
+        let bottom_right = cropped(0.5, 0.5, 0.5, 0.5);
+        let masks = stack(vec![layer.clone()]);
+        assert!(hidden(&masks, &Framing::new(), &bottom_right).is_empty());
+        // Intersected with a disc around the top-left, only that corner's
+        // circle survives, and the same crop takes it out of the frame.
+        assert!(layer.push_part(circle("p3", Join::Intersect, (0.15, 0.15), 0.3)));
+        let masks = stack(vec![layer]);
+        assert_eq!(hidden(&masks, &Framing::new(), &bottom_right), vec!["m1"]);
+    }
+}

core/dr-pipeline/src/preset.rs

@@ -725,6 +725,7 @@ mod tests {
             height: 0.5,
         });
         g.set_param(framing::ID, framing::ANGLE, -2.0);
+        g.set_param(framing::ID, framing::KEYSTONE_V, 40.0);
         g
     }
 
@@ -820,6 +821,14 @@ mod tests {
             Some(0.0),
             "the source's straightening must not travel on this scope"
         );
+        // TRACES: FR-DEV-20
+        // Perspective is composition on the same terms as the crop: it is
+        // withheld by the default scope, not pasted over the target's frame.
+        assert_eq!(
+            target.param(framing::ID, framing::KEYSTONE_V),
+            Some(0.0),
+            "the source's keystone must not travel on this scope"
+        );
     }
 
     #[test]
@@ -829,6 +838,7 @@ mod tests {
         preset.apply(&mut target, Scope::everything());
 
         assert_eq!(target.param(framing::ID, framing::ANGLE), Some(-2.0));
+        assert_eq!(target.param(framing::ID, framing::KEYSTONE_V), Some(40.0));
         assert!(
             (target.crop().width - 0.5).abs() < 1e-5,
             "{:?}",

core/dr-pipeline/src/sidecar.rs

@@ -93,6 +93,9 @@ pub const MAX_RATING: u8 = 5;
 /// Highest flag code: 0 unflagged, 1 pick, 2 reject.
 pub const MAX_FLAG: u8 = 2;
 
+/// Highest colour-label code: purple. Mirrors `dr_catalog::rating::label_code`.
+pub const MAX_LABEL: u8 = 5;
+
 /// TRACES: FR-CAT-8 | FR-NC-8
 /// One image's sidecar: a keyed set of versions.
 ///
@@ -187,6 +190,17 @@ pub struct Version {
     /// so a value moving between the two stores needs no translation table
     /// that could drift.
     pub flag: u8,
+    /// TRACES: FR-CAT-5
+    /// The colour label, `0` for none and `1..=5` red, yellow, green, blue,
+    /// purple — the catalog's `versions.label` codes, for the reason
+    /// [`Self::flag`] shares its encoding.
+    ///
+    /// Here for the reason the rating is: a label that lived only in the
+    /// catalog would go with the catalog, and would never reach the
+    /// photographer's other devices, which learn judgements from this file.
+    /// A build that predates the key keeps it as an unknown line and writes
+    /// it back, so an older device passes it on rather than erasing it.
+    pub label: u8,
     /// The edit itself: `(op, param) -> value`, non-default values only.
     pub params: BTreeMap<(String, String), f32>,
     /// TRACES: FR-DEV-3 | FR-NC-9
@@ -264,6 +278,7 @@ impl Version {
             // it in the "not yet looked at" state a cull resumes from.
             rating: 0,
             flag: 0,
+            label: 0,
             params,
             masks,
             film,
@@ -572,6 +587,10 @@ impl Version {
         // that never had it.
         self.rating = merge_judgement(self.rating, remote.rating, remote_wins);
         self.flag = merge_judgement(self.flag, remote.flag, remote_wins);
+        // TRACES: FR-CAT-5
+        // A label under the same rule: 0 is "none given", so a device that
+        // never labelled a frame cannot clear another's label.
+        self.label = merge_judgement(self.label, remote.label, remote_wins);
 
         // TRACES: FR-DEV-3f
         // The film resolves wholesale to the higher revision, like a mask
@@ -873,6 +892,9 @@ impl Sidecar {
             if v.flag > 0 {
                 let _ = writeln!(out, "flag = {}", v.flag);
             }
+            if v.label > 0 {
+                let _ = writeln!(out, "label = {}", v.label);
+            }
             // TRACES: FR-DEV-3f
             // Before the parameters, because it decides what they mean: the
             // film's exposure slider is a slider on *that stock's* curve.
@@ -1068,6 +1090,17 @@ impl Sidecar {
                         Some(value.to_string());
                 }
                 "flag" => version.flag = value.parse::<u8>().unwrap_or(0).min(MAX_FLAG),
+                // TRACES: FR-CAT-5
+                // A code this build does not know reads as none rather than
+                // being clamped onto purple: a wrong colour is a claim, and
+                // no colour is only a gap.
+                "label" => {
+                    version.label = value
+                        .parse::<u8>()
+                        .ok()
+                        .filter(|l| *l <= MAX_LABEL)
+                        .unwrap_or(0)
+                }
                 // TRACES: FR-DEV-8
                 // Ahead of the `op.param` arm below, which would otherwise try
                 // to read eight numbers as one float and drop the repair with a
@@ -2141,6 +2174,8 @@ mod tests {
             height: 0.6,
         });
         g.set_param(framing::ID, framing::ANGLE, -1.5);
+        g.set_param(framing::ID, framing::KEYSTONE_V, 42.0);
+        g.set_param(framing::ID, framing::KEYSTONE_H, -17.0);
         g.rotate_quarters(1);
 
         let mut sidecar = Sidecar::new();
@@ -2157,6 +2192,12 @@ mod tests {
         assert_eq!(restored.crop(), g.crop());
         assert_eq!(restored.param(framing::ID, framing::ANGLE), Some(-1.5));
         assert_eq!(restored.param(framing::ID, framing::ROTATION), Some(1.0));
+        // TRACES: FR-DEV-20
+        assert_eq!(restored.param(framing::ID, framing::KEYSTONE_V), Some(42.0));
+        assert_eq!(
+            restored.param(framing::ID, framing::KEYSTONE_H),
+            Some(-17.0)
+        );
     }
 
     #[test]
@@ -2828,6 +2869,37 @@ mod tests {
         assert_eq!(back.flag, 1);
     }
 
+    #[test]
+    fn a_label_survives_the_round_trip_and_an_unknown_one_reads_as_none() {
+        // TRACES: FR-CAT-5
+        let mut v = version_of(&EditGraph::default_chain());
+        v.label = 3;
+        let mut sidecar = Sidecar::new();
+        sidecar.put(v);
+        let text = sidecar.to_text();
+        assert!(text.contains("label = 3"), "{text}");
+        let parsed = Sidecar::parse(&text).expect("valid");
+        assert_eq!(parsed.default_version().expect("a version").label, 3);
+
+        let odd = "drsc 1\n\n[version u1]\nname = Default\nrevision = 1\nmodified = 0\n\
+                   label = 9\n";
+        let parsed = Sidecar::parse(odd).expect("valid");
+        assert_eq!(parsed.default_version().expect("a version").label, 0);
+    }
+
+    #[test]
+    fn an_unlabelled_device_cannot_clear_another_devices_label() {
+        // TRACES: FR-CAT-5
+        let mut local = version_of(&EditGraph::default_chain());
+        local.label = 0;
+        local.revision = 9;
+        let mut remote = version_of(&EditGraph::default_chain());
+        remote.label = 1;
+        remote.revision = 1;
+        local.merge(&remote, None);
+        assert_eq!(local.label, 1);
+    }
+
     #[test]
     fn an_unrated_image_writes_no_judgement_lines() {
         // The non-default rule applied to judgement: a library that has never
@@ -2838,6 +2910,7 @@ mod tests {
         let text = sidecar.to_text();
         assert!(!text.contains("rating"), "{text}");
         assert!(!text.contains("flag"), "{text}");
+        assert!(!text.contains("label"), "{text}");
     }
 
     #[test]

core/dr-pipeline/tests/mask_sidecar.rs

@@ -1344,6 +1344,28 @@ fn a_correction_painted_onto_a_subject_survives_a_round_trip() {
     );
 }
 
+/// TRACES: FR-DEV-19a
+/// An intersecting part is written under its own word and reads back as an
+/// intersection. A build from before intersection read that word as a union
+/// (see `an_unknown_join_adds_the_part`), which keeps the part visible and
+/// fixable rather than silently cutting the mask down.
+#[test]
+fn an_intersecting_part_survives_a_round_trip() {
+    let mut graph = EditGraph::default_chain();
+    graph.masks_mut().push(corrected("m1", Join::Intersect));
+
+    let mut sidecar = Sidecar::new();
+    sidecar.put(Version::from_graph("default", "Default", &graph));
+    let text = sidecar.to_text();
+    assert!(text.contains("join = intersect"), "{text}");
+
+    let restored = round_trip(&graph);
+    let layer = &restored.masks().layers()[0];
+    assert_eq!(layer.parts().len(), 2);
+    assert_eq!(layer.parts()[1].join, Join::Intersect);
+    assert_eq!(layer.parts()[0].join, Join::Union, "the base is untouched");
+}
+
 /// TRACES: FR-DEV-19a
 /// A part left out of the build comes back left out, and the file says so
 /// under a word that cannot be confused with the layer's own switch.

core/dr-sync-nextcloud/src/auth.rs

@@ -22,6 +22,11 @@ pub struct LoginFlow {
     pub login_url: String,
     #[serde(rename = "poll")]
     pub poll: PollInfo,
+    /// The server the flow was started against — the address the user typed,
+    /// already normalised. Not part of the response: [`begin`] fills it in so
+    /// [`poll`] can put it in the credentials instead of the server's answer.
+    #[serde(skip)]
+    pub server: String,
 }
 
 #[derive(Debug, Clone, Deserialize)]
@@ -66,9 +71,57 @@ pub async fn begin(
         });
     }
 
-    resp.json::<LoginFlow>()
+    let mut flow = resp
+        .json::<LoginFlow>()
         .await
-        .map_err(|e| RemoteError::Protocol(e.to_string()))
+        .map_err(|e| RemoteError::Protocol(e.to_string()))?;
+
+    // Both URLs are the server's to choose, and neither may be trusted as
+    // sent. The login URL is handed to the operating system to open, where a
+    // `file:` or UNC path is a program launch rather than a web page; the poll
+    // endpoint is where the app password comes back from.
+    flow.login_url = upgraded("login URL", &flow.login_url)?;
+    flow.poll.endpoint = upgraded("poll endpoint", &flow.poll.endpoint)?;
+    flow.server = server.trim_end_matches('/').to_string();
+    Ok(flow)
+}
+
+/// TRACES: NFR-SEC-3
+/// A URL the server sent, upgraded to HTTPS, or refused.
+///
+/// `http` is upgraded rather than refused: a Nextcloud behind a TLS-terminating
+/// proxy without `overwriteprotocol` builds every absolute URL it returns with
+/// `http`, and the same path over `https` is the one that works. Any other
+/// scheme is refused, because the only thing it could be for is reaching
+/// something that is not this server.
+///
+/// The host is not checked. A server reached by its LAN address may answer with
+/// its public name, and nothing here is safer for refusing that: the account
+/// is stored under the address the user typed (see [`poll`]), not under
+/// anything the server said.
+pub(crate) fn upgraded(what: &str, url: &str) -> Result<String, RemoteError> {
+    let mut parsed = url::Url::parse(url).map_err(|e| {
+        RemoteError::Protocol(format!(
+            "the server sent a {what} that is not a URL ({e}): {url}"
+        ))
+    })?;
+    match parsed.scheme() {
+        "https" => {}
+        "http" => parsed.set_scheme("https").map_err(|()| {
+            RemoteError::Protocol(format!("{what} cannot be upgraded to https: {url}"))
+        })?,
+        other => {
+            return Err(RemoteError::Protocol(format!(
+                "the server sent a {what} using {other}:, and only https is accepted: {url}"
+            )))
+        }
+    }
+    if parsed.host_str().is_none_or(str::is_empty) {
+        return Err(RemoteError::Protocol(format!(
+            "the server sent a {what} with no host: {url}"
+        )));
+    }
+    Ok(parsed.into())
 }
 
 /// Poll until the user finishes authenticating in the browser.
@@ -98,10 +151,11 @@ pub async fn poll(
 
         match resp.status().as_u16() {
             200 => {
-                return resp
+                let creds = resp
                     .json::<AppCredentials>()
                     .await
-                    .map_err(|e| RemoteError::Protocol(e.to_string()))
+                    .map_err(|e| RemoteError::Protocol(e.to_string()))?;
+                return Ok(under_typed_server(creds, &flow.server));
             }
             // Still waiting for the user.
             404 => tokio::time::sleep(POLL_INTERVAL).await,
@@ -117,6 +171,26 @@ pub async fn poll(
     Err(RemoteError::AuthFailed)
 }
 
+/// TRACES: NFR-SEC-3
+/// Credentials filed under the address the user typed, not the one the server
+/// reports.
+///
+/// That report is the server's idea of its own URL, and behind a proxy
+/// without `overwriteprotocol` it says `http://` — which, stored, would send
+/// the app password in the clear on every request from then on. The typed
+/// address has just carried the whole flow, so it is known to reach the
+/// server.
+fn under_typed_server(mut creds: AppCredentials, server: &str) -> AppCredentials {
+    if creds.server.trim_end_matches('/') != server {
+        log::info!(
+            "server reports itself as {}; keeping {server}",
+            creds.server
+        );
+    }
+    creds.server = server.to_string();
+    creds
+}
+
 /// Percent-encode a form value.
 fn urlencode(s: &str) -> String {
     s.bytes()
@@ -167,6 +241,49 @@ mod tests {
         assert!(flow.login_url.contains("/login/v2/flow/"));
     }
 
+    /// TRACES: NFR-SEC-3
+    #[test]
+    fn server_urls_are_upgraded_to_https_or_refused() {
+        assert_eq!(
+            upgraded("login URL", "http://cloud.example/login/v2/flow/xyz").unwrap(),
+            "https://cloud.example/login/v2/flow/xyz"
+        );
+        // A non-default port stays, and only the scheme changes.
+        assert_eq!(
+            upgraded("poll endpoint", "http://cloud.example:8443/login/v2/poll").unwrap(),
+            "https://cloud.example:8443/login/v2/poll"
+        );
+        assert_eq!(
+            upgraded("login URL", "https://cloud.example/x").unwrap(),
+            "https://cloud.example/x"
+        );
+        // What `rundll32 url.dll,FileProtocolHandler` would run, and what
+        // `xdg-open` would hand to whatever claims it.
+        for hostile in [
+            "file:///C:/Windows/System32/calc.exe",
+            "\\\\evil\\share\\x.exe",
+            "C:\\x.exe",
+            "javascript:alert(1)",
+            "-v",
+            "",
+        ] {
+            assert!(upgraded("login URL", hostile).is_err(), "{hostile:?}");
+        }
+    }
+
+    /// TRACES: NFR-SEC-3
+    #[test]
+    fn credentials_keep_the_typed_server_not_the_reported_one() {
+        let reported = AppCredentials {
+            server: "http://cloud.example".into(),
+            login_name: "duncan".into(),
+            app_password: "secret-token".into(),
+        };
+        let c = under_typed_server(reported, "https://cloud.example");
+        assert_eq!(c.server, "https://cloud.example");
+        assert_eq!(c.app_password, "secret-token");
+    }
+
     #[test]
     fn form_values_are_encoded() {
         assert_eq!(urlencode("abc123"), "abc123");

core/dr-sync-nextcloud/src/lib.rs

@@ -723,6 +723,15 @@ pub fn http_client(user_agent: &str) -> Result<reqwest::Client, RemoteError> {
         // for months. [`EXTRA_ROOTS`] carries those, and is additive — the
         // webpki-roots set is still installed alongside.
         .tls_certs_only(extra_roots())
+        // TRACES: NFR-SEC-3
+        // Refuse `http` here, below every URL this crate builds, rather than
+        // trusting each place a URL comes from. `normalise_endpoint` upgrades
+        // what the user types, but the login flow's poll endpoint, the account
+        // an older build saved and a redirect all arrive from somewhere else,
+        // and any of them naming `http://` would send the app password in
+        // Basic auth in the clear. reqwest checks this before connecting and
+        // again on every redirect, so a refused request never opens a socket.
+        .https_only(true)
         // A request that hangs forever is indistinguishable from a worker that
         // died, and cost a long time to tell apart once. These turn that into
         // an error the UI can show.
@@ -780,8 +789,16 @@ fn map_send_error(e: reqwest::Error) -> RemoteError {
     let mut detail = e.to_string();
     let mut src: Option<&dyn std::error::Error> = std::error::Error::source(&e);
     let mut tls = false;
+    // A URL the client refused to send — `http` under `https_only`, on the
+    // first request or a redirect. Nothing left the process, so this is the
+    // account's configuration, not the network: reported as `Network` it
+    // would put the app into offline mode over a connection that is fine.
+    let mut refused = e.is_builder();
     while let Some(s) = src {
         let text = s.to_string();
+        if text.contains("URL scheme is not allowed") {
+            refused = true;
+        }
         // rustls surfaces every verification failure through this wording:
         // UnknownIssuer, Expired, NotValidForName, BadSignature.
         if text.contains("invalid peer certificate")
@@ -795,7 +812,9 @@ fn map_send_error(e: reqwest::Error) -> RemoteError {
         src = std::error::Error::source(s);
     }
 
-    if tls {
+    if refused {
+        RemoteError::Configuration(detail)
+    } else if tls {
         RemoteError::Tls(detail)
     } else {
         RemoteError::Network(detail)
@@ -1013,6 +1032,35 @@ mod tests {
         assert!(c.is_ok(), "client must build without a backend");
     }
 
+    /// TRACES: NFR-SEC-3
+    #[tokio::test]
+    async fn plain_http_is_refused_before_a_connection_opens() {
+        // A listener that would take the connection if one were made. The
+        // app password travels in a header, so a request that reached the
+        // socket has already leaked it; failing on the response is too late.
+        let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap();
+        listener.set_nonblocking(true).unwrap();
+        let url = format!("http://{}/remote.php/dav/", listener.local_addr().unwrap());
+
+        let err = http_client("test")
+            .unwrap()
+            .get(&url)
+            .basic_auth("duncan", Some("app-password"))
+            .send()
+            .await
+            .expect_err("http must be refused");
+
+        assert!(
+            matches!(map_send_error(err), RemoteError::Configuration(_)),
+            "a refused scheme is configuration, not the network"
+        );
+        assert_eq!(
+            listener.accept().err().map(|e| e.kind()),
+            Some(std::io::ErrorKind::WouldBlock),
+            "nothing may have connected"
+        );
+    }
+
     #[tokio::test]
     async fn delta_is_unsupported_and_says_why() {
         // Verified absent in the server; the engine must fall back rather

core/dr-sync-nextcloud/src/provider.rs

@@ -96,6 +96,18 @@ impl BackendProvider for NextcloudProvider {
         }
     }
 
+    /// TRACES: NFR-SEC-3
+    /// An `http://` account written before sign-in kept the address the user
+    /// typed: it was stored as the server reported itself, and behind a proxy
+    /// without `overwriteprotocol` that is `http`. The client refuses to send
+    /// to it now, so it is upgraded here rather than left to fail. The
+    /// namespace ignores the scheme, so the catalog stays where it is.
+    fn upgrade_endpoint(&self, stored: &str) -> Option<String> {
+        stored
+            .strip_prefix("http://")
+            .map(|rest| format!("https://{rest}"))
+    }
+
     fn connect(&self, conn: &Connection) -> Result<Box<dyn RemoteBackend>, RemoteError> {
         let creds = Self::credentials(conn)?;
         Ok(Box::new(NextcloudBackend::new(
@@ -132,6 +144,18 @@ mod tests {
         assert!(p.normalise_endpoint("  ").is_err());
     }
 
+    /// TRACES: NFR-SEC-3
+    #[test]
+    fn a_stored_http_endpoint_is_upgraded_and_nothing_else_is_touched() {
+        let p = NextcloudProvider;
+        assert_eq!(
+            p.upgrade_endpoint("http://cloud.example/nextcloud")
+                .as_deref(),
+            Some("https://cloud.example/nextcloud")
+        );
+        assert_eq!(p.upgrade_endpoint("https://cloud.example"), None);
+    }
+
     #[test]
     fn the_account_keeps_the_dav_user_id_apart_from_the_login() {
         // A login can be an email address while the user id is something

core/dr-sync/src/account.rs

@@ -29,7 +29,7 @@ use dr_plat::{SecretError, SecretRef, SecretStore};
 use dr_types::{Format, FormatFilter};
 use serde::{Deserialize, Serialize};
 
-use crate::RemoteError;
+use crate::{BackendRegistry, RemoteError};
 
 /// The connector every account had before there was a choice.
 ///
@@ -510,6 +510,95 @@ impl AccountStore {
         written
     }
 
+    /// TRACES: NFR-SEC-3
+    /// Rewrite the endpoints an older build stored in a form this one would
+    /// not, asking each account's connector
+    /// ([`upgrade_endpoint`](crate::BackendProvider::upgrade_endpoint)).
+    ///
+    /// Per account and best effort: one that cannot be moved — its keyring
+    /// locked, say — is logged and left as it was, and is tried again on the
+    /// next launch, rather than stopping the others or the launch. Returns
+    /// the accounts it rewrote.
+    pub fn upgrade_endpoints(&self, registry: &BackendRegistry) -> Vec<Account> {
+        let mut upgraded = Vec::new();
+        for account in self.list() {
+            let Ok(provider) = registry.for_account(&account) else {
+                continue;
+            };
+            let Some(endpoint) = provider.upgrade_endpoint(&account.endpoint) else {
+                continue;
+            };
+            if endpoint == account.endpoint {
+                continue;
+            }
+            match self.move_endpoint(&account, &endpoint) {
+                Ok(moved) => {
+                    log::info!("account {} moved to {endpoint}", account.describe());
+                    upgraded.push(moved);
+                }
+                Err(e) => log::warn!(
+                    "account {} could not be moved to {endpoint}: {e}",
+                    account.describe()
+                ),
+            }
+        }
+        upgraded
+    }
+
+    /// Move an account to a new endpoint, taking its credential with it.
+    ///
+    /// The endpoint is half of two keys, and both have to be dealt with. The
+    /// credential is filed under it ([`Account::secret_ref`]), so rewriting
+    /// the record alone would strand the app password under the old key and
+    /// sign the user out. And it feeds [`Account::namespace`], so a rewrite
+    /// that changed the namespace would abandon the catalog and everything
+    /// beside it; that is refused outright rather than left to the caller.
+    ///
+    /// Ordered so an interruption at any step leaves something that works:
+    /// the credential is copied before the record names the new key, and the
+    /// old copy is deleted only once nothing names the old one.
+    fn move_endpoint(&self, account: &Account, endpoint: &str) -> Result<Account, AccountError> {
+        let mut moved = account.clone();
+        moved.endpoint = endpoint.to_string();
+        if moved.namespace() != account.namespace() {
+            return Err(AccountError::WouldMoveData {
+                from: account.namespace(),
+                to: moved.namespace(),
+            });
+        }
+
+        let mut config = self.read_config();
+        // Already there — the user signed in again at the new address. That
+        // record and its credential are the newer, so the old one just goes.
+        let duplicate = config.sessions.iter().any(|a| a.is_same_as(&moved));
+
+        let old_ref = account.secret_ref();
+        let secret = match self.secrets.retrieve(&old_ref) {
+            Ok(s) => Some(s),
+            Err(SecretError::NotFound) => None,
+            Err(e) => return Err(e.into()),
+        };
+        if let (Some(s), false) = (&secret, duplicate) {
+            self.secrets.store(&moved.secret_ref(), s)?;
+        }
+
+        if duplicate {
+            config.sessions.retain(|a| !a.is_same_as(account));
+        } else {
+            // In place, not removed and pushed: the last record is the one
+            // the next launch resumes.
+            for a in config.sessions.iter_mut().filter(|a| a.is_same_as(account)) {
+                *a = moved.clone();
+            }
+        }
+        self.write_config(&config)?;
+
+        if secret.is_some() {
+            self.secrets.delete(&old_ref)?;
+        }
+        Ok(moved)
+    }
+
     fn read_config(&self) -> ConfigFile {
         std::fs::read_to_string(&self.config_path)
             .ok()
@@ -545,6 +634,11 @@ pub enum AccountError {
 
     #[error(transparent)]
     Remote(#[from] RemoteError),
+
+    /// A change that would give an account a different local data directory,
+    /// leaving its catalog and caches behind under the old one.
+    #[error("moving the account would leave its local data behind ({from} → {to})")]
+    WouldMoveData { from: String, to: String },
 }
 
 #[cfg(test)]
@@ -574,6 +668,134 @@ mod tests {
         d
     }
 
+    /// A connector that upgrades `http://` endpoints the way Nextcloud's does,
+    /// and every other one not at all.
+    struct Upgrading(&'static str);
+    impl crate::BackendProvider for Upgrading {
+        fn id(&self) -> &'static str {
+            self.0
+        }
+        fn display_name(&self) -> &'static str {
+            "U"
+        }
+        fn endpoint_label(&self) -> &'static str {
+            "Server"
+        }
+        fn endpoint_placeholder(&self) -> &'static str {
+            ""
+        }
+        fn sign_in(&self) -> crate::SignIn {
+            crate::SignIn::Browser
+        }
+        fn normalise_endpoint(&self, i: &str) -> Result<String, String> {
+            Ok(i.into())
+        }
+        fn upgrade_endpoint(&self, stored: &str) -> Option<String> {
+            stored
+                .strip_prefix("http://")
+                .map(|rest| format!("https://{rest}"))
+        }
+        fn connect(&self, _: &Connection) -> Result<Box<dyn crate::RemoteBackend>, RemoteError> {
+            Err(RemoteError::Unsupported("stub"))
+        }
+    }
+
+    fn upgrading(id: &'static str) -> BackendRegistry {
+        let mut r = BackendRegistry::new();
+        r.register(std::sync::Arc::new(Upgrading(id)));
+        r
+    }
+
+    /// TRACES: NFR-SEC-3
+    #[test]
+    fn an_http_account_is_upgraded_and_keeps_its_credential_and_data() {
+        let dir = tmpdir("upgrade");
+        let store = store_in(&dir);
+        let old =
+            Account::new(LEGACY_BACKEND, "http://cloud.example").with_login("duncan", "duncan");
+        store.save(&folder(), None).unwrap();
+        store
+            .save(&old, Some(&Secret::new("secret-token")))
+            .unwrap();
+
+        let moved = store.upgrade_endpoints(&upgrading(LEGACY_BACKEND));
+        assert_eq!(moved.len(), 1);
+
+        let now = store.current().expect("still the account resumed");
+        assert_eq!(now.endpoint, "https://cloud.example");
+        assert_eq!(
+            now.namespace(),
+            old.namespace(),
+            "the catalog directory must not move"
+        );
+        assert_eq!(
+            store
+                .connection(&now, true)
+                .unwrap()
+                .require_secret()
+                .unwrap()
+                .expose(),
+            "secret-token",
+            "the credential moves with the account"
+        );
+        assert!(
+            store.connection(&old, true).is_err(),
+            "nothing is left under the old key"
+        );
+        assert_eq!(store.list().len(), 2, "the folder library is untouched");
+
+        // And a second launch has nothing to do.
+        assert!(store
+            .upgrade_endpoints(&upgrading(LEGACY_BACKEND))
+            .is_empty());
+    }
+
+    /// TRACES: NFR-SEC-3
+    #[test]
+    fn a_move_that_would_change_the_data_directory_is_refused() {
+        // A scheme change keeps the namespace, which is what makes the upgrade
+        // safe; a host change does not, and would give the library a new,
+        // empty data directory. The account is left exactly as it was.
+        let dir = tmpdir("upgrade-refused");
+        let store = store_in(&dir);
+        let old = nextcloud();
+        store
+            .save(&old, Some(&Secret::new("secret-token")))
+            .unwrap();
+
+        assert!(matches!(
+            store.move_endpoint(&old, "https://elsewhere.example"),
+            Err(AccountError::WouldMoveData { .. })
+        ));
+        assert_eq!(store.current().unwrap().endpoint, old.endpoint);
+        assert!(store.connection(&old, true).is_ok());
+    }
+
+    /// TRACES: NFR-SEC-3
+    #[test]
+    fn an_upgrade_onto_an_existing_sign_in_keeps_the_newer_one() {
+        let dir = tmpdir("upgrade-duplicate");
+        let store = store_in(&dir);
+        let old =
+            Account::new(LEGACY_BACKEND, "http://cloud.example").with_login("duncan", "duncan");
+        let new =
+            Account::new(LEGACY_BACKEND, "https://cloud.example").with_login("duncan", "duncan");
+        store.save(&old, Some(&Secret::new("stale"))).unwrap();
+        store.save(&new, Some(&Secret::new("fresh"))).unwrap();
+
+        store.upgrade_endpoints(&upgrading(LEGACY_BACKEND));
+        assert_eq!(store.list().len(), 1);
+        assert_eq!(
+            store
+                .connection(&new, true)
+                .unwrap()
+                .require_secret()
+                .unwrap()
+                .expose(),
+            "fresh"
+        );
+    }
+
     #[test]
     fn a_saved_account_survives_reopening() {
         let dir = tmpdir("survives");

core/dr-sync/src/provider.rs

@@ -83,6 +83,20 @@ pub trait BackendProvider: Send + Sync {
     /// wrong rather than naming a type.
     fn normalise_endpoint(&self, input: &str) -> Result<String, String>;
 
+    /// The form a *stored* endpoint should take now, where an older build
+    /// wrote one this build would not.
+    ///
+    /// Not [`normalise_endpoint`](Self::normalise_endpoint) run again: that
+    /// judges what a person typed, and may touch the world to do it — a folder
+    /// is canonicalised and must exist — so rerunning it on every launch would
+    /// fail a library whose disk is unplugged, or rename one whose path now
+    /// resolves differently. This is a pure rewrite of the string, and `None`
+    /// means leave it alone, which is the answer for almost every connector.
+    fn upgrade_endpoint(&self, stored: &str) -> Option<String> {
+        let _ = stored;
+        None
+    }
+
     /// Build an account from a normalised endpoint alone.
     ///
     /// Only meaningful for [`SignIn::EndpointOnly`]; a browser flow produces

core/dr-types/src/place.rs

@@ -115,6 +115,10 @@ pub enum PlaceScope {
 #[serde(default)]
 pub struct StoredFilter {
     pub min_rating: u8,
+    /// TRACES: FR-UI-5
+    /// The top of a star range. A record from before ranges existed has none,
+    /// which reads as no ceiling — what it meant when it was written.
+    pub max_rating: Option<u8>,
     pub unjudged: bool,
     pub flag: Option<FlagState>,
     pub local_only: bool,
@@ -131,6 +135,10 @@ pub struct StoredFilter {
     /// Travels: it is a narrowing like `local_only`, and a record without it
     /// — from a build before it existed — reads as off.
     pub eyes_open: bool,
+    /// TRACES: FR-CAT-6
+    /// The colour label the grid was narrowed to. A record without it reads
+    /// as none, as every term added after the first does.
+    pub label: Option<crate::ColourLabel>,
 }
 
 /// Where the photographer was, at the moment they were there.

docker/android/assemble-apk.sh

@@ -332,6 +332,29 @@ else
     echo "    assets: no models found (face indexing and the scene tab will be off on the device)"
 fi
 
+# The manual: the rendered page and its pictures, read in place by
+# ManualActivity's WebView as file:///android_asset/manual/index.html. Not
+# unpacked like the models: a WebView reads an asset straight out of the APK,
+# and relative links to media/ resolve inside the same asset tree, so the page
+# costs no first-launch copy and no second copy on /data.
+#
+# About 27 MB, stored below like the models — a GIF or PNG is already
+# compressed, and deflating it again buys nothing. The pictures are LFS
+# objects, and unlike a missing model a pointer would not fail loudly: it
+# ships as a manual full of broken images. So it stops the build here.
+rm -rf "${OUT}/staging/assets/manual"
+mkdir -p "${OUT}/staging/assets/manual/media"
+cp "${REPO}/docs/manual/index.html" "${OUT}/staging/assets/manual/"
+for f in "${REPO}/docs/manual/media"/*; do
+    if head -c 40 "${f}" | grep -q '^version https://git-lfs'; then
+        echo "error: $(basename "${f}") is an LFS pointer, not a picture." >&2
+        echo "       run: git lfs pull --include='docs/manual/media/**'" >&2
+        exit 1
+    fi
+    cp "${f}" "${OUT}/staging/assets/manual/media/"
+done
+echo "    manual: index.html and $(ls "${OUT}/staging/assets/manual/media" | wc -l) picture(s), $(du -sh "${OUT}/staging/assets/manual" | cut -f1)"
+
 # -0 "" stores the .so without compression so Android can mmap it directly
 # (extractNativeLibs=false territory); for a 37 MB library that also keeps
 # install times sane.

docker/windows/package.sh

@@ -69,6 +69,23 @@ for dir in face scene inpaint; do
 done
 echo "==> staged $(ls "${STAGE}/models" | wc -l) model file(s)"
 
+# The manual: the rendered page and its pictures, beside the executable where
+# dr_ui::manual looks on Windows (dr_plat::system_data_dirs is the exe's own
+# directory there). The pictures are LFS objects; a pointer is ~130 bytes of
+# text that every browser draws as a broken image, so refuse it here rather
+# than ship a manual with no pictures in it.
+mkdir -p "${STAGE}/manual/media"
+cp "${REPO}/docs/manual/index.html" "${STAGE}/manual/"
+for f in "${REPO}/docs/manual/media"/*; do
+    if head -c 40 "${f}" | grep -q '^version https://git-lfs'; then
+        echo "error: $(basename "${f}") is an LFS pointer, not a picture." >&2
+        echo "       run: git lfs pull --include='docs/manual/media/**'" >&2
+        exit 1
+    fi
+    cp "${f}" "${STAGE}/manual/media/"
+done
+echo "==> staged the manual and $(ls "${STAGE}/manual/media" | wc -l) picture(s)"
+
 # One installer in the output directory, the one just built. The directory
 # is cached between CI runs, so after a version bump a glob over it would find
 # two and the smoke test would hand Wine both names as one path.

docs/README.md

@@ -7,8 +7,8 @@ second.
 
 | | |
 |---|---|
-| [The manual](manual/README.md) | Every feature, pictured from the application itself — opening a library, rating and filing, developing, local masks, repair, film, panoramas, export |
-| [How it is driven](gestures.md) | Every gesture and shortcut, by screen. Generated from the code, so it cannot describe one the application does not have |
+| [The manual](manual/README.md) | Every feature, pictured from the application itself — opening a library, rating, labelling and filing, duplicate originals, developing, local masks, repair, film, panoramas, export. The application carries it and opens it from Help and from Settings |
+| [How it is driven](gestures.md) | Every gesture and shortcut, by screen. Generated from the code, so it cannot describe one the application does not have. The same list is the in-app help sheet: `Help` or `F1` in the grid, `?` or `F1` in develop |
 
 The [top-level README](../README.md) says what DarkRoom is, how to get it on
 each platform, and what is still missing.
@@ -74,11 +74,13 @@ recollection.
 
 ## Conventions
 
-Two files here are generated and must not be edited by hand:
-`gestures.md` and `dev/traceability.md`. Both come from
-`cargo run -p traceability` and the pre-commit hook keeps them in step with
-the tree. The manual's pictures are recorded by
-[`tools/manual`](../tools/manual/README.md) and live in LFS.
+Three files here are generated and must not be edited by hand:
+`gestures.md`, `dev/traceability.md` and `manual/index.html`, the page the
+packages install, rendered from `manual/README.md`. All three come from
+`cargo run -p traceability`, the pre-commit hook keeps them in step with the
+tree, and CI fails when one is not what the tree generates. The manual's
+pictures are recorded by [`tools/manual`](../tools/manual/README.md) and live
+in LFS.
 
 A design document links to the requirements it satisfies and to the code
 that satisfies them. When the code moves, the link moves with it; a document

docs/dev/architecture.md

@@ -11,9 +11,10 @@ and records the decisions and constraints behind the design.
 ## 1. Overview
 
 DarkRoom is a Rust application with a Slint interface. On Linux it renders through wgpu to Vulkan.
-On Android it renders through Skia to OpenGL — not by preference but because wgpu's Vulkan
-swapchain cannot pre-rotate, which tears a portrait window on a landscape-mounted panel
-([technical-debt.md TD-1](technical-debt.md)). The compute passes are wgpu on both. The design is organised around four ideas, each of which the rest of this
+On Android it renders through Skia, also on wgpu's Vulkan swapchain, which a patched wgpu-hal and
+Skia renderer pre-rotate for a landscape-mounted panel ([technical-debt.md TD-1](technical-debt.md),
+[third_party/](../../third_party/README.md)). The compute passes are wgpu on both, on the same
+device the compositor draws with. The design is organised around four ideas, each of which the rest of this
 document elaborates:
 
 1. **Pixels stay on the GPU.** From decode to display, image data never round-trips through the
@@ -51,7 +52,7 @@ darkroom/
 │   ├── dr-types        SourceRef, ImageId, VersionId, ParamValue — shared vocabulary
 │   ├── dr-catalog      SQLite index, scan, query, metadata
 │   ├── dr-sidecar      the authoritative edit store (§6.12)
-│   ├── dr-decode       RawDecoder trait, rawler impl, embedded-preview extraction
+│   ├── dr-decode       Decoder trait (§3.2), rawler impl, embedded-preview extraction
 │   ├── dr-pipeline     Operation trait, descriptors, edit graph, registry
 │   ├── dr-gpu          wgpu device, tile scheduler, WGSL shaders, mask rasteriser
 │   ├── dr-colour       lcms2 bindings, camera profiles, working-space transforms
@@ -131,6 +132,15 @@ Four separate entry points because the caller's needs differ sharply by phase. C
 `RawImage` carries CFA-pattern sensor data plus black/white levels and camera colour matrices — it
 is *not* demosaiced. Demosaic is a GPU pipeline stage (§5.2).
 
+> **As built (FR-RAW-2, 0.15.0).** The trait is `dr_decode::Decoder`, and it takes bytes rather
+> than a reader: `header_bytes` says how much of a file metadata needs, `metadata` and
+> `orientation` read it, `locate_preview` says where the embedded preview sits so the caller's
+> storage can fetch that range, and `preview` and `decode` take the bytes fetched. The split this
+> section argues for survives; what moved is who reads the file, which is the storage layer
+> (§3.1's `read_range`), not the decoder. `dr_decode::Rawler` is the one implementation, and only
+> the places that start a job name `dr_decode::default()`; everything below them takes a
+> `&dyn Decoder`.
+
 ### 3.3 Operation and descriptors
 
 The develop pipeline is a sequence of operations with a uniform interface. Polymorphism is by
@@ -1125,12 +1135,11 @@ At 4K the shader finishes in 0.28 ms and then 7.15 ms is spent moving pixels thr
 26× overhead that scales with area, which is why an uncapped window resize falls off a cliff. The
 constraint is not a stylistic preference; it is the dominant cost in the frame.
 
-**One exception, on Android only, and it is debt rather than a revision.** The develop view there
-reads the frame back rather than handing over a texture, because zero-copy requires Slint to draw
-with wgpu and wgpu's Android swapchain tears a portrait window. The reasoning, the measurements
-that forced it and what would remove it are in [technical-debt.md TD-1](technical-debt.md). The
-constraint above still governs every other path, including the desktop develop view and the export
-pipeline, and the Android exception is expected to be temporary.
+**No exceptions since 0.15.0.** Android read the develop frame back until then, because wgpu's
+Vulkan swapchain could not pre-rotate and a portrait window tore on the tablet. Two local patches
+removed the need ([technical-debt.md TD-1](technical-debt.md)), so the frame reaches the compositor
+as a texture on both platforms. The constraint governs every display path. The export pipeline
+reads pixels back by design, because a file is its output.
 
 ### 6.2 Tiling from day one
 

docs/dev/catalog.md

@@ -249,6 +249,13 @@ what you actually have.
 only when something needs it: import duplicate detection (FR-CAT-11), or reconnecting a moved source
 (FR-CAT-9). Never during a routine scan.
 
+Consolidating the duplicates a library already holds (FR-CAT-11a, 0.16.0) does not wait for it. It
+proves a group the same by `content_hash` where every copy has one, and otherwise by a digest of
+each copy's first and last megabyte, read by range through the backend and kept in `dedup_probes`
+keyed on the size and mtime it was taken at, so a second review reads nothing.
+`dedup_probes` is created on first use rather than by a migration, because a schema bump would make
+an older build refuse this catalog's snapshot at sync (`core/dr-catalog/src/duplicates.rs`).
+
 ---
 
 ## 4. The library view
@@ -711,8 +718,9 @@ because if the SQLite path proves troublesome, this is the fallback with a known
   membership needs to be *stable* — for manual ordering, or for a pinned set that must not shift
   under the user — it needs materialising with an invalidation rule. Deferred until there is a
   concrete need.
-- **Multi-root capture-time collisions.** FR-CAT-11 detects duplicates on import; the same image
-  catalogued under two roots is a related but distinct case, not yet specified.
+- **Multi-root capture-time collisions.** FR-CAT-11 detects duplicates on import, and FR-CAT-11a
+  consolidates the copies one root holds in several folders; the same image catalogued under two
+  roots is a related but distinct case, not yet specified.
 - **Timeline granularity selection.** Which bucket size the UI picks for a given zoom is a UI
   concern, but the catalog should probably suggest one from the query's date span rather than have
   the UI guess.

docs/dev/display-and-extension.md

@@ -22,7 +22,7 @@ and the reason is that some of the work is done and untagged.
 | FR-DSP-1 proxy rendering | **Done.** The develop view renders at viewport resolution, not source. |
 | FR-DSP-2 tiled computation | **Absent, and §2 now says it should stay that way.** Measured: the fused pass is inside the budget everywhere. See [frame-budget.md](frame-budget.md). |
 | FR-DSP-3 interactive latency | **Measured and asserted** for the fused path — `core/dr-gpu/tests/frame_budget.rs`. Missed by one operation, clarity, for the reason recorded as TD-4. |
-| FR-DSP-4 progressive refinement | **Absent**, and §4's condition did not fire. Every render is full quality and can afford to be. |
+| FR-DSP-4 progressive refinement | **Built** (0.15.0), although §4's condition did not fire on the fused path: a half-resolution draft while a gesture moves, one sharp frame 120 ms after it stops, the histogram dimmed while it lags, and the draft faded out over 150 ms — `ui/dr-ui/src/refine.rs`. See [frame-budget.md](frame-budget.md). |
 | FR-DSP-5 zoom and pan | **Done and tagged**, against tests that fail if the behaviour is removed — `core/dr-gpu/tests/zoom_resolution.rs`. `Framing::view` shrinks the sampled region while the render target keeps its size, so zooming *raises* the resolution the pipeline works at. That is FR-DSP-5's requirement, arrived at without tiles. |
 | FR-DSP-6 colour management | **Done.** Output space is a parameter of composition. |
 | FR-DSP-7 histogram and clipping | **Done**, GPU-side, no per-frame readback. |

docs/dev/distribution.md

@@ -45,10 +45,11 @@ somewhere:
   `project_license` is GPL-3.0-or-later; those differ on purpose — see the
   comment in the file.
 - **Vulkan is a requirement, not a preference.** The develop pipeline is
-  compute shaders through wgpu, and NFR-R8 — how far a CPU fallback goes — is
-  still open, so today there is nothing behind it. A package that installs onto
-  a machine with no working ICD produces an application that starts and cannot
-  develop.
+  compute shaders through wgpu, and NFR-R8 was decided on 2026-09-19 against a
+  CPU pipeline: without a device the library, the grid and the judgements work
+  on embedded previews, and develop and export say they are unavailable. A
+  package that installs onto a machine with no working ICD produces an
+  application that starts and cannot develop.
 - **A Secret Service implementation, or an honest degraded mode.** FR-NC-2 is
   explicit that the absence of a secrets daemon is a stated degraded mode and
   never a silent fall back to plaintext. Packages express this as an optional
@@ -59,6 +60,12 @@ somewhere:
   the Flatpak manifest check the file size and refuse, because the alternative
   is a package whose face indexing fails inside the graph loader on a user's
   machine rather than on the packager's.
+- **So are the manual's pictures.** Since 0.15.0 the Arch package, the Windows
+  installer and the APK carry `docs/manual/index.html` and its `media/`, where
+  `dr_ui::manual` looks for them (`/usr/share/darkroom/manual`, `manual\` beside
+  `darkroom.exe`, the APK's `assets/manual`), and each refuses a pointer where a
+  picture should be — shipped, a pointer is a manual of broken images that
+  nothing reports. The Flatpak manifest does not install it yet.
 
 ---
 

docs/dev/frame-budget.md

@@ -264,6 +264,14 @@ base, computed at reduced resolution and correct at any moment the user stops.
 "Render coarse while dragging, sharpen when it settles" would paper over the same
 34 ms with a visible swap. Fix the stage.
 
+> **Since, 0.15.0.** The stage was fixed (§ The reduced base, below), and the
+> draft was built anyway, in the form this section would accept: the develop
+> view renders at half resolution while a gesture moves and once at full
+> resolution 120 ms after it stops (`ui/dr-ui/src/refine.rs`), the histogram
+> dims while it describes an older frame, and the last draft fades out over
+> 150 ms rather than being swapped. It is not a mask over a slow stage; it
+> spares a drag the full-resolution frames it does not need.
+
 ---
 
 ## Which GPU, on a machine with more than one
@@ -431,3 +439,38 @@ difference between a quarter-scale and a half-scale base to 0.03 stops of peak
 excursion and 2% of frame reach. But it is a change in reach rather than only
 in cost, it is what a tile scheduler would be handed, and it is worth knowing
 that the number moved rather than discovering it later as a seam.
+
+---
+
+## The fit view, again — 2026-09-25
+
+**Status:** Measured in the commits named, not re-run for this file.
+
+Two changes to the fused path made the `fit` rows above cheaper again,
+each with its before and after in its commit message. Both were measured
+on the laptop RTX 3050 with its clocks held at 420/810 MHz by the power
+cap, on the synthetic 60 MP source of `examples/frame_budget.rs`, median of
+five alternated runs; both leave the rgba8 output bit-identical.
+
+**The source gather is read once per framing** (`1dc7b45`). At fit every
+output pixel reads one texel on a stride through a source three or four
+times its width, and that gather was most of the fused pass. The pass now
+keeps a render-sized `rgba16float` cache of it, keyed on the framing, and
+reads it back while only the adjustments move:
+
+| scene | before | after |
+|---|---:|---:|
+| neutral, 2560 × 1600 fit | 10.62 ms | 3.88 ms |
+| neutral, 3840 × 2160 fit | 21.05 ms | 7.11 ms |
+| clarity, 3840 × 2160 fit | 42.20 ms | 27.88 ms |
+| neutral, 2560 × 1600 1:1 (control) | 3.83 ms | 3.84 ms |
+
+An interpolated read (straightening, lens warps, CA) is not cached, and the
+cache is written on the second frame with a given key, so a crop or zoom
+drag pays nothing for it.
+
+**A detail pass that changes nothing is dropped** (`d430ec9`). Capture
+sharpening at a scale too coarse to draw its radius emits an empty pass,
+which cost a full read and write when another neighbourhood operation
+followed it: sharpen with clarity at 2560 × 1600 fit went from 18.66 ms to
+14.16 ms, and at 3840 × 2160 from 37.93 ms to 27.88 ms.

docs/dev/mask-editing.md

@@ -269,7 +269,7 @@ The set operations are already expressible in fixed-function blending over
 |------|-------------------------------|--------|-------|
 | Union | `One`, `One`, `Max` | `max(dst, src)` | M1 |
 | Subtract | `Zero`, `OneMinusSrc`, `Add` | `dst · (1 − src)` | M1 |
-| Intersect | `Zero`, `Src`, `Add` | `dst · src` | M2 |
+| Intersect | `Zero`, `Src`, `Add` | `dst · src` | M2 (built 2026-09-24) |
 
 The middle row is `brush_erase`, already constructed in
 [`MaskPass::new`](../../core/dr-gpu/src/mask.rs). The other two are the same
@@ -292,6 +292,16 @@ rendering as it did. `an_erase_stroke_holes_its_own_part_and_not_the_mask` in
 [`local_adjustments.rs`](../../core/dr-gpu/tests/local_adjustments.rs) is the test
 that holds this in place.
 
+Intersection, as built, is exactly the table's row: a third pipeline beside
+the other two in `MaskPass::new`, `Join::apply` stating the three on the CPU,
+and `a_part_unioned_subtracted_and_intersected_gives_the_three_fields` and
+`the_joins_match_their_definition` in `local_adjustments.rs` holding the GPU
+to it. It is a product, not a minimum: the two agree wherever either side is
+fully in or out, and between two soft edges the product is the softer reading,
+which is the right way for an overlap of two partial selections to be wrong.
+Its blend is `Add`, not `Max`, so it adds nothing to the Android question
+below.
+
 `Max` blending on `r8unorm` is core WGPU and universally supported on the
 desktop backends; **verify it on the Android adapter before M2 lands**, since
 that is the platform where a blend mode is most likely to be quietly emulated
@@ -689,6 +699,14 @@ nothing, and every report of it came back as "the masks do not work".
 per-part distance fields (§5.5), the part list with its join chips, folding
 two layers.
 
+Done (2026-09-24): `Intersect` — the join, its blend state, the sidecar word
+`intersect`, the part row's chip cycling + / − / ∩, and an "∩ Intersect"
+button beside Add and Subtract (§7.1). The part list with its chips and a
+per-part eye was already there from M1. Outstanding: joining a model,
+gradient or range part from the panel (the buttons still join a painted
+part, so an intersection is painted where the mask should survive), per-part
+distance fields, and folding two layers.
+
 **M3 — Push, and cling.** The warp pass (§5.3) and edge-aware deposit (§5.6).
 Both are refinements of a tool that already works, which is the right place
 for the two riskiest pieces.

docs/dev/outstanding.md

@@ -28,6 +28,12 @@ Android memory pressure and lost-root recovery (FR-PLAT-AND-5, FR-PLAT-AND-2), i
 recorded where it appears rather than deleted, because a requirement that is *half* met is the
 one most likely to be reported as closed.
 
+**Swept again on 2026-09-26, for 0.16.0.** Progressive refinement (FR-DSP-4) and the panorama
+(§3.11) were built and are struck below; consolidating duplicate originals (FR-CAT-11a) now sits
+beside re-import detection; the accessibility and localisation counts in §6 had not been read
+against the tree since 2026-08-30 and are replaced; and §4a records what the develop and keyboard
+work of 0.15.0 and 0.16.0 left open.
+
 ---
 
 ## 1. Plugins — post-v1 since 2026-09-19
@@ -80,8 +86,8 @@ somebody reads the matrix.
 
 ## 2. Culling — the stated differentiator, half built
 
-[D11](requirements.md) names culling "the core differentiator". FR-CULL-1, -2, -3, -4 and -8 through
--12 are built. Three are not.
+[D11](requirements.md) names culling "the core differentiator". FR-CULL-1 through -5 and -8 through
+-13 are built. Two are not.
 
 **FR-CULL-3 — Raw-truth overlays. Built, all three bullets.** Focus peaking is
 `core/dr-gpu/src/focus.rs` and `ui/dr-ui/src/peaking.rs`; the raw histogram and the raw clipping
@@ -115,8 +121,13 @@ labels — that comment was a forward reference and is now simply wrong, rather
 And `dr_catalog::bursts::choose_representative` is written and tested but bound to no gesture, so
 today the only override is expanding the burst.
 
-`core/dr-catalog/src/dedup.rs` remains a different thing: re-import detection under FR-CAT-11,
-matching a file against one already catalogued, not two photographs against each other.
+Neither is deduplication, which is about one file held twice rather than two frames that look
+alike. `core/dr-catalog/src/dedup.rs` is re-import detection under FR-CAT-11, matching a file on
+a card against one already catalogued. Its other half, FR-CAT-11a, is built since 0.16.0:
+`core/dr-catalog/src/duplicates.rs` groups the copies a library already holds (same root, camera,
+capture instant and size), `ui/dr-ui/src/duplicates.rs` proves each group the same by digest and
+compares their edits, and a group is folded onto one survivor with the others trashed in one
+transaction, from the Duplicate originals review in the sidebar and in Settings.
 
 **FR-CULL-6 — Compare and survey.** Absent. No side-by-side view, no synchronised zoom or pan.
 This is the one of the four with no adjacent machinery at all, and it is also the one that most
@@ -166,10 +177,15 @@ Two measurements say it costs more than it saves on the interactive path. Neithe
 about the export path or about a device under memory pressure, which is where the case for it
 actually lives — and that is spike S6, which has not run.
 
-**FR-DSP-4 — Progressive refinement.** Unbuilt. FR-DSP-1's proxy rendering and TD-4's
-quarter-resolution base are adjacent and are not it: both are fixed choices about what resolution to
-compute at, where FR-DSP-4 asks for a first frame that is deliberately cheap and a second that
-replaces it. Nothing tracks a "this frame is provisional" state.
+**FR-DSP-4 — Progressive refinement. Built in 0.15.0.** While a gesture moves the canvas renders
+a half-resolution draft, and the sharp frame lands once, 120 ms after the last movement: the
+decision is `ui/dr-ui/src/refine.rs`, a debounce whose every draft re-arms the settle timer, driven
+through simulated timelines in its tests. The draft flag reaches the interface (`canvas-draft`,
+`Levels.provisional`), so the histogram dims while it describes a frame older than the one on
+screen, and the last draft is kept and faded out over 150 ms when the sharp frame arrives, so
+refinement is not a swap. Draft frames themselves date from 2026-08-09; what this entry missed,
+and 0.15.0 finished, was a settle that waited for the gesture to stop, a provisional state the
+interface could see, and a refinement that was not a jarring swap.
 
 **NFR-RES-2 — Images larger than GPU memory.** Half answered. NFR-R8's "decide explicitly" was
 decided on 2026-09-19: there is no CPU render pipeline, the degraded mode is the viewer on
@@ -181,11 +197,42 @@ settle both this and FR-DSP-2, and there is no evidence it has run.
 
 ---
 
+## 4a. Develop, masks and the keyboard — what the 0.15.0 and 0.16.0 work left open
+
+Most of what these two releases built closed a clause outright, and is listed here only so that
+nobody looks for it below: perspective correction (FR-DEV-20, Vertical and Horizontal in Compose),
+the crop that says when it orphans a mask (FR-DEV-17), intersection as a third mask join
+(FR-DEV-19a), the keyboard vocabulary for develop (FR-DEV-16, met and held by `gestures-check` in
+both directions), colour labels set, shown and filtered (NFR-A11Y-3's first example), the RAW
+decoder behind a trait (FR-RAW-2), and duplicate originals (FR-CAT-11a, §2 above). What they left:
+
+- **FR-UI-5's pointer half.** Keys cover navigation, rating and common adjustments in every view,
+  and judging works in develop on the open photograph; but no develop slider takes the scroll
+  wheel, which the clause names. Its status note in the register says so rather than rounding up.
+- **FR-RAW-2's second decoder.** The trait is built and a stub decoder is proved to reach the
+  scan, the preview ladder and export without a caller changing; LibRaw itself (D2) is not
+  built, and S7 is what would say when it is needed.
+- **FR-DEV-19's M2 remainder** ([mask-editing.md §11](mask-editing.md)). The panel's Add, Subtract
+  and Intersect buttons join a painted part only, so an intersection with a gradient or a range is
+  painted where the mask should survive; per-part distance fields and folding two layers are
+  unbuilt. From M1, the brush has no ring drawn on the photograph and a layer's whole stroke
+  history is redrawn per dab.
+- **FR-DEV-17's blind spot, by design.** A layer that adds a range or a region selection is never
+  reported, because the check has no label map to measure it with (`dr_pipeline::orphan`); a
+  false alarm on the common path would teach the notice to be dismissed unread.
+
+The desktop's scrollbars (the develop column, the grid, the sidebar, Settings, the film list and
+the help sheet) are drawn only where `dr_plat::is_touch_first()` is false; on Android the lists
+still scroll by flick alone, which is deliberate rather than outstanding.
+
+---
+
 ## 5. Android beyond running, and Flatpak
 
 The Android app is not a stub — it builds an APK, runs the whole application, unpacks bundled face
-models, and has been measured on a tablet ([faces.md §12.1](faces.md),
-[technical-debt.md TD-1](technical-debt.md)). What is missing is the platform contract around it.
+models, has been measured on a tablet ([faces.md §12.1](faces.md)), and since 0.15.0 draws the
+develop view zero-copy as the desktop does ([technical-debt.md TD-1](technical-debt.md), paid off).
+It carries the manual and opens it in a WebView. What is missing is the platform contract around it.
 
 **FR-PLAT-AND-1 is untagged, and what it was tagged for was intent rather than code.**
 The requirement demands that library access be obtained *exclusively* through the Storage Access
@@ -267,38 +314,44 @@ fixes users will need — is unaddressed, and there is no update mechanism of an
 
 ## 6. Accessibility and internationalisation — the hard half is done and the easy half is not
 
-**NFR-A11Y-1 — Localisation.** `@tr(` appears **zero** times across 14,482 lines of Slint. That
-number overstates the problem, because the part that is genuinely architectural was got right:
-`LocalizedKey` keeps display strings out of `core/` entirely, every operation publishes a key rather
-than a label, and `labels::resolve` is the single point where a key becomes text. What that single
-point does, however, is a hardcoded English `match` in Rust source — so changing a translation
-requires a recompile, which is the one thing the requirement explicitly forbids. There is no message
-catalogue in any format, no locale-resolution rule, and no decision recorded about RTL.
+**NFR-A11Y-1 — Localisation.** One screen of twenty-eight Slint files is converted: the launch
+screen wraps its strings in `@tr(`, and `ui/dr-ui/build.rs` records the mechanism — Slint's bundled
+translations, a `.po` per language under `ui/dr-ui/lang/`, extracted with `slint-tr-extractor` —
+and why bundling rather than gettext (Android has no path a `.mo` could sit at). No `.po` exists
+yet, so every build is the original English, and the rest of the interface (about 23,600 lines of
+markup) is unwrapped. The part that is genuinely architectural was got right: `LocalizedKey` keeps
+display strings out of `core/` entirely, every operation publishes a key rather than a label, and
+`labels.rs` is the single point where a key becomes text — but that point is a hardcoded English
+`match`, and `build.rs` says the Rust-side mechanism is not built. Two things the requirement asks
+remain against it even when the conversion is finished: a bundled translation is compiled in, so
+changing one still needs a rebuild, which the clause forbids; and there is no locale-resolution rule
+and no decision recorded about RTL.
 
-The work left is therefore smaller than it looks and entirely mechanical: a catalogue format, a load
-path behind `resolve`, and `@tr(` around the Slint literals. The design it needs already exists.
+**NFR-A11Y-2 — Accessibility. Tagged, and half tested.** `accessible-*` properties now appear on
+some 127 lines across eleven Slint files — the shared controls in `widgets.slint` and
+`controls.slint`, every slider, the rail, the grid's cells (named by file), the rating stars, the
+labels and the duplicates review — and `ui/dr-ui/tests/ui_controls_are_accessible.rs` fails when a
+shared control stops declaring its role or a slider is added without a name. The manual's recording
+scripts find every control they press through the same names (`dr_ui::automation`), so a control a
+screen reader cannot name is also one the manual cannot record. What is unchecked is stated in the
+test: whether the words are good, and whether Android exposes any of it — spike S13, which has not
+run. The requirement's other two clauses are debt with entries of their own: platform font scaling
+([TD-7](technical-debt.md)) and WCAG AA contrast ([TD-6](technical-debt.md)).
 
-**NFR-A11Y-2 — Accessibility.** `accessible-*` appears five times in the whole interface, all five
-on one control — the parameter slider in `adjust.slint` — and nothing is set from the Rust side at
-all. Everything else in eighteen Slint files is unnamed to AT-SPI and TalkBack. The requirement's own
-caveat, that Slint's Android accessibility needs verifying, is spike S13, which has not run.
+**NFR-A11Y-3 — Colour-independent status.** Built, and now asserted. The clipping readout pairs a
+marker that appears or disappears with a figure in words; the rating strip is a solid star against
+an outline in an achromatic palette; the pick/reject mark is a tick against a cross; the
+focus-peaking colour chips say "Red" and "Cyan" rather than showing swatches; and colour labels —
+the requirement's first named example — now have an interface built with a shape from the start:
+every mark carries its label's initial (R, Y, G, B, P) on its colour, and every chip, picker and the
+develop bar name the label in words.
 
-**NFR-A11Y-3 — Colour-independent status.** Built where a control exists, and now tagged: the
-clipping readout pairs a marker that appears or disappears with a figure in words, the rating strip
-is a solid star against an outline in an achromatic palette, the pick/reject mark is a tick against
-a cross, and the focus-peaking colour chips say "Red" and "Cyan" rather than showing swatches. Each
-of those already carried the reasoning in a comment naming this requirement and simply had no
-`TRACES` line.
-
-Two caveats, because the tag now says more than the evidence does. **Only the clipping clause has a
-test** — `a_clipping_figure_distinguishes_none_from_nearly_none`, which pins `<0.1%` apart from `0%`
-so the figure cannot contradict the lit marker beside it. The three Slint components are
-inspected-and-argued, not asserted, and nothing would fail if a future edit made a star differ only
-in tint. **And the requirement's first named example has no interface at all**: catalog colour
-labels are a nullable `label INTEGER` column on the versions table and are set and shown nowhere, so
-the clause about them is untestable rather than satisfied. That clause closes when the label UI is
-built, not before, and it should be built with a shape from the outset — which is the same argument
-as below, for doing this alongside NFR-A11Y-2 rather than after it.
+The clipping clause is pinned by `a_clipping_figure_distinguishes_none_from_nearly_none`. The other
+four are pinned by `ui/dr-ui/tests/status_is_not_colour_alone.rs`, which reads the markup and fails
+if a star, a flag or a label comes to differ in tint alone — if the glyph a state selects stops
+depending on the state, if two states select the same drawing, if two labels share a letter, or if
+the peaking chips stop being words. It is a structural check, not a perceptual one: it cannot say
+whether the letters are legible at 14px on a given screen, which is still a matter of looking.
 
 ---
 
@@ -455,18 +508,23 @@ whether something *should* be built — which is the opposite of the order §9 a
 
 ---
 
-## 11. Merging — specified 2026-09-19, nothing built
+## 11. Merging — the panorama built, three clauses open
 
 §3.11 was written on 2026-09-19 under D18, undeferring the panorama from §7 and leaving HDR merge
-and focus stacking there with their data model decided. Eleven `FR-MRG` clauses and two `NFR-MRG`
-figures entered the register at once with no code behind any of them, which is why the coverage
-figure fell from 83.0% to 77.2% on the same day — a specification, not a regression.
+and focus stacking there with their data model decided. Its clauses entered the register with no
+code behind them, which is why the coverage figure fell from 83.0% to 77.2% on that day — and the
+panorama was then built in the same week: alignment, projections, the chunked composite written as
+a DNG beside its sources, the auto-crop and the model's border fill.
+[panorama.md](panorama.md) §11 and §13 are where it stands.
 
-[panorama.md](panorama.md) is the design, and its §10 is the order of work. Nothing starts before
-**S15**: whether rawler reads back a linear DNG the application writes, whether XFeat loads under
-tract at a fixed shape, whether the working-space texture can be tapped where FR-MRG-2 needs it,
-and what a chunked blend of a 100 MP composite costs on the tablet. The first two are a day each
-and either can change the design, which is the reason they come first.
+Three clauses carry no tag:
+
+- **FR-MRG-9 — the tablet.** Android runs the same code, but the stated ceiling on frame count and
+  source resolution, refused with a message rather than an out-of-memory kill, is not set.
+- **NFR-MRG-1 — merge latency.** Measured on the desktop and inside its 60 s (panorama.md §11);
+  the tablet's figure is S15.4's open half, and nothing asserts either per commit.
+- **NFR-MRG-2 — reproducible merges.** Nothing checks that the same sources and settings give a
+  byte-identical composite.
 
 ## 12. D12, which governed all of the above
 

docs/dev/requirements.md

@@ -179,6 +179,17 @@ card into the library. Both are needed.
 original filename) and by content hash, offering skip or import-as-new. Camera filenames wrap at
 `IMG_9999`, so filename alone is insufficient. Existing catalog duplicates are detectable on demand.
 
+**FR-CAT-11a — Consolidating catalog duplicates.** Files already in the library more than once
+(same root, camera, capture time and size) shall be listed on demand as groups, from the library
+and from Settings. Before anything moves, each group shall be proved the same file by a stored full
+digest or by a digest of the first and last megabyte of every copy, and its copies' develop edits
+compared; a group that differs in either is left out and the review says why. One copy survives —
+not under a backup-looking folder, then camera-named, then oldest, overridable per group — and the
+others' collections, keywords, highest rating, agreed flag and label, and faces are merged onto it,
+with disagreements reported rather than decided silently. Each group is merged and its other copies
+moved to the trash (FR-CAT-15) in one transaction: wholly consolidated or untouched. Nothing is
+deleted.
+
 **FR-CAT-12 — Versions (virtual copies).** An image may carry multiple named `Version`s, each with
 an independent edit graph, without duplicating source data. Versions are creatable, nameable,
 deletable, and independently exportable; one is the default.
@@ -243,9 +254,15 @@ read metadata, and `import.rs` fetches exactly that range through `Storage::read
 calling `dr_decode::metadata`. The decoder states its requirement and the storage layer satisfies
 it; a decoder holding its own `SourceRef` would have had to implement the range policy itself.
 
-What is genuinely not built is the trait. There is one decoder, reached through free functions, so
-"without changing callers" is a claim nothing yet tests. The clause stands as written and is
-outstanding work, not a satisfied one.
+*Status (2026-09-24).* The trait is built: `dr_decode::Decoder`, over bytes — `header_bytes`,
+`metadata`, `orientation`, `locate_preview`, `preview` and `decode` — with `dr_decode::Rawler` as
+its one implementation, delegating to the free functions that were there before. The catalog scan,
+the thumbnail ladder, import, the viewer, export, merge and repairs take a `&dyn Decoder`; only the
+places that start a job name `dr_decode::default()`. "Without changing callers" is tested by
+`dr-ui`'s `decoder_seam` tests, which hand a stub decoder for a container no real decoder reads to
+the scan, the ladder and export, and fail if any of them reaches past the trait. Nothing in the
+trait takes a path or a `SourceRef`. The second decoder itself (LibRaw, D2) is not built; S7 (#48)
+is what would say when it is needed.
 
 **FR-RAW-3 — Sensor data handling.** Correctly apply per-camera black/white levels, CFA pattern
 identification, and camera-native colour matrices. Demosaic quality shall be selectable, with at
@@ -544,9 +561,17 @@ decade; a mask that cannot be corrected is the reason an edit leaves this applic
 editor.
 
 **FR-DEV-19a — Mask composition.** A layer's mask is an ordered list of parts, each naming a source
-and how it joins the mask before it — added to it, or taken out of it. A layer of one part is
-exactly the layer that existed before this, and reads and writes the same sidecar. The parts of a
-layer merge under FR-NC-9 as the layer does, and each carries its own edge treatment.
+and how it joins the mask before it — added to it, taken out of it, or intersected with it, keeping
+only where both agree. A layer of one part is exactly the layer that existed before this, and reads
+and writes the same sidecar. The parts of a layer merge under FR-NC-9 as the layer does, and each
+carries its own edge treatment.
+
+*Intersection added 2026-09-24* (issue #9, `mask-editing.md` M2). Union and subtraction were built
+first; the selections that most need composing — the sky that is also bright, the subject that is
+also skin — are intersections, and spelling one as a subtraction needs a part that selects the
+complement. Intersection is the product of the two coverages, which is the minimum wherever either
+side is fully in or out. A sidecar written before it never names it; a build from before it reads
+the word as a union, which keeps the part visible rather than dropping it.
 
 **FR-DEV-19b — Hand correction.** A part may be painted, with add and erase strokes, at a radius,
 hardness and flow the photographer sets. Strokes are stored as normalised source coordinates and
@@ -569,6 +594,17 @@ which knows nothing of a layer's shaping and nothing at all about a gradient, a
 so choosing a subject or a category produced a layer whose extent was invisible, and every control
 in FR-DEV-19a and FR-DEV-19b acted on something the photographer could not see.
 
+**FR-DEV-17 — A crop that orphans a mask says so.** When a crop is let go — the end of the drag,
+not any frame of it — and it has left a mask layer's coverage entirely or mostly outside the frame,
+the photographer is told how many layers, and which, and offered to keep the crop or take it back.
+The crop is applied either way and is never refused; taking it back is the ordinary undo, so the
+crop and its warning are one step in the history. A crop that strands nothing, and a layer already
+outside the frame before the crop moved, produce no notice at all.
+Mask geometry is stored in source coordinates, so a tighter crop does not destroy a layer: it makes
+it invisible, and nothing announces it. That silent loss is the whole case for cropping last; a
+notice at the moment it happens turns it into a decision, and lets the crop stay where the histogram
+argues it belongs — early.
+
 **FR-DEV-16 — A keyboard vocabulary for develop.** Every develop gesture that can be reached from
 the keyboard is bound, tagged beside its implementation, and generated into the gesture book
 (FR-UI-4): stepping through the folder, fit and 1:1, holding the original, undo and redo, copying
@@ -580,6 +616,28 @@ Editing rhythm depends on the hands staying put: reaching for a menu breaks the
 edit the same way it breaks the pace of a cull. A binding nobody can discover is the same as no
 binding, which is why the generated book is part of the requirement rather than documentation of it.
 
+*Status (2026-09-24).* Met. Every binding named above is bound in develop's key handler and tagged
+beside it, and develop also answers zoom in, out, fit and 1:1, panning a magnified view, rating,
+flagging and labelling the open photograph, nudging the control last moved (the framing sliders
+included), turning a mask part's join, keeping a crop that hid a mask, and going back to the grid.
+"Cannot describe a binding the application does not have" is now checked rather than argued:
+`traces gestures-check` reads the chord every handler compares against (`Keys.chord`, in
+`ui/dr-ui/ui/keys.slint`) and fails when a bound key has no tag in its place or a tagged key has no
+handler (`tools/traceability/src/keymap.rs`).
+
+**FR-DEV-20 — Perspective correction.** Framing shall include perspective correction — a vertical
+and a horizontal keystone — applied before the crop. It runs in the coordinate chain after the
+crop and the straightening and before the stored orientation and the lens warp, so that "vertical"
+means vertical in the photograph as it is shown and the lens is still corrected on the full frame
+it projected. It is part of framing and carries its `Compose` attribute, so a settings paste
+withholds it by default for the same reason it withholds the crop. The correction never exposes
+undefined area on its own; where it is combined with a straightening angle, the crop that avoids
+the empty corners (`max_inscribed_crop`) accounts for it, and the crop is refitted when the gesture
+ends.
+Converging verticals are the commonest geometric fault in architectural and interior work, and
+correcting them reframes the image — which is why the order relative to the crop is not a
+preference, and why the correction belongs to composition and not to the colour operations.
+
 ### 3.4 Display and interaction
 
 **FR-DSP-1 — Proxy-resolution rendering.** The develop view renders at the resolution actually
@@ -609,6 +667,12 @@ asynchronously, and the proxy result remains on screen until it is ready.
 quality or resolution, refining to full quality when interaction settles. Refinement is visually
 smooth, not a jarring swap.
 
+*Status (2026-09-26).* Met in 0.15.0. A gesture renders half-resolution drafts, and the sharp frame
+lands once, 120 ms after the last movement rather than on a timer counted from the first
+(`ui/dr-ui/src/refine.rs`). While a draft is up the histogram's display reading is dimmed, since it
+describes the last settled frame; the last draft is kept over the sharp one and faded out in
+150 ms, so the refinement is not a swap.
+
 **FR-DSP-5 — Zoom and pan.** Fit, 1:1, and arbitrary zoom levels. At 1:1 and above, the pipeline
 operates on the visible crop at full source resolution.
 
@@ -695,6 +759,15 @@ rating, and common adjustments. Neither is required for any operation to be reac
 > wherever they can be set, and on the roll's cells, so stepping along a set shows what has been
 > judged.
 
+*Status (2026-09-24).* The keyboard half and the amendment are met; the pointer half is not yet.
+Keys cover navigation, rating and common adjustments in the grid, develop, the collections sidebar,
+People and the export and copy sheets, each one in the gesture book and held to its handler by the
+gestures gate. Rating and flagging work in develop on the open photograph without advancing, with
+stars and Pick/Reject in the top bar, and the roll's cells carry each frame's flag and stars; pick
+and reject also gained a pointer and touch route in the grid (Flag in the selection bar), where they
+had been keyboard-only. Outstanding: scroll-wheel adjustment on numeric controls — the develop
+sliders do not take the wheel.
+
 **FR-UI-6 — Shared component library.** Touch and desktop presentations are variants of shared
 components, not parallel implementations. A new operation (FR-DEV-3c) becomes usable on both
 without frontend work.

docs/dev/storage.md

@@ -195,6 +195,7 @@ pub trait BackendProvider: Send + Sync {
     fn endpoint_placeholder(&self) -> &'static str;
     fn sign_in(&self) -> SignIn;
     fn normalise_endpoint(&self, input: &str) -> Result<String, String>;
+    fn upgrade_endpoint(&self, stored: &str) -> Option<String>;            // defaulted: None
     fn account_for(&self, endpoint: &str) -> Result<Account, RemoteError>;  // defaulted
     fn connect(&self, conn: &Connection) -> Result<Box<dyn RemoteBackend>, RemoteError>;
 }
@@ -215,6 +216,18 @@ pub enum SignIn {
   Nextcloud provider upgrades `http://` to `https://` here (`NFR-SEC-3`); the
   folder provider canonicalises the path, so two spellings of one directory do
   not become two accounts indexing the same photographs.
+- **`upgrade_endpoint` is for accounts already saved,** and is not a second
+  call to `normalise_endpoint`: the folder provider's needs the path to exist,
+  so running it on every launch would fail a library on an unplugged disk.
+  `AccountStore::upgrade_endpoints` runs it at launch; only Nextcloud answers,
+  rewriting an `http://` account an older build saved to `https://`
+  (#65). The keyring entry is filed under the endpoint, so the secret is
+  copied across before the record changes, and a move that would change
+  `namespace()` is refused rather than performed. Below both, the Nextcloud
+  client sets `https_only`, so no URL it sends — a typed one, the login flow's
+  poll endpoint, a redirect — can carry the app password in the clear, and
+  the login flow keeps the address the user typed rather than the server's
+  idea of its own URL.
 - **`connect` is synchronous and cheap.** It validates configuration and builds
   a client; it does not talk to the remote. Workers call it per task.
 - **`SignIn` is a shape, not a method.** It would be tidier to expose

docs/dev/technical-debt.md

@@ -14,7 +14,7 @@ has quietly stopped being necessary.
 
 ---
 
-## TD-1 — The Android develop view reads pixels back through the CPU
+## TD-1 — The Android develop view reads pixels back through the CPU ✅ PAID OFF
 
 **Breaks:** [architecture.md §12 / 6.1](architecture.md) — GPU results never round-trip through the
 CPU — and AC-8, on Android only. Desktop is unaffected and keeps the zero-copy path.
@@ -88,6 +88,41 @@ Any one of these removes it:
 `unstable-wgpu-29` to non-Android, the `#[cfg(target_os = "android")]` arm of
 `DevelopSession::render` is gone, and the tablet is clean in portrait.
 
+### Paid off
+
+By none of the three routes above. It came from a fourth that the list missed: patching both
+halves locally. "wgpu cannot do [the rotation] on Skia's behalf" was true and beside the point,
+because Skia can rotate its own canvas. Slint's Skia renderer already does it for rotated panels on
+linuxkms, just not on its wgpu surface. So two small patches, carried in `third_party/`
+([README](../../third_party/README.md)), do it for us:
+
+- **wgpu-hal 29.0.4.** `vulkan::Surface::set_pre_transform` lets a caller choose the swapchain's
+  `preTransform`, and `current_transform` reads the surface's. It is opt-in, and the default stays
+  `IDENTITY`, so desktop is unchanged.
+- **i-slint-renderer-skia 1.17.1.** On Android the wgpu surface sizes its swapchain in the panel's
+  orientation, promises the display's transform, and concatenates the matching rotation onto the
+  canvas before Slint draws. The imported develop texture goes through the same matrix as
+  everything else. The surface re-reads the transform every frame, because a half turn does not
+  resize the window. The item renderer's pixel snapping now accepts right-angle rotations,
+  or portrait would have lost it everywhere.
+
+With those in place, `unstable-wgpu-29` is common to both platforms, `shared_gpu` hands the one
+device to Slint on Android too, and both readbacks are gone (the frame through `export_pixels` and
+the focus overlay through `read_overlay`). The develop frame and its overlay reach the compositor
+as textures, as they do on desktop.
+
+**Verified by eye, not by instrument.** On 2026-09-25 the user checked the release-signed build on
+the tablet and found it clean. That is the portrait tear this entry was opened over. The
+`dumpsys SurfaceFlinger` readings that would complete the table above (composition and
+`bufferTransform` per orientation, expected `DEVICE` in all three) were not taken, because adb
+would not hold the device that morning. Nor was the develop frame time measured before or after.
+The readback's cost was never measured either, so the saving is reasoned, not measured.
+
+**The cost moved rather than vanished:** two upstream crates are pinned by path. A Slint or wgpu
+bump has to carry the patches forward (the README says how), and cargo only *warns* when a patch no
+longer matches, then silently builds the unpatched crate. Both patches should go upstream: the
+Skia half is the Android counterpart of a feature Slint already has, and the wgpu half is #3345.
+
 ---
 
 ## TD-2 — Thumbnails are fetched one at a time

docs/dev/ui-navigation.md

@@ -684,7 +684,8 @@ three columns of equal width side by side, each its own scroll:
    with them because it is not in their column.
 2. **Sliders** — `GroupStrip` above `AdjustPanel`, exactly as in the column.
    With a group selected this is one screen of sliders; with All it scrolls.
-3. **The mode's panels** — `ComposePanel` and `TransferPanel` in photo mode,
+3. **The mode's panels** — `ComposePanel` in photo mode (copy and paste
+   moved to the top bar, 2026-09-22),
    `SpotPanel` in repair, `MaskPanel` in local. Empty otherwise, which is
    a signal of its own about which mode the view is in (§1.1).
 

docs/dev/windows.md

@@ -285,6 +285,8 @@ $LOCALAPPDATA\Programs\DarkRoom\
     scrfd_500m_640.int8.onnx   scrfd_2.5g_640.int8.onnx   scrfd_10g_640.int8.onnx
     yolo26s-sem-ade20k.onnx   yolo26s-sem-ade20k.classes.json   categories.txt
     migan-512.onnx
+  manual\
+    index.html   media\   (the rendered manual and its pictures)
   LICENSE
   uninstall.exe
 ```

docs/gestures.md

@@ -5,7 +5,7 @@
 
 Every entry here is extracted from the comment beside the code that implements it, so this file cannot describe a gesture the application does not have. Add one by writing a `GESTURE:` block next to the implementation; there is nowhere else to write it.
 
-41 gestures, in 4 places.
+76 gestures, in 7 places.
 
 ## Develop
 
@@ -13,121 +13,254 @@ Every entry here is extracted from the comment beside the code that implements i
 
 - **Touch** — Press "pick" in the group's heading, then tap something neutral in the picture
 - **Pointer** — Press "pick", then click something neutral
+- **See it** — [in the manual](manual/README.md#white-balance-from-the-photograph)
 
 Sampling a neutral is the first move of the tonal pass — every colour judgement afterwards is measured against where the grey was put — and guessing at two sliders until a wall stops looking green is the wrong way round. One click is one sample and one step to undo; the sliders stay, because a sampled neutral is where the decision starts rather than where it ends.
 
-<sub>`ui/dr-ui/ui/adjust.slint:158`</sub>
+<sub>`ui/dr-ui/ui/adjust.slint:159`</sub>
+
+### Choose a film stock from the keyboard
+
+- **Touch** — Tap the Film row, flick the list, tap a stock
+- **Pointer** — Click the Film row, then a stock; the wheel or a drag scrolls the list
+- **Keyboard** — With the list open, `↑` and `↓` move along it, `Enter` chooses, and `Escape` or `Back` closes it unchanged
+- **See it** — [in the manual](manual/README.md#film)
+
+The list is longer than it is tall, so a way to walk it that cannot be lost to the column's scrolling is part of it being reachable at all. Focus goes back to the photograph's keys when it closes.
+
+<sub>`ui/dr-ui/ui/adjust.slint:1476`</sub>
 
 ### Magnify the photograph by any amount
 
 - **Touch** — Pinch it with two fingers
 - **Pointer** — The scroll wheel over it
+- **Keyboard** — `Ctrl+=` or `Ctrl+Plus` in, `Ctrl+-` out, about the middle of the view
+- **See it** — [in the manual](manual/README.md#looking-closer)
 
-Anchored on the fingers' midpoint, and on the pointer, so the gesture reads as magnifying the picture rather than sliding it about. Double-tap is the way to an exact 1:1; this is the way to everything in between.
+Anchored on the fingers' midpoint, and on the pointer, so the gesture reads as magnifying the picture rather than sliding it about. Double-tap is the way to an exact 1:1; this is the way to everything in between. Past 1:1 the pixels are shown as they are, square and unsmoothed; below it, filtered.
 
-<sub>`ui/dr-ui/ui/app.slint:1762`</sub>
+<sub>`ui/dr-ui/ui/app.slint:1908`</sub>
 
 ### Move a magnified photograph about
 
 - **Touch** — Drag it
 - **Pointer** — Drag it
+- **Keyboard** — `Shift+←`, `Shift+→`, `Shift+↑` and `Shift+↓`, a fifth of the view at a time
+- **See it** — [in the manual](manual/README.md#looking-closer)
 
 Only once there is something outside the viewport to reach, which is why the cursor becomes a hand exactly then. The view is clamped to the frame: panning past the edge would show undefined area beside the photograph, and that reads as a rendering fault rather than as the end of the picture.
 
-<sub>`ui/dr-ui/ui/app.slint:1853`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2004`</sub>
 
 ### Paint a mask by hand
 
 - **Touch** — Choose Paint or Erase, then drag on the photograph
 - **Pointer** — Choose Paint or Erase, then drag
+- **See it** — [in the manual](manual/README.md#local-adjustments)
 
 A model's mask stops inside a shoulder and leaks into the hair, and no single edge control fixes two errors that go opposite ways. The whole stroke is one step in the history, so taking a mark back costs one press however long it took to make.
 
-<sub>`ui/dr-ui/ui/app.slint:1940`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2095`</sub>
+
+### Open this list
+
+- **Touch** — Press "?" in the top bar, beside Settings, and Done to put it away
+- **Pointer** — Press "?" in the top bar, beside Settings, and Done to put it away
+- **Keyboard** — `F1`; the key that closes any sheet puts it away
+
+Most of the keys are develop's, and a reference that could only be opened from the grid had to be looked up before opening the photograph they were wanted for.
+
+<sub>`ui/dr-ui/ui/app.slint:2321`</sub>
 
 ### Take back the last change
 
 - **Touch** — Tap the step above the current one in the History list
 - **Pointer** — Click it, or press Undo in the History header
-- **Keyboard** — Ctrl+Z
+- **Keyboard** — `Ctrl+Z`
+- **See it** — [in the manual](manual/README.md#history-snapshots-presets)
 
 A whole drag is one step, so undo takes back a decision rather than a frame of a gesture. The list is there because arriving six steps back costs what arriving from one does.
 
-<sub>`ui/dr-ui/ui/app.slint:2160`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2351`</sub>
 
 ### Do it again after taking it back
 
 - **Touch** — Tap the step below the current one in the History list
 - **Pointer** — Click it, or press Redo in the History header
-- **Keyboard** — Ctrl+Shift+Z
+- **Keyboard** — `Ctrl+Shift+Z`, or `Ctrl+Y`
+- **See it** — [in the manual](manual/README.md#history-snapshots-presets)
 
-<sub>`ui/dr-ui/ui/app.slint:2173`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2365`</sub>
+
+### Remove a repair
+
+- **Touch** — Tap it, then Delete Repair
+- **Pointer** — Click it, then Delete Repair
+- **Keyboard** — `Delete` or `Backspace`, while repairing
+
+<sub>`ui/dr-ui/ui/app.slint:2385`</sub>
 
 ### Copy the settings from this photograph
 
-- **Touch** — Press Copy in the Settings panel
-- **Pointer** — Press Copy in the Settings panel
-- **Keyboard** — Ctrl+C
+- **Touch** — Press Copy in the top bar
+- **Pointer** — Press Copy in the top bar
+- **Keyboard** — `Ctrl+C`
+- **See it** — [in the manual](manual/README.md#copying-settings)
 
-The panel is the copy that has to work: a tablet has no modifier key to hold and no menu bar to hang the action from. The shortcut is an accelerator for a control that is on screen either way.
+The button is the copy that has to work: a tablet has no modifier key to hold and no menu bar to hang the action from. The shortcut is an accelerator for a control that is on screen either way.
 
-<sub>`ui/dr-ui/ui/app.slint:2206`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2404`</sub>
 
 ### Paste the settings onto this photograph
 
-- **Touch** — Press Paste in the Settings panel
-- **Pointer** — Press Paste in the Settings panel
-- **Keyboard** — Ctrl+V
+- **Touch** — Press Paste in the top bar
+- **Pointer** — Press Paste in the top bar
+- **Keyboard** — `Ctrl+V`
+- **See it** — [in the manual](manual/README.md#copying-settings)
 
 The button names what would be pasted — "3 adjustments", and whether the crop is coming with it — which the shortcut cannot say. Both paste the same scope.
 
-<sub>`ui/dr-ui/ui/app.slint:2218`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2417`</sub>
+
+### Choose which kinds of edit a copy carries
+
+- **Touch** — Open Presets and toggle the kinds
+- **Pointer** — Open Presets and toggle the kinds
+- **Keyboard** — `Ctrl+Shift+C`, which offers Copy beside them
+- **See it** — [in the manual](manual/README.md#copying-settings)
+
+Lightroom's Copy Settings. Pasting a look across a shoot usually means leaving each frame's crop and rotation alone, and that is a choice to make at the moment of copying.
+
+<sub>`ui/dr-ui/ui/app.slint:2435`</sub>
+
+### Export this photograph as the last one was
+
+- **Touch** — Press Export in the top bar
+- **Pointer** — Press Export in the top bar
+- **Keyboard** — `Ctrl+Shift+E`
+- **See it** — [in the manual](manual/README.md#export)
+
+Every export runs on the defaults in Settings, so "as the last one was" is what the button already does. The chord is Lightroom's and darktable's, kept so hands that learned it there need not learn it again.
+
+<sub>`ui/dr-ui/ui/app.slint:2460`</sub>
+
+### Choose how to export, then export
+
+- **Touch** — Open Settings, then Export defaults
+- **Pointer** — Open Settings, then Export defaults
+- **Keyboard** — `Ctrl+E`
+- **See it** — [in the manual](manual/README.md#export)
+
+The export sheet is the export defaults alone with an Export button. What is chosen there is kept, so it is also what the next Ctrl+Shift+E uses.
+
+<sub>`ui/dr-ui/ui/app.slint:2473`</sub>
+
+### Keep a crop that leaves a mask outside
+
+- **Touch** — Press "Keep crop" on the notice, or "Undo crop" to take it back
+- **Pointer** — Press "Keep crop" on the notice, or "Undo crop" to take it back
+- **Keyboard** — `Enter` keeps it; `Ctrl+Z` takes the crop back, like any other step
+
+<sub>`ui/dr-ui/ui/app.slint:2538`</sub>
+
+### Go back to the grid
+
+- **Touch** — Press "‹ Library" in the top bar
+- **Pointer** — Press "‹ Library" in the top bar
+- **Keyboard** — `G`
+
+Lightroom's key for the grid. Escape gets there too, but a step at a time — out of a mode, then out of a zoom — where this goes straight back.
+
+<sub>`ui/dr-ui/ui/app.slint:2555`</sub>
+
+### Nudge the control last moved
+
+- **Touch** — Drag its track
+- **Pointer** — Drag its track
+- **Keyboard** — `=` or `Plus` up and `-` down, a hundredth of its travel at a time; hold for more
+
+Lightroom's keys for the selected slider. There is no focus ring on a slider here, so "selected" is the last one moved — the same control `R` puts back — which covers the framing sliders, perspective included, as well as the adjustments.
+
+<sub>`ui/dr-ui/ui/app.slint:2584`</sub>
 
 ### Change which group of adjustments is on screen
 
 - **Touch** — Tap a group in the rail down the left
 - **Pointer** — Click a group in the strip above the develop column
-- **Keyboard** — [ and ] step through them, wrapping round through "everything"
+- **Keyboard** — `[` and `]` step through them, wrapping round through "everything"
+- **See it** — [in the manual](manual/README.md#developing-a-photograph)
 
 The groups are whatever the operation set declares itself to be about, so there are as many as the pipeline has and no key can be assigned to one of them by name. Stepping is the binding that survives a node being added.
 
-<sub>`ui/dr-ui/ui/app.slint:2246`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2612`</sub>
 
 ### Look at the photograph at 1:1
 
 - **Touch** — Double-tap the photograph
 - **Pointer** — Double-click it, or press the zoom readout floating over the canvas
-- **Keyboard** — Z
+- **Keyboard** — `Z` goes in and back out; `Ctrl+1` goes to 1:1 and `Ctrl+0` back to the whole frame
+- **See it** — [in the manual](manual/README.md#looking-closer)
 
-Noise reduction and capture sharpening are judgements about single pixels, and a fitted view averages several of the file's into each one on screen — so the frame looks softer than it is and the correction goes too far. The point and the magnification survive opening the next photograph, which is what makes checking the same eye across forty portraits forty keystrokes rather than forty pans.
+Noise reduction and capture sharpening are judgements about single pixels, and a fitted view averages several of the file's into each one on screen — so the frame looks softer than it is and the correction goes too far. The point and the magnification survive opening the next photograph, which is what makes checking the same eye across forty portraits forty keystrokes rather than forty pans. From 1:1 on the photograph is drawn as its own pixels, each a hard-edged square, rather than smoothed into a blur.
 
-<sub>`ui/dr-ui/ui/app.slint:2281`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2648`</sub>
+
+### Rate this photograph
+
+- **Touch** — Tap a star in the top bar
+- **Pointer** — Click a star in the top bar
+- **Keyboard** — `0`–`5`
+
+<sub>`ui/dr-ui/ui/app.slint:2705`</sub>
+
+### Pick or reject this photograph
+
+- **Touch** — Press Pick or Reject in the top bar; again to take the flag off
+- **Pointer** — Press Pick or Reject in the top bar; again to take the flag off
+- **Keyboard** — `P` picks, `X` rejects and `U` takes the flag off
+
+The grid's keys, on the photograph that is open (FR-UI-5, 2026-09-19). Judging here does not move on to the next frame: that belongs to culling, and in develop the photograph in front of you is the one being worked on.
+
+<sub>`ui/dr-ui/ui/app.slint:2711`</sub>
+
+### Give this photograph a colour label
+
+- **Touch** — Tap Label in the top bar, then a colour
+- **Pointer** — Click Label in the top bar, then a colour
+- **Keyboard** — `6` red, `7` yellow, `8` green, `9` blue; the same key again takes it off
+- **See it** — [in the manual](manual/README.md#rating-and-flagging)
+
+The grid's keys, on the photograph that is open, so labelling while stepping through a folder is one hand's work. The bar names the label in words beside its mark.
+
+<sub>`ui/dr-ui/ui/app.slint:2741`</sub>
 
 ### Move to the next or previous photograph
 
 - **Touch** — Tap a frame in the roll along the foot of the canvas
 - **Pointer** — Click a frame in the roll
-- **Keyboard** — Right arrow or space for the next, left arrow for the one before
+- **Keyboard** — `→`, `D` or `Space` for the next; `←` or `A` for the one before
+- **See it** — [in the manual](manual/README.md#moving-between-photographs)
 
-The edit on screen is saved on the way out, so stepping through a folder is as much a departure as going back to the grid and loses nothing.
+The edit on screen is saved on the way out, so stepping through a folder is as much a departure as going back to the grid and loses nothing. A and D as well as the arrows, so the left hand steps along the roll while the right stays on the mouse.
 
-<sub>`ui/dr-ui/ui/app.slint:2333`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2766`</sub>
 
 ### See the photograph before you edited it
 
 - **Touch** — Press and hold "Before"
 - **Pointer** — Press and hold "Before"
-- **Keyboard** — Hold \
+- **Keyboard** — Hold `\`
+- **See it** — [in the manual](manual/README.md#light)
 
 Held rather than toggled, and no split screen: a split halves the working image on the tablet the column was sized for, and the comparison photographers describe making is a flick back and forth. It takes no history step, so checking whether a frame is overcooked costs nothing to undo afterwards.
 
-<sub>`ui/dr-ui/ui/app.slint:2457`</sub>
+<sub>`ui/dr-ui/ui/app.slint:2896`</sub>
 
 ### Put one control back to its default
 
 - **Touch** — Double-tap its track
 - **Pointer** — Double-click its track, or right-click it
-- **Keyboard** — R, for the control last moved
+- **Keyboard** — `R`, for the control last moved
 
 The column is 280px wide and the colour mixer alone puts thirty-six of these in it, so a reset button per row would be most of the width. Two ways in with a pointer because right-click is the one a hand already reaches for and double-click is the one that needs no second button. A group's own reset is in its heading; this is the single control.
 
@@ -137,6 +270,7 @@ The column is 280px wide and the colour mixer alone puts thirty-six of these in
 
 - **Touch** — Press and hold the eye beside the snapshot
 - **Pointer** — Press and hold the eye beside the snapshot
+- **See it** — [in the manual](manual/README.md#history-snapshots-presets)
 
 The same hold as "Before", against a point the photographer chose rather than the file: "the version I liked twenty minutes ago" is how a choice between two treatments is actually made. It takes no history step and changes nothing; letting go puts the edit back.
 
@@ -146,38 +280,81 @@ The same hold as "Before", against a point the photographer chose rather than th
 
 - **Touch** — Type a name in the History panel and press Snapshot
 - **Pointer** — Type a name in the History panel and press Snapshot
+- **See it** — [in the manual](manual/README.md#history-snapshots-presets)
 
 The history is forgotten with the sitting, on purpose; a snapshot is the photographer saying this one should not be. It is written into the sidecar as a version of the edit, so it survives a restart and reaches the other device. Pressing a snapshot puts the photograph back to it, as one step that undo takes back whole.
 
-<sub>`ui/dr-ui/ui/history.slint:307`</sub>
+<sub>`ui/dr-ui/ui/history.slint:308`</sub>
 
 ### Show or hide one mask layer
 
 - **Touch** — Tap the ring at the head of its row
 - **Pointer** — Click the ring at the head of its row
-- **Keyboard** — H, for the selected layer history step, unlike holding "Before" — the layer really is off until it is switched back on.
+- **Keyboard** — `H`, for the selected layer
+- **See it** — [in the manual](manual/README.md#local-adjustments)
 
-Disabling a layer is the before-and-after a local edit constantly wants, so it is one press away rather than inside the row. It is an edit and does take a
+Disabling a layer is the before-and-after a local edit constantly wants, so it is one press away rather than inside the row. It is an edit and does take a history step, unlike holding "Before" — the layer really is off until it is switched back on.
 
-<sub>`ui/dr-ui/ui/masks.slint:212`</sub>
+<sub>`ui/dr-ui/ui/masks.slint:213`</sub>
 
 ### Show or hide one mask on the photograph
 
 - **Touch** — Tap the eye on its row
 - **Pointer** — Click the eye on its row
+- **See it** — [in the manual](manual/README.md#local-adjustments)
 
 A mask is judged by seeing where it falls, and two are judged by seeing where they meet — so each row has its own eye rather than the panel having one, and the eye is drawn in the colour the mask shows in, so the row says which shape on the picture is its. Nothing about the edit changes: this is how the photograph is looked at, and takes no history step.
 
-<sub>`ui/dr-ui/ui/masks.slint:279`</sub>
+<sub>`ui/dr-ui/ui/masks.slint:280`</sub>
+
+### Change how a part joins its mask
+
+- **Touch** — Tap the + / − / ∩ chip on the part's row
+- **Pointer** — Click the + / − / ∩ chip on the part's row
+- **Keyboard** — `J` turns the selected part's chip, while masking
+- **See it** — [in the manual](manual/README.md#local-adjustments)
+
+A chip that cycles rather than a menu, because a photographer flips a join while looking at the picture, not at the panel: add, take away, keep only where both agree, and round.
+
+<sub>`ui/dr-ui/ui/masks.slint:367`</sub>
 
 ### Leave one part out of a mask, and put it back
 
 - **Touch** — Tap the ring on the part's row
 - **Pointer** — Click the ring on the part's row
+- **See it** — [in the manual](manual/README.md#local-adjustments)
 
 The question a correction raises is whether it did what it was for — whether the stroke filled the shoulder, whether the subtracted gradient took only the sky. Removing it answers that and loses it. The same ring the layer wears, one row down, because it is the same question about a smaller thing.
 
-<sub>`ui/dr-ui/ui/masks.slint:394`</sub>
+<sub>`ui/dr-ui/ui/masks.slint:409`</sub>
+
+## Everywhere
+
+### Close what is open, or go back a step
+
+- **Touch** — The system Back gesture, or the Back button
+- **Pointer** — The Close or Back button on whatever is open
+- **Keyboard** — `Escape`, or `Back` where the device has one
+
+One key for "up one", innermost first: a question before the sheet under it, a sheet before the view, a view before the library. Nothing is left behind a dialogue that the key walked straight past.
+
+<sub>`ui/dr-ui/ui/app.slint:987`</sub>
+
+### Do what a sheet offers
+
+- **Touch** — Press its button — Export, or Copy
+- **Pointer** — Press its button — Export, or Copy
+- **Keyboard** — `Enter`, on the export and copy sheets
+
+<sub>`ui/dr-ui/ui/app.slint:997`</sub>
+
+### Scroll by the scrollbar
+
+- **Pointer** — Drag the bar along the right-hand edge of a list, or click the track above or below it to move by a page
+
+A list cut off at its edge looks, to a mouse, like a list that ends there — the film stocks past the tenth read as deleted. The bar says there is more and where the view is in it, and it is the one way to scroll that needs neither a wheel nor a drag on the content, which may be a row that would take the click.
+
+<sub>`ui/dr-ui/ui/widgets.slint:1074`</sub>
 
 ## Collections sidebar
 
@@ -185,37 +362,94 @@ The question a correction raises is whether it did what it was for — whether t
 
 - **Touch** — Press and hold it until it lifts, then drag it
 - **Pointer** — Drag it, or hold it until it lifts and then drag
+- **See it** — [in the manual](manual/README.md#collections)
 
 The tree is inside a Flickable, which claims any drag beginning inside it — so with a finger a drag on a row is a scroll until something says otherwise. The hold is that something, and it is what every mobile list already uses to pick a row up. The row lifts the moment it fires, so the gesture says it has been understood before anything moves.
 
-<sub>`ui/dr-ui/ui/collections.slint:335`</sub>
+<sub>`ui/dr-ui/ui/collections.slint:352`</sub>
 
 ### Act on a collection — rename, nest, un-nest, delete
 
 - **Touch** — Press and hold the collection, then let go without moving
 - **Pointer** — Right-click it
+- **See it** — [in the manual](manual/README.md#collections)
 
 The hold arms a drag and opens this menu, and which one you get is decided by whether you moved — the same fork the grid uses. One menu for everything done to a row, because there is one hold per row: while the hold opened the offline question by itself, nothing else the tree can do had a touch route.
 
-<sub>`ui/dr-ui/ui/collections.slint:346`</sub>
+<sub>`ui/dr-ui/ui/collections.slint:364`</sub>
 
 ### Take a collection back out of the one it is nested in
 
 - **Touch** — Hold it, then drag it onto "All photographs" — or let go and choose "Move to top level"
 - **Pointer** — Drag it onto "All photographs", or right-click it and choose "Move to top level"
+- **See it** — [in the manual](manual/README.md#collections)
 
 Nesting is a drag of one row onto another, and its inverse had no gesture at all: "All photographs" refused every drop, which is right for a photograph — it is already in the library — and wrong for a collection, which has a top level to be returned to. Without it a collection dragged into another was in there permanently.
 
-<sub>`ui/dr-ui/ui/collections.slint:356`</sub>
+<sub>`ui/dr-ui/ui/collections.slint:375`</sub>
 
 ### Rename a collection
 
 - **Touch** — Hold the collection, then "Rename"
 - **Pointer** — Double-click its name, or right-click it and choose "Rename"
+- **Keyboard** — Type the name and press `Enter`; `Escape` abandons it
+- **See it** — [in the manual](manual/README.md#collections)
 
 Double-click is what a file manager and a Lightroom panel use for the same thing, so it needs no discovering — but nothing on screen says so, which is what the menu item is for.
 
-<sub>`ui/dr-ui/ui/collections.slint:369`</sub>
+<sub>`ui/dr-ui/ui/collections.slint:389`</sub>
+
+### Review duplicate originals
+
+- **Touch** — Tap "Duplicate originals" under the trash
+- **Pointer** — Click "Duplicate originals" under the trash
+- **See it** — [in the manual](manual/README.md#duplicate-originals)
+
+Under the trash because the trash is where the spare copies go, and only while the catalog holds any: a row that is always there and usually empty is noise.
+
+<sub>`ui/dr-ui/ui/collections.slint:884`</sub>
+
+## Duplicate originals
+
+### Check duplicate originals are the same file
+
+- **Touch** — Tap "Check"
+- **Pointer** — Click "Check"
+- **See it** — [in the manual](manual/README.md#duplicate-originals)
+
+Nothing is moved on the catalog's say-so. The check reads the first and last megabyte of each copy and its sidecar, keeps what it read, and drops any group whose bytes or edits differ.
+
+<sub>`ui/dr-ui/ui/duplicates.slint:233`</sub>
+
+### Move the spare copies to the trash
+
+- **Touch** — Tap "Move N copies to trash"
+- **Pointer** — Click "Move N copies to trash"
+- **See it** — [in the manual](manual/README.md#duplicate-originals)
+
+The one verb, named with its count, and off until a check has proved something. Each group is merged onto the copy that stays and the others go to the trash together or not at all; the trash view gives them back.
+
+<sub>`ui/dr-ui/ui/duplicates.slint:248`</sub>
+
+### Choose which copy of a duplicate stays
+
+- **Touch** — Tap the copy's path
+- **Pointer** — Click the copy's path
+- **See it** — [in the manual](manual/README.md#duplicate-originals)
+
+The rule picks the copy outside a backup folder that still has the camera's name. The path is the evidence, so the path is what is tapped to overrule it.
+
+<sub>`ui/dr-ui/ui/duplicates.slint:313`</sub>
+
+### Leave a duplicate group as it is
+
+- **Touch** — Untick "Include" on the group
+- **Pointer** — Untick "Include" on the group
+- **See it** — [in the manual](manual/README.md#duplicate-originals)
+
+The plan is every group the check proved; a group you want to keep in two places is taken out of it rather than argued with.
+
+<sub>`ui/dr-ui/ui/duplicates.slint:323`</sub>
 
 ## People
 
@@ -223,37 +457,59 @@ Double-click is what a file manager and a Lightroom panel use for the same thing
 
 - **Touch** — Tap the faces that do not belong, then "Split off"
 - **Pointer** — Click the faces that do not belong, then "Split off"
+- **See it** — [in the manual](manual/README.md#people)
 
 Grouping over-merges on siblings, on parents and children, and on the same person a decade apart, so splitting is as prominent as merging. A tool that can only merge makes its own errors permanent.
 
-<sub>`ui/dr-ui/ui/identity.slint:188`</sub>
+<sub>`ui/dr-ui/ui/identity.slint:189`</sub>
 
 ### Rule on a suggested face
 
 - **Touch** — Tick to confirm it, cross to reject it
 - **Pointer** — Tick to confirm it, cross to reject it
+- **See it** — [in the manual](manual/README.md#people)
 
 A face is either the system's guess or the user's judgement, and the two are never conflated. A rejection is remembered, so the face is not suggested for that person again. The gesture note above is the whole label: a tick and a cross are only "confirm" and "reject" to someone who can see the suggestion they sit beside, and `IconButton`'s fallback would announce them as "check" and "cross" — two icon names that say nothing about which person is being ruled on.
 
-<sub>`ui/dr-ui/ui/identity.slint:208`</sub>
+<sub>`ui/dr-ui/ui/identity.slint:210`</sub>
+
+### Move between people
+
+- **Touch** — Tap a person on the rail
+- **Pointer** — Click a person on the rail
+- **Keyboard** — `↑` and `↓`, along the rail
+
+<sub>`ui/dr-ui/ui/identity.slint:385`</sub>
+
+### Name a person
+
+- **Touch** — Tap the name field over their faces, type, and press Enter
+- **Pointer** — Click the name field over their faces, type, and press Enter
+- **Keyboard** — `F2` puts the name field under the keys
+
+The rename key everywhere else. Enter finishes the name and hands the keys back to the rail, so naming a run of people is `Down`, `F2`, a name, Enter, over and over.
+
+<sub>`ui/dr-ui/ui/identity.slint:391`</sub>
 
 ### See a person's photographs
 
 - **Touch** — Choose them in the rail, then "Show photos"
 - **Pointer** — Choose them in the rail, then "Show photos"
+- **See it** — [in the manual](manual/README.md#people)
 
 This is the point of having identified anybody. Without it the screen is a filing cabinet with no drawer handles.
 
-<sub>`ui/dr-ui/ui/identity.slint:635`</sub>
+<sub>`ui/dr-ui/ui/identity.slint:689`</sub>
 
 ### Change how faces are grouped
 
 - **Touch** — "Grouping…", move the dials, then Regroup
 - **Pointer** — "Grouping…", move the dials, then Regroup
+- **See it** — [in the manual](manual/README.md#people)
 
 The right match confidence is a property of your library, not of the model. "What would this do?" answers for this library without writing anything; names, confirmations and the groups you have set aside are kept whatever the dials say.
 
-<sub>`ui/dr-ui/ui/identity.slint:672`</sub>
+<sub>`ui/dr-ui/ui/identity.slint:727`</sub>
 
 ## Library grid
 
@@ -261,133 +517,273 @@ The right match confidence is a property of your library, not of the model. "Wha
 
 - **Touch** — Press and hold a photograph, or press Select in the header
 - **Pointer** — Ctrl-click, or press Select in the header
+- **See it** — [in the manual](manual/README.md#selecting-several)
 
 Touch has no ctrl, so without a mode there is no way to select a second photograph — the first tap would open it. The hold is the fast way in and the button is the one that can be found.
 
-<sub>`ui/dr-ui/ui/library.slint:1572`</sub>
+<sub>`ui/dr-ui/ui/library.slint:1701`</sub>
 
 ### Add or remove one photograph
 
 - **Touch** — While selecting, tap it
 - **Pointer** — Ctrl-click it
+- **See it** — [in the manual](manual/README.md#selecting-several)
 
 While selecting, a tap never opens. That is the whole point of the mode: one meaning per gesture at a time. Press Done to get tap-to-open back.
 
-<sub>`ui/dr-ui/ui/library.slint:1581`</sub>
+<sub>`ui/dr-ui/ui/library.slint:1711`</sub>
 
 ### Leave selecting
 
 - **Touch** — Press Done in the header
 - **Pointer** — Press Done in the header
-- **Keyboard** — Escape
+- **Keyboard** — `Escape`, or `Back`; an open sheet closes first
+- **See it** — [in the manual](manual/README.md#selecting-several)
 
-<sub>`ui/dr-ui/ui/library.slint:1589`</sub>
+<sub>`ui/dr-ui/ui/library.slint:1720`</sub>
 
 ### Pick a photograph up to drag it
 
 - **Touch** — Press and hold it until a ring opens around it, then drag
 - **Pointer** — Drag it
+- **See it** — [in the manual](manual/README.md#collections)
 
 A finger on a photograph might be starting a scroll, and for the first half-second the grid assumes it is. Holding says otherwise, and the ring is the grid saying it heard — from there the drag cannot be lost to a scroll. A mouse never waits: the cursor is precise enough that a sideways drag is unambiguous from the first pixel.
 
-<sub>`ui/dr-ui/ui/library.slint:1619`</sub>
+<sub>`ui/dr-ui/ui/library.slint:1751`</sub>
 
 ### Select a range
 
 - **Touch** — While selecting, press "Select to…", then tap the last photograph of the run
 - **Pointer** — Shift-click the last photograph of the run
+- **Keyboard** — Shift with any key that walks the grid — `Shift+←`, `Shift+→`, `Shift+↑`, `Shift+↓`, `Shift+Page Up`, `Shift+Page Down`, `Shift+Home`, `Shift+End`
+- **See it** — [in the manual](manual/README.md#selecting-several)
 
 This replaced a double tap, which had no visible state and could take forty photographs by accident. The run is resolved by the catalog rather than by what is on screen, so the grid can scroll between the two taps — the ranges that hurt on a tablet are longer than a screenful, which is exactly where a finger sweep runs out.
 
-<sub>`ui/dr-ui/ui/library.slint:1684`</sub>
+<sub>`ui/dr-ui/ui/library.slint:1817`</sub>
 
 ### Take the blinks out of a burst
 
 - **Touch** — Narrow to a person, then tap "Eyes open" beside their name on the filter bar
 - **Pointer** — Narrow to a person, then click "Eyes open" beside their name on the filter bar
+- **See it** — [in the manual](manual/README.md#bursts)
 
 Face indexing reads each face's eyes. The chip drops frames where the chosen people are caught blinking, and leaves sunglasses and eyes it could not read alone.
 
-<sub>`ui/dr-ui/ui/library.slint:2383`</sub>
+<sub>`ui/dr-ui/ui/library.slint:2534`</sub>
 
 ### Find photographs with two people in them
 
 - **Touch** — Open the People chip on the filter bar, tap each name, then switch the chip beside them to "all of them"
 - **Pointer** — Open the People chip on the filter bar, click each name, then switch the chip beside them to "all of them"
+- **See it** — [in the manual](manual/README.md#people)
 
 "Any of them" is a union and "all of them" is an intersection. The tray is where both terms and the choice between them live, because a filter belongs on the filter bar.
 
-<sub>`ui/dr-ui/ui/library.slint:2412`</sub>
+<sub>`ui/dr-ui/ui/library.slint:2564`</sub>
+
+### Show only photographs with one colour label
+
+- **Touch** — Tap its chip in the filter bar
+- **Pointer** — Click its chip in the filter bar
+- **See it** — [in the manual](manual/README.md#rating-and-flagging)
+
+Each chip is the label's mark and its name, so the one you want is found by reading it; tap the lit chip again to show every label.
+
+<sub>`ui/dr-ui/ui/library.slint:2688`</sub>
+
+### Export the selection as the last export was
+
+- **Touch** — Select them, then Export in the selection bar
+- **Pointer** — Select them, then Export in the selection bar
+- **Keyboard** — `Ctrl+Shift+E`, or `Ctrl+E` to see the export settings first
+- **See it** — [in the manual](manual/README.md#export)
+
+Lightroom's and darktable's chords. Every export runs on the saved defaults, so the plain chord opens them beside an Export button and the shifted one skips straight to exporting.
+
+<sub>`ui/dr-ui/ui/library.slint:3158`</sub>
+
+### Paste copied settings onto the selection
+
+- **Touch** — Select them, then "Paste to N" in the selection bar
+- **Pointer** — Select them, then "Paste to N"
+- **Keyboard** — `Ctrl+V`
+- **See it** — [in the manual](manual/README.md#copying-settings)
+
+<sub>`ui/dr-ui/ui/library.slint:3182`</sub>
+
+### Keyword the selection
+
+- **Touch** — Select them, then Keywords in the selection bar
+- **Pointer** — Select them, then Keywords in the selection bar
+- **Keyboard** — `Ctrl+K`
+
+Lightroom's keywording chord. The sheet opens with its field ready for typing, so the keys that judge in the grid are out of the way until it closes.
+
+<sub>`ui/dr-ui/ui/library.slint:3211`</sub>
+
+### Show only photographs with some number of stars
+
+- **Touch** — Tap a star chip in the filter bar
+- **Pointer** — Click a star chip in the filter bar
+- **Keyboard** — Hold `F` and tap a digit, `0`–`5`, for exactly that many stars, or two digits for everything between them; tap `F` alone to show every rating again
+- **See it** — [in the manual](manual/README.md#rating-and-flagging)
+
+The chips say "this many or more". A range with a ceiling — the twos and threes still to be decided — is the keyboard's alone, and the bar says so in words while it holds.
+
+<sub>`ui/dr-ui/ui/library.slint:3245`</sub>
+
+### Give photographs a colour label
+
+- **Touch** — Select them, then Label in the selection bar and tap a colour
+- **Pointer** — Select them, then Label in the selection bar and click a colour
+- **Keyboard** — `6` red, `7` yellow, `8` green, `9` blue, with the pointer over it or on the selection; the same key again takes the label off
+- **See it** — [in the manual](manual/README.md#rating-and-flagging)
+
+Lightroom's keys, so hands that learned them there need not learn them again. Purple has no key there either, and is on the bar. Every mark carries its label's initial, so the label is read without telling the colours apart.
+
+<sub>`ui/dr-ui/ui/library.slint:3295`</sub>
+
+### Pick or reject a photograph
+
+- **Touch** — Select them, then Flag in the selection bar and Pick, Reject or No flag
+- **Pointer** — Select them, then Flag in the selection bar and Pick, Reject or No flag
+- **Keyboard** — `P` picks, `X` rejects and `U` takes the flag off, with the pointer over it or on the selection
+
+The keys every culling tool uses, so muscle memory built elsewhere works here.
+
+<sub>`ui/dr-ui/ui/library.slint:3319`</sub>
+
+### Move photographs to the trash
+
+- **Touch** — Tap the bin at the start of a cell's stars
+- **Pointer** — Hover the cell and click the bin before its stars
+- **Keyboard** — `Delete` or `Backspace`, on the selection
+
+The bin acts on one photograph, so a stray click cannot trash a selection; the key acts on the selection because that is what every file manager's Delete does. Both are undone from the trash view.
+
+<sub>`ui/dr-ui/ui/library.slint:3346`</sub>
+
+### Open this list
+
+- **Touch** — Press Help in the header, and Done to put it away
+- **Pointer** — Press Help in the header, and Done to put it away
+- **Keyboard** — `F1`, and `Escape` to put it away
+
+<sub>`ui/dr-ui/ui/library.slint:3371`</sub>
+
+### Rename the collection the grid is showing
+
+- **Touch** — Hold it in the sidebar, then "Rename"
+- **Pointer** — Double-click it in the sidebar
+- **Keyboard** — `F2`
+
+<sub>`ui/dr-ui/ui/library.slint:3379`</sub>
+
+### Move through the grid
+
+- **Touch** — Scroll, and tap a photograph
+- **Pointer** — Scroll, and click a photograph
+- **Keyboard** — `←`, `→`, `↑` and `↓` move one photograph; `Page Up` and `Page Down` a screenful; `Home` and `End` to the first and the last
+
+The cursor selects what it lands on, so walking and judging are one hand's work.
+
+<sub>`ui/dr-ui/ui/library.slint:3399`</sub>
 
 ### Resize the thumbnails
 
 - **Touch** — Pinch the grid with two fingers
 - **Pointer** — Ctrl and the scroll wheel
+- **Keyboard** — `=` or `Plus` for larger, `-` for smaller
+- **See it** — [in the manual](manual/README.md#getting-about)
 
 There is no wheel on a tablet, so without the pinch the cell size could only be changed by a control a finger cannot reach.
 
-<sub>`ui/dr-ui/ui/library.slint:3072`</sub>
+<sub>`ui/dr-ui/ui/library.slint:3530`</sub>
 
 ### File photographs in a collection
 
 - **Touch** — Drag a photograph — or a whole selection — onto a collection in the sidebar. Starting a drag stops the press becoming a hold, so it cannot leave you in selection mode.
 - **Pointer** — Drag a photograph — or a whole selection — onto a collection in the sidebar
+- **See it** — [in the manual](manual/README.md#collections)
 
 The selection is what the drag carries, which is why selecting several is worth the mode: forty photographs file in one gesture.
 
-<sub>`ui/dr-ui/ui/library.slint:3269`</sub>
+<sub>`ui/dr-ui/ui/library.slint:3729`</sub>
 
 ### Open a photograph
 
 - **Touch** — Tap it — a single tap, any length
 - **Pointer** — Click it
+- **Keyboard** — `Enter`, on the photograph the arrow keys have walked to
+- **See it** — [in the manual](manual/README.md#developing-a-photograph)
 
 A tap opens; a tap that *moved* does not. Travel is what separates a deliberate tap from a hand brushing past, and it is the only thing that does: the two are the same length. An earlier version required the finger to dwell 120 ms instead, and that rejected ordinary taps — a real tap is often quicker than a brush.
 
-<sub>`ui/dr-ui/ui/library.slint:3537`</sub>
+<sub>`ui/dr-ui/ui/library.slint:4034`</sub>
 
 ### Rate a photograph without opening it
 
 - **Touch** — Tap a star on the cell
 - **Pointer** — Hover the cell, then click a star
-- **Keyboard** — 0 to 5 on the selection
+- **Keyboard** — `0`–`5` with the pointer over it, or on the selection
+- **See it** — [in the manual](manual/README.md#rating-and-flagging)
 
 A star has to take the press without it also reaching the cell, or every rating throws the user into develop.
 
-<sub>`ui/dr-ui/ui/library.slint:3657`</sub>
+<sub>`ui/dr-ui/ui/library.slint:4157`</sub>
 
 ### Choose the frame a folded burst shows
 
 - **Touch** — Open the burst, then tap the ring on the frame you want
 - **Pointer** — Open the burst, then click the ring on the frame you want
+- **See it** — [in the manual](manual/README.md#bursts)
 
 A folded burst draws its earliest frame, which is a fact about the clock and not a judgement about the photograph — nothing in this application ranks a frame (FR-CULL-5). But the point of a burst is that one of the twelve is better than the other eleven, and the photographer is the only one who knows which. So the choice is offered on the frames themselves, while they are open and side by side, which is the one moment the alternatives are on screen to be compared.
 
-<sub>`ui/dr-ui/ui/library.slint:3788`</sub>
+<sub>`ui/dr-ui/ui/library.slint:4290`</sub>
 
 ### Drop the selection but keep selecting
 
 - **Touch** — Press Clear in the selection strip
 - **Pointer** — Press Clear in the selection strip
+- **Keyboard** — `Ctrl+D` or `Ctrl+Shift+A`
+- **See it** — [in the manual](manual/README.md#selecting-several)
 
 Distinct from Done, which leaves the mode entirely. Clearing keeps it, so the next selection can start straight away.
 
-<sub>`ui/dr-ui/ui/library.slint:4454`</sub>
+<sub>`ui/dr-ui/ui/library.slint:4981`</sub>
 
 ### Select everything the grid is showing
 
 - **Touch** — While selecting, press "Select all"
 - **Pointer** — While selecting, press "Select all"
+- **Keyboard** — `Ctrl+A`
+- **See it** — [in the manual](manual/README.md#selecting-several)
 
 A scoped grid of two hundred frames is two hundred taps otherwise, and "all of them, except those three" is a far more common shape than the taps it took to say it.
 
-<sub>`ui/dr-ui/ui/library.slint:4471`</sub>
+<sub>`ui/dr-ui/ui/library.slint:5000`</sub>
 
 ### Take photographs out of a collection
 
 - **Touch** — Select them, then "Collections…" in the selection bar
 - **Pointer** — Select them, then "Collections…" in the selection bar
+- **See it** — [in the manual](manual/README.md#collections)
 
 The badge on a cell says a photograph is filed in three collections and never which. This is the sheet that names them, and the only way out of one the grid is not currently scoped to.
 
-<sub>`ui/dr-ui/ui/library.slint:4595`</sub>
+<sub>`ui/dr-ui/ui/library.slint:5181`</sub>
+
+## Settings
+
+### Review duplicate originals
+
+- **Touch** — Tap "Review duplicate originals"
+- **Pointer** — Click "Review duplicate originals"
+- **See it** — [in the manual](manual/README.md#duplicate-originals)
+
+Beside the other whole-library passes, because it is one; the sidebar offers the same page under the trash.
+
+<sub>`ui/dr-ui/ui/settings.slint:533`</sub>

docs/manual/README.md

@@ -33,20 +33,43 @@ wrote first.
 
 Photographs are ordered by capture time, with a month heading where each
 begins. The strip on the left is the timeline — drag it to jump to a year.
-The bar above the grid filters by rating, flag and where the file is (on this
-device, or only on the server).
+The bar above the grid filters by rating, flag, colour label and where the
+file is (on this device, or only on the server).
 
 ### Rating and flagging
 
 Hover a cell and the stars appear; click one. The filter chips above the grid
-count what each rating holds, and clicking `3+` shows only those.
+count what each rating holds, and clicking `3+` shows only those. From the
+keyboard, `0` to `5` set the stars on the photograph under the pointer or on
+the selection; `P` picks, `X` rejects and `U` takes the flag off, and `Flag`
+on the selection bar does the same with the pointer. In
+develop the same keys judge the open photograph and stay on it, and the top
+bar carries its stars and `Pick` and `Reject`; the roll marks each frame's
+flag and stars.
 
 ![Rating two photographs, then filtering the grid to three stars and more](media/library-rating.gif)
 
+Colour labels work as Lightroom's do: `6` red, `7` yellow, `8` green, `9`
+blue, on the photograph under the pointer or on the selection, and the same
+key again takes the label off. `Label` on the selection bar offers all five,
+purple included, and `None`. Each label is drawn with its initial on it, so
+it reads without telling the colours apart, and the filter bar has a chip for
+each. In develop, the top bar names the open photograph's label and sets it,
+and the same keys work there.
+
+![Labelling four frames with 6, 7, 8 and 9, taking one off with the same key, and filtering the grid to green](media/library-labels.gif)
+
 ### Getting about
 
 Drag the timeline to scrub through years; Ctrl and the wheel resize the
-thumbnails.
+thumbnails. On a desktop a scrollbar beside the grid says how far through the
+library the view is — drag its thumb, or click the track to move a page — and
+the sidebar, the develop column and Settings have one too whenever they run
+past the window. On a tablet they scroll by flick alone.
+
+`Help` in the header, or `F1`, opens the controls and shortcuts: every key and
+gesture, screen by screen, with `See it` beside those this page shows, and
+`Manual` to open this page. In develop it is the `?` beside `Settings`.
 
 ![Scrubbing the timeline](media/library-timeline.gif)
 
@@ -86,6 +109,49 @@ move back to the top level, keep it offline, delete.
 
 ![The menu on a collection](media/library-collection-menu.png)
 
+### Bursts
+
+Frames taken a moment apart that look alike fold into one cell, with a badge
+counting them. Click the badge — tap it, on a tablet — to open the burst in
+the grid, and again to fold it back. A folded burst shows its earliest frame;
+to have it show another, open it and click the ring on the frame you want.
+
+With faces indexed, narrowing the grid to a person puts `Eyes open` beside
+their name on the filter bar, which leaves out the frames where they blinked.
+
+### Duplicate originals
+
+A library put together by hand often holds the same RAW more than once — a
+dated folder, a `bck` folder beside it, a renamed copy another program
+exported. When the catalog finds files with the same camera, capture time
+and size in more than one place, `Duplicate originals` appears under the
+trash in the sidebar with how many there are; Settings offers the same page
+beside the other whole-library passes.
+
+The page lists each group with its picture and its paths. One copy is
+marked `Stays`: the one outside a folder named like a backup (`bck`,
+`backup`, `copy`, `old`…), then the one still named the way the camera named
+it (`_MG_4623`, `IMG_0001`, `DSC_0042`), then the one catalogued first. Tap
+another path to keep that copy instead, and untick `Include` to leave a
+group alone.
+
+Nothing moves until `Check` has read the first and last megabyte of every
+copy and each copy's sidecar. A group whose files differ, or whose copies
+carry two different edits, is marked skipped and says why — two edits of
+one frame are kept for virtual copies. What the check reads is kept, so a
+second visit costs nothing. The line under each group says what the copy
+that stays will gain: the highest rating, every keyword and collection, a
+flag or label the copies agree on, faces and the names on them, and the
+edit if only a copy had one. Where the copies disagree on a flag or a
+label, the one that stays keeps its own and the line says so.
+
+![Two frames with a copy each in a bck folder, checked and proved the same, the copy that stays marked on each](media/duplicates.png)
+
+`Move N copies to trash` does it, one group at a time: each group is merged
+and its spare copies moved together, or not at all. The copies go to the
+trash, not away — select them in the trash view and `Restore` puts them
+back where they were.
+
 ## Developing a photograph
 
 Click a thumbnail to open it. The column on the right is every adjustment;
@@ -105,9 +171,19 @@ Hold `Before` to see the photograph as it was.
 ### Looking closer
 
 Double-click for 1:1; drag to move about; double-click again to fit. The
-wheel zooms to any amount in between.
+wheel zooms to any amount in between. Past 1:1 the file's own pixels are
+drawn as hard-edged blocks rather than smoothed, so what you see is what
+the sensor recorded.
 
-![Zooming to 1:1, panning, and back](media/develop-zoom.gif)
+![Zooming to 1:1 with a double-click, panning, then further in with the wheel](media/develop-zoom.gif)
+
+### Moving between photographs
+
+The roll along the foot of the canvas holds the photographs the grid was
+showing; click one to open it. The right arrow, `D` or space opens the next,
+and the left arrow or `A` the one before; held down, they go on past the
+stretch the roll has loaded, through everything the grid would show. The
+edit on screen is saved on the way, so stepping along a shoot loses nothing.
 
 ### White balance from the photograph
 
@@ -126,6 +202,17 @@ ratio from the chips. `Done composing` returns to the photograph.
 
 ![Cropping, straightening and choosing a ratio](media/compose.gif)
 
+`Vertical` and `Horizontal`, under `Straighten`, correct the lines that
+converge when the camera is tilted up at a building; the crop refits to
+what is left.
+
+![Standing towers shot from below upright with the Vertical slider](media/compose-perspective.gif)
+
+A crop that leaves a mask wholly outside the frame says so, and offers to
+take the crop back or keep it.
+
+![A stroke in a corner, a crop that leaves it outside, and the notice with Undo crop and Keep crop](media/crop-orphan.gif)
+
 ### Local adjustments
 
 `Local` in the rail turns the column into a mask stack. `Find subjects` runs a
@@ -144,7 +231,11 @@ tinted, as alpha, or as an outline; the eye on its row switches it off.
 ![Looking at the mask three ways, painting into it, then growing its edge](media/local-paint.gif)
 
 Linear and radial gradients, a tone range and a colour range are the other
-ways to make one; each can be combined with any other.
+ways to make one; each can be combined with any other. `∩ Intersect` keeps
+only where the new part and the mask agree: below, a gradient over the
+lower half, then two strokes that survive only where the gradient is.
+
+![A linear gradient, then Intersect and two painted strokes](media/local-intersect.gif)
 
 ### Repair
 
@@ -158,9 +249,11 @@ opacity are in the panel. Heal blends; clone copies.
 
 The `Film` chooser at the head of Adjust applies a spectral simulation of a
 named stock; below it, the print exposure and push controls a film has and a
-sensor does not.
+sensor does not. The list opens over the column and scrolls on its own — by
+wheel, drag or flick, or with `Up`, `Down` and `Enter` — down to the
+black-and-white stocks at its end.
 
-![Choosing Velvia, then holding Before](media/film.gif)
+![Opening the film list, scrolling it, choosing Velvia, then holding Before](media/film.gif)
 
 ### History, snapshots, presets
 
@@ -170,6 +263,15 @@ apply elsewhere, and imports `.xmp` from other applications.
 
 ![The presets sheet](media/presets.png)
 
+### Copying settings
+
+`Copy` in the top bar, or Ctrl+C, takes this photograph's settings; `Paste`,
+or Ctrl+V, puts them on another, and says what it would paste — how many
+adjustments, and whether the crop comes too. In the grid, `Paste to N` on the
+selection bar pastes onto every photograph selected. Which kinds of edit a
+copy carries is chosen in `Presets…`, or with Ctrl+Shift+C: a look carried
+across a shoot usually leaves each frame's own crop alone.
+
 ## Merging a panorama
 
 Select the frames, then `Merge to panorama` from the selection bar. The
@@ -200,9 +302,10 @@ set is refused, and the header says so.
 
 ![The settings page: background activity, indexing, storage, display](media/settings.png)
 
-Background activity with progress, thumbnail and face indexing, storage on
-this device, display and colour, export defaults, what is written to XMP
-sidecars, and a diagnostics bundle for a bug report.
+Background activity with progress, thumbnail and face indexing, how many
+duplicate originals the library holds and the way to their review, storage
+on this device, display and colour, export defaults, what is written to XMP
+sidecars, the manual, and a diagnostics bundle for a bug report.
 
 ## People
 
@@ -230,6 +333,11 @@ a private X server and records each scene; `record.sh <library>` re-makes
 every picture here. Run it after a change to the interface and commit what
 changed. The pictures are in LFS.
 
+The application carries this page. `cargo run -p traceability -- manual`
+renders it to `index.html` beside it, which the packages install with the
+pictures and the app opens from Help, from Settings, and from the "See it"
+link beside a gesture on the help sheet. CI fails when the two differ.
+
 Making it the first time turned up nine faults, each fixed in its own
 commit before the pictures were taken: the folder picker could not choose
 the top level, month headings overprinted each other, a category mask

docs/manual/index.html

@@ -0,0 +1,388 @@
+<!DOCTYPE html>
+<!-- GENERATED FILE — do not edit by hand. -->
+<!-- Source: docs/manual/README.md. Regenerate: cargo run -p traceability -- manual -->
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<meta name="color-scheme" content="light dark">
+<title>DarkRoom, shown</title>
+<style>
+:root {
+  --bg: #fbfaf8;
+  --ink: #1d1c1a;
+  --ink-dim: #5c5955;
+  --rule: #dedad4;
+  --accent: #8a4b12;
+  --panel: #f1eee9;
+  --mark: #f6e3c7;
+}
+@media (prefers-color-scheme: dark) {
+  :root {
+    --bg: #161514;
+    --ink: #e9e6e1;
+    --ink-dim: #a39e97;
+    --rule: #34312d;
+    --accent: #e8a25c;
+    --panel: #201e1c;
+    --mark: #43321f;
+  }
+}
+* { box-sizing: border-box; }
+body {
+  margin: 0;
+  background: var(--bg);
+  color: var(--ink);
+  font: 17px/1.6 system-ui, -apple-system, "Segoe UI", Roboto, sans-serif;
+}
+.page {
+  display: grid;
+  grid-template-columns: 15rem minmax(0, 46rem);
+  gap: 3rem;
+  justify-content: center;
+  padding: 2rem 1.5rem 4rem;
+}
+.toc {
+  position: sticky;
+  top: 1.5rem;
+  align-self: start;
+  max-height: calc(100vh - 3rem);
+  overflow-y: auto;
+  font-size: 0.9rem;
+}
+.toc-title {
+  margin: 0 0 0.5rem;
+  color: var(--ink-dim);
+  font-size: 0.75rem;
+  font-weight: 700;
+  letter-spacing: 0.08em;
+  text-transform: uppercase;
+}
+.toc ul { list-style: none; margin: 0; padding: 0; }
+.toc ul ul { padding-left: 0.9rem; }
+.toc li { margin: 0.2rem 0; }
+.toc a { color: var(--ink-dim); text-decoration: none; }
+.toc a:hover { color: var(--accent); }
+main { min-width: 0; }
+h1, h2, h3 { line-height: 1.25; scroll-margin-top: 1rem; }
+h1 { font-size: 2.1rem; margin: 0 0 1rem; }
+h2 { font-size: 1.5rem; margin: 2.6rem 0 0.8rem; padding-top: 1rem; border-top: 1px solid var(--rule); }
+h3 { font-size: 1.15rem; margin: 1.8rem 0 0.6rem; }
+h2:target, h3:target { background: var(--mark); border-radius: 4px; padding-left: 0.3rem; margin-left: -0.3rem; }
+a { color: var(--accent); }
+code {
+  font: 0.88em ui-monospace, "Cascadia Mono", "DejaVu Sans Mono", monospace;
+  background: var(--panel);
+  border: 1px solid var(--rule);
+  border-radius: 4px;
+  padding: 0.05em 0.3em;
+}
+figure { margin: 1.4rem 0; }
+figure img, main img {
+  display: block;
+  width: 100%;
+  height: auto;
+  aspect-ratio: auto 16 / 11;
+  border-radius: 6px;
+  border: 1px solid var(--rule);
+}
+figcaption { margin-top: 0.4rem; color: var(--ink-dim); font-size: 0.9rem; }
+table { border-collapse: collapse; width: 100%; font-size: 0.95rem; }
+th, td { text-align: left; padding: 0.4rem 0.6rem; border-bottom: 1px solid var(--rule); vertical-align: top; }
+th { color: var(--ink-dim); font-weight: 600; }
+@media (max-width: 52rem) {
+  .page { grid-template-columns: minmax(0, 1fr); gap: 1rem; padding: 1rem 16px 3rem; }
+  .toc { position: static; max-height: none; border: 1px solid var(--rule); border-radius: 6px; padding: 0.8rem 1rem; background: var(--panel); }
+  body { font-size: 16px; }
+}
+</style>
+</head>
+<body>
+<div class="page">
+<nav class="toc" aria-label="Contents">
+<p class="toc-title">Contents</p>
+<ul>
+<li><a href="#opening-a-library">Opening a library</a></li>
+<li><a href="#the-library">The library</a>
+<ul>
+<li><a href="#rating-and-flagging">Rating and flagging</a></li>
+<li><a href="#getting-about">Getting about</a></li>
+<li><a href="#selecting-several">Selecting several</a></li>
+<li><a href="#collections">Collections</a></li>
+<li><a href="#bursts">Bursts</a></li>
+<li><a href="#duplicate-originals">Duplicate originals</a></li>
+</ul>
+</li>
+<li><a href="#developing-a-photograph">Developing a photograph</a>
+<ul>
+<li><a href="#light">Light</a></li>
+<li><a href="#looking-closer">Looking closer</a></li>
+<li><a href="#moving-between-photographs">Moving between photographs</a></li>
+<li><a href="#white-balance-from-the-photograph">White balance from the photograph</a></li>
+<li><a href="#composing">Composing</a></li>
+<li><a href="#local-adjustments">Local adjustments</a></li>
+<li><a href="#repair">Repair</a></li>
+<li><a href="#film">Film</a></li>
+<li><a href="#history-snapshots-presets">History, snapshots, presets</a></li>
+<li><a href="#copying-settings">Copying settings</a></li>
+</ul>
+</li>
+<li><a href="#merging-a-panorama">Merging a panorama</a></li>
+<li><a href="#export">Export</a></li>
+<li><a href="#settings">Settings</a></li>
+<li><a href="#people">People</a></li>
+<li><a href="#where-things-are-written-down">Where things are written down</a></li>
+<li><a href="#how-this-page-is-made">How this page is made</a></li>
+</ul>
+</nav>
+<main>
+<h1 id="darkroom-shown">DarkRoom, shown</h1>
+<p>A tour of what the application does, one picture per thing. Every image on
+this page was captured from the desktop build driving itself — nothing is a
+mock-up, and nothing has been retouched outside DarkRoom. Where a feature is
+better seen moving, it moves.</p>
+<p>The requirements behind each feature are in <a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/requirements.md">requirements.md</a>;
+the reasoning is in the design documents linked from each section. This page
+is only about what you see.</p>
+<p>The photographs are the author's. None show a person.</p>
+<h2 id="opening-a-library">Opening a library</h2>
+<p>DarkRoom opens on a library: a folder on this machine, a folder a sync
+client keeps, or a Nextcloud account. A folder needs no password and uploads
+nothing.</p>
+<figure><img loading="lazy" src="media/launch.png" alt="The launch screen: a server field, a folder field, and which formats to scan for"><figcaption>The launch screen: a server field, a folder field, and which formats to scan for</figcaption></figure>
+<p>Once a folder is named, it is the library — you are not asked for it again,
+and <code>Open library</code> opens it whole. <code>Subfolder…</code> narrows the scan to part of
+it. The formats ticked are what the scan looks for; RAW is on and JPEG off
+by default, because a RAW editor's sensible default is the file the camera
+wrote first.</p>
+<figure><img loading="lazy" src="media/launch-folder.png" alt="A folder chosen: the library, whether to scan a subfolder, and the formats"><figcaption>A folder chosen: the library, whether to scan a subfolder, and the formats</figcaption></figure>
+<h2 id="the-library">The library</h2>
+<figure><img loading="lazy" src="media/library.png" alt="The grid: collections on the left, the timeline beside it, the roll of thumbnails, and the filter bar above"><figcaption>The grid: collections on the left, the timeline beside it, the roll of thumbnails, and the filter bar above</figcaption></figure>
+<p>Photographs are ordered by capture time, with a month heading where each
+begins. The strip on the left is the timeline — drag it to jump to a year.
+The bar above the grid filters by rating, flag, colour label and where the
+file is (on this device, or only on the server).</p>
+<h3 id="rating-and-flagging">Rating and flagging</h3>
+<p>Hover a cell and the stars appear; click one. The filter chips above the grid
+count what each rating holds, and clicking <code>3+</code> shows only those. From the
+keyboard, <code>0</code> to <code>5</code> set the stars on the photograph under the pointer or on
+the selection; <code>P</code> picks, <code>X</code> rejects and <code>U</code> takes the flag off, and <code>Flag</code>
+on the selection bar does the same with the pointer. In
+develop the same keys judge the open photograph and stay on it, and the top
+bar carries its stars and <code>Pick</code> and <code>Reject</code>; the roll marks each frame's
+flag and stars.</p>
+<figure><img loading="lazy" src="media/library-rating.gif" alt="Rating two photographs, then filtering the grid to three stars and more"><figcaption>Rating two photographs, then filtering the grid to three stars and more</figcaption></figure>
+<p>Colour labels work as Lightroom's do: <code>6</code> red, <code>7</code> yellow, <code>8</code> green, <code>9</code>
+blue, on the photograph under the pointer or on the selection, and the same
+key again takes the label off. <code>Label</code> on the selection bar offers all five,
+purple included, and <code>None</code>. Each label is drawn with its initial on it, so
+it reads without telling the colours apart, and the filter bar has a chip for
+each. In develop, the top bar names the open photograph's label and sets it,
+and the same keys work there.</p>
+<figure><img loading="lazy" src="media/library-labels.gif" alt="Labelling four frames with 6, 7, 8 and 9, taking one off with the same key, and filtering the grid to green"><figcaption>Labelling four frames with 6, 7, 8 and 9, taking one off with the same key, and filtering the grid to green</figcaption></figure>
+<h3 id="getting-about">Getting about</h3>
+<p>Drag the timeline to scrub through years; Ctrl and the wheel resize the
+thumbnails. On a desktop a scrollbar beside the grid says how far through the
+library the view is — drag its thumb, or click the track to move a page — and
+the sidebar, the develop column and Settings have one too whenever they run
+past the window. On a tablet they scroll by flick alone.</p>
+<p><code>Help</code> in the header, or <code>F1</code>, opens the controls and shortcuts: every key and
+gesture, screen by screen, with <code>See it</code> beside those this page shows, and
+<code>Manual</code> to open this page. In develop it is the <code>?</code> beside <code>Settings</code>.</p>
+<figure><img loading="lazy" src="media/library-timeline.gif" alt="Scrubbing the timeline"><figcaption>Scrubbing the timeline</figcaption></figure>
+<figure><img loading="lazy" src="media/library-thumbsize.gif" alt="Resizing the thumbnails with Ctrl and the wheel"><figcaption>Resizing the thumbnails with Ctrl and the wheel</figcaption></figure>
+<h3 id="selecting-several">Selecting several</h3>
+<p><code>Select</code> in the header — or Ctrl-click — starts a selection. Shift-click picks
+a range. The bar at the foot of the grid is everything a selection can be done
+to: collections, keywords, presets, export, and merging to a panorama.</p>
+<figure><img loading="lazy" src="media/library-selection.png" alt="Twelve photographs selected, with the selection bar along the foot of the grid"><figcaption>Twelve photographs selected, with the selection bar along the foot of the grid</figcaption></figure>
+<p><code>Keywords</code> on that bar opens a sheet; type a word and press return, and it
+is on every photograph selected. The list below the field is every keyword
+the library has, ticked where the selection carries it.</p>
+<figure><img loading="lazy" src="media/library-keywords.gif" alt="Keywording twelve frames"><figcaption>Keywording twelve frames</figcaption></figure>
+<h3 id="collections">Collections</h3>
+<p><code>+</code> at the head of the sidebar makes one. Drag a photograph — or the whole
+selection — onto its row to file it there; click the row to see it. A
+photograph can be in several, and the badge on its cell counts them.</p>
+<figure><img loading="lazy" src="media/library-collections.gif" alt="Filing photographs in a collection by dragging them onto it"><figcaption>Filing photographs in a collection by dragging them onto it</figcaption></figure>
+<p>Collections nest. Drag one onto another to put it inside; <code>+</code> with a
+collection selected — or <code>New collection inside</code> from its menu — makes a
+child. A parent shows everything its children hold, and its count says so.
+Right-click a row (hold it, on a tablet) for the menu: rename, nest,
+move back to the top level, keep it offline, delete.</p>
+<figure><img loading="lazy" src="media/library-nesting.gif" alt="Making Trips, nesting Alps and New York inside it, and opening the parent"><figcaption>Making Trips, nesting Alps and New York inside it, and opening the parent</figcaption></figure>
+<figure><img loading="lazy" src="media/library-nesting.png" alt="Trips showing both of its children's photographs"><figcaption>Trips showing both of its children's photographs</figcaption></figure>
+<figure><img loading="lazy" src="media/library-collection-menu.png" alt="The menu on a collection"><figcaption>The menu on a collection</figcaption></figure>
+<h3 id="bursts">Bursts</h3>
+<p>Frames taken a moment apart that look alike fold into one cell, with a badge
+counting them. Click the badge — tap it, on a tablet — to open the burst in
+the grid, and again to fold it back. A folded burst shows its earliest frame;
+to have it show another, open it and click the ring on the frame you want.</p>
+<p>With faces indexed, narrowing the grid to a person puts <code>Eyes open</code> beside
+their name on the filter bar, which leaves out the frames where they blinked.</p>
+<h3 id="duplicate-originals">Duplicate originals</h3>
+<p>A library put together by hand often holds the same RAW more than once — a
+dated folder, a <code>bck</code> folder beside it, a renamed copy another program
+exported. When the catalog finds files with the same camera, capture time
+and size in more than one place, <code>Duplicate originals</code> appears under the
+trash in the sidebar with how many there are; Settings offers the same page
+beside the other whole-library passes.</p>
+<p>The page lists each group with its picture and its paths. One copy is
+marked <code>Stays</code>: the one outside a folder named like a backup (<code>bck</code>,
+<code>backup</code>, <code>copy</code>, <code>old</code>…), then the one still named the way the camera named
+it (<code>_MG_4623</code>, <code>IMG_0001</code>, <code>DSC_0042</code>), then the one catalogued first. Tap
+another path to keep that copy instead, and untick <code>Include</code> to leave a
+group alone.</p>
+<p>Nothing moves until <code>Check</code> has read the first and last megabyte of every
+copy and each copy's sidecar. A group whose files differ, or whose copies
+carry two different edits, is marked skipped and says why — two edits of
+one frame are kept for virtual copies. What the check reads is kept, so a
+second visit costs nothing. The line under each group says what the copy
+that stays will gain: the highest rating, every keyword and collection, a
+flag or label the copies agree on, faces and the names on them, and the
+edit if only a copy had one. Where the copies disagree on a flag or a
+label, the one that stays keeps its own and the line says so.</p>
+<figure><img loading="lazy" src="media/duplicates.png" alt="Two frames with a copy each in a bck folder, checked and proved the same, the copy that stays marked on each"><figcaption>Two frames with a copy each in a bck folder, checked and proved the same, the copy that stays marked on each</figcaption></figure>
+<p><code>Move N copies to trash</code> does it, one group at a time: each group is merged
+and its spare copies moved together, or not at all. The copies go to the
+trash, not away — select them in the trash view and <code>Restore</code> puts them
+back where they were.</p>
+<h2 id="developing-a-photograph">Developing a photograph</h2>
+<p>Click a thumbnail to open it. The column on the right is every adjustment;
+the strip at its head narrows it to one group.</p>
+<figure><img loading="lazy" src="media/develop.png" alt="The develop view: the photograph, the histogram, and the adjustment column"><figcaption>The develop view: the photograph, the histogram, and the adjustment column</figcaption></figure>
+<figure><img loading="lazy" src="media/develop-groups.gif" alt="Switching between the Optics, Light, Colour, Effects and Detail groups"><figcaption>Switching between the Optics, Light, Colour, Effects and Detail groups</figcaption></figure>
+<h3 id="light">Light</h3>
+<p>Exposure, contrast, highlights, shadows, blacks, whites and a tone curve.
+Hold <code>Before</code> to see the photograph as it was.</p>
+<figure><img loading="lazy" src="media/develop-light.gif" alt="Raising exposure, pulling the highlights, lifting the shadows, then holding Before"><figcaption>Raising exposure, pulling the highlights, lifting the shadows, then holding Before</figcaption></figure>
+<h3 id="looking-closer">Looking closer</h3>
+<p>Double-click for 1:1; drag to move about; double-click again to fit. The
+wheel zooms to any amount in between. Past 1:1 the file's own pixels are
+drawn as hard-edged blocks rather than smoothed, so what you see is what
+the sensor recorded.</p>
+<figure><img loading="lazy" src="media/develop-zoom.gif" alt="Zooming to 1:1 with a double-click, panning, then further in with the wheel"><figcaption>Zooming to 1:1 with a double-click, panning, then further in with the wheel</figcaption></figure>
+<h3 id="moving-between-photographs">Moving between photographs</h3>
+<p>The roll along the foot of the canvas holds the photographs the grid was
+showing; click one to open it. The right arrow, <code>D</code> or space opens the next,
+and the left arrow or <code>A</code> the one before; held down, they go on past the
+stretch the roll has loaded, through everything the grid would show. The
+edit on screen is saved on the way, so stepping along a shoot loses nothing.</p>
+<h3 id="white-balance-from-the-photograph">White balance from the photograph</h3>
+<p>Press <code>pick</code> in the White Balance group, then click something neutral —
+a white wall, a grey card, the air conditioner here. The picker sets the
+sliders from the photograph, not from where they were: below, the frame
+is dragged cold first and one click puts it right. A blown highlight is
+refused, since a clipped pixel has no colour left to balance.</p>
+<figure><img loading="lazy" src="media/develop-wb.gif" alt="Cooling the frame with the slider, then picking a white air conditioner to set the white balance"><figcaption>Cooling the frame with the slider, then picking a white air conditioner to set the white balance</figcaption></figure>
+<h3 id="composing">Composing</h3>
+<p>Crop by dragging the frame's corners, straighten with the slider, lock a
+ratio from the chips. <code>Done composing</code> returns to the photograph.</p>
+<figure><img loading="lazy" src="media/compose.gif" alt="Cropping, straightening and choosing a ratio"><figcaption>Cropping, straightening and choosing a ratio</figcaption></figure>
+<p><code>Vertical</code> and <code>Horizontal</code>, under <code>Straighten</code>, correct the lines that
+converge when the camera is tilted up at a building; the crop refits to
+what is left.</p>
+<figure><img loading="lazy" src="media/compose-perspective.gif" alt="Standing towers shot from below upright with the Vertical slider"><figcaption>Standing towers shot from below upright with the Vertical slider</figcaption></figure>
+<p>A crop that leaves a mask wholly outside the frame says so, and offers to
+take the crop back or keep it.</p>
+<figure><img loading="lazy" src="media/crop-orphan.gif" alt="A stroke in a corner, a crop that leaves it outside, and the notice with Undo crop and Keep crop"><figcaption>A stroke in a corner, a crop that leaves it outside, and the notice with Undo crop and Keep crop</figcaption></figure>
+<h3 id="local-adjustments">Local adjustments</h3>
+<p><code>Local</code> in the rail turns the column into a mask stack. <code>Find subjects</code> runs a
+segmentation model over the photograph; what it recognises appears as a list
+of categories with how much of the frame each covers. Click one and it is a
+mask — then every slider below edits only that region.</p>
+<figure><img loading="lazy" src="media/local-categories.png" alt="What the model found in an urban scene: ground, architecture, sky, vegetation"><figcaption>What the model found in an urban scene: ground, architecture, sky, vegetation</figcaption></figure>
+<figure><img loading="lazy" src="media/local-segment.png" alt="The sky chosen: tinted on the photograph, and the column now scoped to it"><figcaption>The sky chosen: tinted on the photograph, and the column now scoped to it</figcaption></figure>
+<p>A mask is a stack of parts. Paint into it, subtract a gradient from it, grow
+or shrink its edge, choose how it falls off. <code>Show masks as</code> draws the mask
+tinted, as alpha, or as an outline; the eye on its row switches it off.</p>
+<figure><img loading="lazy" src="media/local-paint.gif" alt="Looking at the mask three ways, painting into it, then growing its edge"><figcaption>Looking at the mask three ways, painting into it, then growing its edge</figcaption></figure>
+<p>Linear and radial gradients, a tone range and a colour range are the other
+ways to make one; each can be combined with any other. <code>∩ Intersect</code> keeps
+only where the new part and the mask agree: below, a gradient over the
+lower half, then two strokes that survive only where the gradient is.</p>
+<figure><img loading="lazy" src="media/local-intersect.gif" alt="A linear gradient, then Intersect and two painted strokes"><figcaption>A linear gradient, then Intersect and two painted strokes</figcaption></figure>
+<h3 id="repair">Repair</h3>
+<p><code>Repair</code> in the rail: click a mark and it is covered from a source DarkRoom
+chooses beside it. Drag either circle to move it; the size, feather and
+opacity are in the panel. Heal blends; clone copies.</p>
+<figure><img loading="lazy" src="media/repair.gif" alt="Covering marks on a road"><figcaption>Covering marks on a road</figcaption></figure>
+<h3 id="film">Film</h3>
+<p>The <code>Film</code> chooser at the head of Adjust applies a spectral simulation of a
+named stock; below it, the print exposure and push controls a film has and a
+sensor does not. The list opens over the column and scrolls on its own — by
+wheel, drag or flick, or with <code>Up</code>, <code>Down</code> and <code>Enter</code> — down to the
+black-and-white stocks at its end.</p>
+<figure><img loading="lazy" src="media/film.gif" alt="Opening the film list, scrolling it, choosing Velvia, then holding Before"><figcaption>Opening the film list, scrolling it, choosing Velvia, then holding Before</figcaption></figure>
+<h3 id="history-snapshots-presets">History, snapshots, presets</h3>
+<p>Every change is a step; <code>Undo</code> and the History panel walk them. <code>Snapshot</code>
+keeps the current state under a name. <code>Presets…</code> saves the settings to
+apply elsewhere, and imports <code>.xmp</code> from other applications.</p>
+<figure><img loading="lazy" src="media/presets.png" alt="The presets sheet"><figcaption>The presets sheet</figcaption></figure>
+<h3 id="copying-settings">Copying settings</h3>
+<p><code>Copy</code> in the top bar, or Ctrl+C, takes this photograph's settings; <code>Paste</code>,
+or Ctrl+V, puts them on another, and says what it would paste — how many
+adjustments, and whether the crop comes too. In the grid, <code>Paste to N</code> on the
+selection bar pastes onto every photograph selected. Which kinds of edit a
+copy carries is chosen in <code>Presets…</code>, or with Ctrl+Shift+C: a look carried
+across a shoot usually leaves each frame's own crop alone.</p>
+<h2 id="merging-a-panorama">Merging a panorama</h2>
+<p>Select the frames, then <code>Merge to panorama</code> from the selection bar. The
+frames are read, aligned, and drawn on the suggested projection with each one
+outlined where it landed — twelve hand-held portrait frames across an alpine
+valley, here. Change the projection (a 150° sweep on a flat perspective is
+what the middle of the film shows, and why cylindrical is suggested), ask for
+the border to be filled rather than cropped, then <code>Merge</code>. The composite is
+written beside its sources as a DNG and appears in the grid with the merge
+as the first step in its history.</p>
+<figure><img loading="lazy" src="media/panorama.gif" alt="Twelve frames aligned, the projections tried, and the border filled"><figcaption>Twelve frames aligned, the projections tried, and the border filled</figcaption></figure>
+<figure><img loading="lazy" src="media/panorama-aligned.png" alt="The alignment on a cylinder, each frame outlined where it landed"><figcaption>The alignment on a cylinder, each frame outlined where it landed</figcaption></figure>
+<figure><img loading="lazy" src="media/panorama-filled.png" alt="The same, with the ragged border filled by the model rather than cropped away"><figcaption>The same, with the ragged border filled by the model rather than cropped away</figcaption></figure>
+<h2 id="export">Export</h2>
+<p><code>Export</code> in the develop header, or <code>Export N</code> from a selection. Format,
+size, colour space, sharpening, naming and where the file goes are in
+Settings, and apply to every export until changed. An export with no folder
+set is refused, and the header says so.</p>
+<figure><img loading="lazy" src="media/settings-export.png" alt="Export defaults in Settings"><figcaption>Export defaults in Settings</figcaption></figure>
+<h2 id="settings">Settings</h2>
+<figure><img loading="lazy" src="media/settings.png" alt="The settings page: background activity, indexing, storage, display"><figcaption>The settings page: background activity, indexing, storage, display</figcaption></figure>
+<p>Background activity with progress, thumbnail and face indexing, how many
+duplicate originals the library holds and the way to their review, storage
+on this device, display and colour, export defaults, what is written to XMP
+sidecars, the manual, and a diagnostics bundle for a bug report.</p>
+<h2 id="people">People</h2>
+<p>Face detection and identity run over the library and group faces by person;
+the <code>Identity</code> page is where suggestions are confirmed, rejected and split,
+and <code>People</code> on the filter bar narrows the grid to someone. Not pictured
+here, for the obvious reason — <a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/faces.md">faces.md</a> has the design.</p>
+<h2 id="where-things-are-written-down">Where things are written down</h2>
+<table><thead><tr><th>Feature</th><th>Design</th></tr></thead><tbody>
+<tr><td>Local masks and segmentation</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/segmentation.md">segmentation.md</a>, <a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/mask-editing.md">mask-editing.md</a></td></tr>
+<tr><td>Repair</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/spot-removal.md">spot-removal.md</a></td></tr>
+<tr><td>Panorama</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/panorama.md">panorama.md</a></td></tr>
+<tr><td>Faces and identity</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/faces.md">faces.md</a></td></tr>
+<tr><td>Gestures, generated from the code</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/gestures.md">gestures.md</a></td></tr>
+<tr><td>Navigation and layout</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/ui-navigation.md">ui-navigation.md</a></td></tr>
+<tr><td>Sync and storage</td><td><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/docs/dev/storage.md">storage.md</a></td></tr>
+</tbody></table>
+<h2 id="how-this-page-is-made">How this page is made</h2>
+<p><a href="https://gitea.tourolle.paris/dtourolle/DarkRoom/src/branch/master/tools/manual/README.md"><code>tools/manual/</code></a> drives the desktop build on
+a private X server and records each scene; <code>record.sh &lt;library&gt;</code> re-makes
+every picture here. Run it after a change to the interface and commit what
+changed. The pictures are in LFS.</p>
+<p>The application carries this page. <code>cargo run -p traceability -- manual</code>
+renders it to <code>index.html</code> beside it, which the packages install with the
+pictures and the app opens from Help, from Settings, and from the "See it"
+link beside a gesture on the help sheet. CI fails when the two differ.</p>
+<p>Making it the first time turned up nine faults, each fixed in its own
+commit before the pictures were taken: the folder picker could not choose
+the top level, month headings overprinted each other, a category mask
+widened the column off the window, the mask tint outlived its mode, the
+first sync uploaded an empty thumbnail shard, a merge that ran out of GPU
+memory left the page on Stop for ever, the export settings promised to ask
+for a folder and did not, the keyword sheet sent its keys to the grid, and
+an empty trash told you to check your library folder.</p>
+</main>
+</div>
+</body>
+</html>