fix(anki): respect configured fields for word-card enrichment

- Keep dedicated Lapis/Kiku sentence-card fields unchanged

.agents/skills/subminer-release/SKILL.md

@@ -0,0 +1,34 @@
+---
+name: subminer-release
+description: Prepare, cut, publish, or repair SubMiner stable and prerelease releases. Use for hands-on release work; do not use for general release questions.
+---
+
+# SubMiner release
+
+Carry out the requested release phase using the repository's current release process.
+
+## Source of truth
+
+Read `docs/RELEASING.md` completely before changing files or release state. Treat it as canonical. Read `changes/README.md` when the work touches change fragments or generated release notes.
+
+Do not copy release commands or policy into this skill. If this skill disagrees with the release guide, follow the guide and reconcile the skill before handoff.
+
+## Workflow
+
+1. Identify whether the request is for a stable release, prerelease, release preparation, publication, or repair.
+2. Inspect the current branch, worktree status, package version, pending change fragments, relevant tags, and latest CI state before making changes.
+3. Follow the matching procedure in `docs/RELEASING.md` in order. Review generated changelog and release-note Markdown before it can be committed or published.
+4. Run every required gate for the requested release phase. Do not treat a cheaper test lane as a substitute for the documented release gate.
+5. Before a stable tag, confirm the package and tag versions match and no pending `changes/*.md` fragments remain. Preserve fragments for prereleases as documented.
+6. Report the resulting version, completed checks, local commit and tag state, remote publication state, skipped platform checks, and any remaining manual work.
+
+## Authorization boundaries
+
+- A request to prepare a release stops before commit, tag, push, or remote publication unless the user also authorizes those actions.
+- A clear request to cut or publish a release includes the documented commit, tag, and push steps. Ask before the first remote mutation when the wording is ambiguous.
+- Do not edit an existing GitHub release, publish to the AUR, change secrets, or alter signing configuration unless the user explicitly requests that operation.
+- Do not switch branches without consent.
+
+## Stop conditions
+
+Stop and report the blocker when required CI or a release gate fails, authentication is missing, versions disagree, required artifacts are absent, or the worktree contains unexpected changes that overlap the release. Do not tag or publish a partially verified release.

.github/workflows/docs-pages.yml

@@ -32,9 +32,11 @@ jobs:
       - name: Guard stable docs tag shape
         id: tag_guard
         if: github.ref_type == 'tag'
+        env:
+          TAG_NAME: ${{ github.ref_name }}
         run: |
-          if [[ ! "${{ github.ref_name }}" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
-            echo "::notice::Skipping non-stable docs tag ${{ github.ref_name }}"
+          if [[ ! "$TAG_NAME" =~ ^v[0-9]+\.[0-9]+\.[0-9]+$ ]]; then
+            echo "::notice::Skipping non-stable docs tag $TAG_NAME"
             echo "stable_tag=false" >> "$GITHUB_OUTPUT"
             exit 0
           fi

.github/workflows/prerelease.yml

@@ -274,7 +274,8 @@ jobs:
             config.example.jsonc \
             plugin/subminer \
             plugin/subminer.conf \
-            assets/themes/subminer.rasi
+            assets/themes/subminer.rasi \
+            assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer
 
       - name: Generate checksums
         run: |
@@ -296,15 +297,22 @@ jobs:
         run: echo "VERSION=${GITHUB_REF#refs/tags/}" >> "$GITHUB_OUTPUT"
 
       - name: Verify committed prerelease notes
+        env:
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           if [ ! -s release/prerelease-notes.md ]; then
             echo "::error::release/prerelease-notes.md is missing or empty. Run 'bun run changelog:prerelease-notes --version <version>' locally and commit the file before tagging."
             exit 1
           fi
+          if ! bun run changelog:check-prerelease-notes --version "$RELEASE_VERSION"; then
+            echo "::error::release/prerelease-notes.md was not generated for $RELEASE_VERSION. Rerun 'bun run changelog:prerelease-notes --version $RELEASE_VERSION' locally, commit, and retag."
+            exit 1
+          fi
 
       - name: Publish Prerelease
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           set -euo pipefail
 
@@ -326,27 +334,27 @@ jobs:
             exit 1
           fi
 
-          if gh release view "${{ steps.version.outputs.VERSION }}" >/dev/null 2>&1; then
-            gh release edit "${{ steps.version.outputs.VERSION }}" \
+          if gh release view "$RELEASE_VERSION" >/dev/null 2>&1; then
+            gh release edit "$RELEASE_VERSION" \
               --draft \
               --prerelease \
-              --title "${{ steps.version.outputs.VERSION }}" \
+              --title "$RELEASE_VERSION" \
               --notes-file release/prerelease-notes.md
           else
-            gh release create "${{ steps.version.outputs.VERSION }}" \
+            gh release create "$RELEASE_VERSION" \
               --draft \
               --latest=false \
               --prerelease \
-              --title "${{ steps.version.outputs.VERSION }}" \
+              --title "$RELEASE_VERSION" \
               --notes-file release/prerelease-notes.md
           fi
 
           for asset in "${artifacts[@]}"; do
-            gh release upload "${{ steps.version.outputs.VERSION }}" "$asset" --clobber
+            gh release upload "$RELEASE_VERSION" "$asset" --clobber
           done
 
-          gh release edit "${{ steps.version.outputs.VERSION }}" \
+          gh release edit "$RELEASE_VERSION" \
             --draft=false \
             --prerelease \
-            --title "${{ steps.version.outputs.VERSION }}" \
+            --title "$RELEASE_VERSION" \
             --notes-file release/prerelease-notes.md

.github/workflows/release.yml

@@ -273,7 +273,8 @@ jobs:
             config.example.jsonc \
             plugin/subminer \
             plugin/subminer.conf \
-            assets/themes/subminer.rasi
+            assets/themes/subminer.rasi \
+            assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer
 
       - name: Generate checksums
         run: |
@@ -295,33 +296,40 @@ jobs:
         run: echo "VERSION=${GITHUB_REF#refs/tags/}" >> "$GITHUB_OUTPUT"
 
       - name: Guard against pending changelog fragments
+        env:
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           if find changes -maxdepth 1 -name '*.md' -not -name README.md -print -quit | grep -q .; then
-            echo "::error::Pending changelog fragments detected. Run 'bun run changelog:build --version ${{ steps.version.outputs.VERSION }}' locally and commit the polished CHANGELOG.md before tagging. CI no longer auto-builds the changelog because the polish step requires the local 'claude' CLI."
+            echo "::error::Pending changelog fragments detected. Run 'bun run changelog:build --version $RELEASE_VERSION' locally and commit the polished CHANGELOG.md before tagging. CI no longer auto-builds the changelog because the polish step requires the local 'claude' CLI."
             exit 1
           fi
 
       - name: Verify changelog is ready for tagged release
-        run: bun run changelog:check --version "${{ steps.version.outputs.VERSION }}"
+        env:
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
+        run: bun run changelog:check --version "$RELEASE_VERSION"
 
       - name: Generate release notes from changelog
-        run: bun run changelog:release-notes --version "${{ steps.version.outputs.VERSION }}"
+        env:
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
+        run: bun run changelog:release-notes --version "$RELEASE_VERSION"
 
       - name: Publish Release
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           set -euo pipefail
 
-          if gh release view "${{ steps.version.outputs.VERSION }}" >/dev/null 2>&1; then
+          if gh release view "$RELEASE_VERSION" >/dev/null 2>&1; then
             # Do not pass the prerelease flag here; gh defaults to a normal release.
-            gh release edit "${{ steps.version.outputs.VERSION }}" \
+            gh release edit "$RELEASE_VERSION" \
               --draft=false \
-              --title "${{ steps.version.outputs.VERSION }}" \
+              --title "$RELEASE_VERSION" \
               --notes-file release/release-notes.md
           else
-            gh release create "${{ steps.version.outputs.VERSION }}" \
-              --title "${{ steps.version.outputs.VERSION }}" \
+            gh release create "$RELEASE_VERSION" \
+              --title "$RELEASE_VERSION" \
               --notes-file release/release-notes.md
           fi
 
@@ -344,7 +352,7 @@ jobs:
           fi
 
           for asset in "${artifacts[@]}"; do
-            gh release upload "${{ steps.version.outputs.VERSION }}" "$asset" --clobber
+            gh release upload "$RELEASE_VERSION" "$asset" --clobber
           done
 
   aur-publish:
@@ -420,9 +428,10 @@ jobs:
         if: steps.aur_prereqs.outputs.skip != 'true' && steps.aur_ssh.outputs.skip != 'true' && steps.aur_clone.outputs.skip != 'true'
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           set -euo pipefail
-          version="${{ steps.version.outputs.VERSION }}"
+          version="$RELEASE_VERSION"
           install -dm755 .tmp/aur-release-assets
           gh release download "$version" \
             --dir .tmp/aur-release-assets \
@@ -432,15 +441,17 @@ jobs:
 
       - name: Update AUR packaging metadata
         if: steps.aur_prereqs.outputs.skip != 'true' && steps.aur_ssh.outputs.skip != 'true' && steps.aur_clone.outputs.skip != 'true'
+        env:
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           set -euo pipefail
-          version_no_v="${{ steps.version.outputs.VERSION }}"
+          version_no_v="$RELEASE_VERSION"
           version_no_v="${version_no_v#v}"
           cp packaging/aur/subminer-bin/PKGBUILD aur-subminer-bin/PKGBUILD
           cp packaging/aur/subminer-bin/.SRCINFO aur-subminer-bin/.SRCINFO
           bash scripts/update-aur-package.sh \
             --pkg-dir aur-subminer-bin \
-            --version "${{ steps.version.outputs.VERSION }}" \
+            --version "$RELEASE_VERSION" \
             --appimage ".tmp/aur-release-assets/SubMiner-${version_no_v}.AppImage" \
             --wrapper ".tmp/aur-release-assets/subminer" \
             --assets ".tmp/aur-release-assets/subminer-assets.tar.gz"
@@ -450,6 +461,7 @@ jobs:
         working-directory: aur-subminer-bin
         env:
           GIT_SSH_COMMAND: ssh -i ~/.ssh/aur -o IdentitiesOnly=yes
+          RELEASE_VERSION: ${{ steps.version.outputs.VERSION }}
         run: |
           set -euo pipefail
           if git diff --quiet -- PKGBUILD .SRCINFO; then
@@ -459,7 +471,7 @@ jobs:
           git config user.name "github-actions[bot]"
           git config user.email "41898282+github-actions[bot]@users.noreply.github.com"
           git add PKGBUILD .SRCINFO
-          git commit -m "Update to ${{ steps.version.outputs.VERSION }}"
+          git commit -m "Update to $RELEASE_VERSION"
 
           attempts=3
           for attempt in $(seq 1 "$attempts"); do

.gitignore

@@ -49,6 +49,7 @@ tests/*
 !.agents/skills/
 .agents/skills/*
 !.agents/skills/subminer-change-verification/
+!.agents/skills/subminer-release/
 !.agents/skills/subminer-scrum-master/
 .agents/skills/subminer-change-verification/*
 !.agents/skills/subminer-change-verification/SKILL.md
@@ -56,6 +57,8 @@ tests/*
 .agents/skills/subminer-change-verification/scripts/*
 !.agents/skills/subminer-change-verification/scripts/classify_subminer_diff.sh
 !.agents/skills/subminer-change-verification/scripts/verify_subminer_change.sh
+.agents/skills/subminer-release/*
+!.agents/skills/subminer-release/SKILL.md
 .agents/skills/subminer-scrum-master/*
 !.agents/skills/subminer-scrum-master/SKILL.md
 favicon.png

CHANGELOG.md

@@ -1,5 +1,46 @@
 # Changelog
 
+## v0.19.4 (2026-08-25)
+
+### Added
+- **Library Merge & Move**: Duplicate library cards for the same show can now be combined. Select cards in the library grid and use "Merge Selected" to pick which entry to keep and move every episode onto it, preserving sessions, mined cards, and watch time. Episodes can also be reassigned individually via the "→" button, useful when a file lands under a stray title; manual assignments survive later filename parsing, Jellyfin refreshes, and season repair. Exact AniList title matches with compatible seasons now merge automatically, while fuzzy matches surface as dismissible "Possible duplicate" reviews instead of merging silently.
+- **Duplicate Line Cleanup Tool**: The Vocabulary tab's new "Duplicates" button scans a chosen time window (7 days through all time) for old karaoke/typeset duplicate-line bursts, shows what it found, and collapses each run to one line once confirmed; `subminer stats cleanup --duplicate-lines` does the same from the terminal, with `--dry-run` and `--lookback-days <n>` options. Watch time and lines-seen totals are left unchanged.
+
+### Changed
+- **Prerelease Release Notes**: Prerelease notes now open with a "Changes since" section listing only what changed versus the previous beta/RC of the same version, above the cumulative highlights, and CI rejects prerelease tags whose committed notes were generated for a different beta/RC.
+
+### Fixed
+- **Subtitle & Karaoke Duplication**:
+  - Karaoke and animated signs are reconstructed once from their authored text and shown only while actually sung, with original word spacing preserved, instead of flooding the overlay, subtitle sidebar, immersion history, mining, or stats with glyph fragments, per-frame color phases, and repeated animation events.
+  - Decorative layers (highlight sweeps, glow/shadow copies, symbol-font decoration, particle swarms, hidden or zero-scaled text) stay out of published text, while ordinary repeated dialogue, positioned signs, wrapped lyric rows, and multi-row CC-style blocks still display correctly.
+  - Embedded subtitle tracks on network-mounted (SMB/NFS) media are extracted and parsed again instead of falling back to live-text-only, restoring karaoke reconstruction, sidebar cues, and mining for releases that only ship subtitles inside the container.
+  - Secondary subtitles go through the same deduplication pipeline as primary subtitles and no longer clip display after about four lines.
+  - Event-heavy karaoke files that previously stalled subtitle loading for several seconds now parse in well under a second.
+- **Character Dictionary Reliability**:
+  - Generation, merged rebuilds, and imports no longer freeze the app on large dictionaries; snapshot I/O, archive building, and image/name lookup caches moved off the UI's critical path.
+  - Dictionaries are reused instead of regenerated when MeCab finds no name splits.
+  - Cached portraits restore correctly after the portrait index finishes loading post-tokenization.
+  - Desktop progress notifications on Linux AppImage installs update in place instead of flickering, fixing a bug where the AppImage's bundled libraries broke the system notification helper.
+- **Overlay Startup & Modals**:
+  - The macOS window-tracking helper targets macOS 12.0+ instead of requiring the build machine's exact macOS version, fixing crashes on older systems like Ventura that left the overlay stuck on "Overlay loading".
+  - mpv IPC connection attempts time out and retry, showing an actionable error if content still isn't ready after 30 seconds.
+  - Dedicated overlay modals are prewarmed on macOS and Windows so shortcuts open them promptly.
+  - On macOS, reused modals and the stats window open above fullscreen mpv on its current Space instead of jumping to another desktop.
+- **Wayland File Drop**: Fixed native Wayland drag-and-drop from file managers such as Thunar, so subtitle and video files dropped on the visible overlay are resolved and forwarded to mpv.
+- **Windows Mouse Lag**: Fixed system-wide mouse lag on Windows while SubMiner is running, caused by the overlay's global mouse hook for click-through forwarding and by the mpv window tracker blocking the app on repeated PowerShell lookups.
+- **Sentence Mining Audio & Clips**: Sentence-audio generation no longer times out on slow network-mounted media with many subtitle/font streams (bounded FFmpeg probing, two-minute extraction budget, clearer error reporting), and mined audio/animated AVIF clips now capture the subtitle line that was actually mined by snapshotting the clip range at lookup time instead of reading live mpv state later.
+- **Stats Performance & Reliability**: Immersion stats storage now sets its SQLite busy timeout before WAL setup, avoiding transient lock errors under concurrent writes. Deletes in the stats dashboard no longer freeze the UI, run proportional to what's deleted instead of rebuilding full lifetime summaries, retry safely if the delete worker crashes, and no longer rescan the whole library when deleting very common words; a new index also makes large session deletes drop from minutes to milliseconds. Library merges, video moves, and AniList reassignments got the same lifetime-summary fix.
+- **Vocabulary Stats Accuracy**: Vocabulary totals and charts now count all tracked vocabulary instead of only the first page, new-word history uses corrected daily rollups (fixing legacy timestamp and time-zone issues), summary cards refresh automatically after edits to the exclusion list, and rapid exclusion edits no longer race each other.
+- **Rofi MKV Thumbnails**: Fixed missing MKV thumbnails in the Linux rofi picker when system thumbnailer registrations only advertise legacy Matroska MIME aliases.
+
+<details>
+<summary>Internal changes</summary>
+
+### Internal
+- Docs Site Indexing: Excluded the `/main/` and `/v/<version>/` docs trees from search indexing (self-referential canonical, `noindex,follow`, matching `X-Robots-Tag`) so crawlers focus on current docs instead of ~30 archived copies of every page, and restored `<lastmod>` dates in the docs sitemap that were silently dropped by production builds.
+
+</details>
+
 ## v0.19.3 (2026-08-13)
 
 ### Added

Makefile

@@ -2,8 +2,10 @@
 
 APP_NAME := subminer
 THEME_SOURCE := assets/themes/subminer.rasi
+THUMBNAILER_SOURCE := assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer
 LAUNCHER_OUT := dist/launcher/$(APP_NAME)
 THEME_FILE := subminer.rasi
+THUMBNAILER_FILE := subminer-ffmpegthumbnailer.thumbnailer
 
 # Default install prefix for the wrapper script.
 PREFIX ?= $(HOME)/.local
@@ -221,11 +223,13 @@ docs-dev: ensure-bun
 
 
 install-linux: build-launcher
-	@printf '%s\n' "[INFO] Installing Linux wrapper/theme artifacts"
+	@printf '%s\n' "[INFO] Installing Linux wrapper/support artifacts"
 	@install -d "$(BINDIR)"
 	@install -m 0755 "$(LAUNCHER_OUT)" "$(BINDIR)/$(APP_NAME)"
 	@install -d "$(LINUX_DATA_DIR)/themes"
 	@install -m 0644 "./$(THEME_SOURCE)" "$(LINUX_DATA_DIR)/themes/$(THEME_FILE)"
+	@install -d "$(LINUX_DATA_DIR)/thumbnailers"
+	@install -m 0644 "./$(THUMBNAILER_SOURCE)" "$(LINUX_DATA_DIR)/thumbnailers/$(THUMBNAILER_FILE)"
 	@install -d "$(LINUX_DATA_DIR)/plugin/subminer"
 	@cp -R ./plugin/subminer/. "$(LINUX_DATA_DIR)/plugin/subminer/"
 	@if [ -n "$(APPIMAGE_SRC)" ]; then \
@@ -234,7 +238,7 @@ install-linux: build-launcher
 		printf '%s\n' "[WARN] No release/SubMiner-*.AppImage found; skipping AppImage install"; \
 		printf '%s\n' "       Build one with: make build"; \
 	fi
-	@printf '%s\n' "Installed to:" "  $(BINDIR)/subminer" "  $(LINUX_DATA_DIR)/themes/$(THEME_FILE)"
+	@printf '%s\n' "Installed to:" "  $(BINDIR)/subminer" "  $(LINUX_DATA_DIR)/themes/$(THEME_FILE)" "  $(LINUX_DATA_DIR)/thumbnailers/$(THUMBNAILER_FILE)"
 
 install-macos: build-launcher
 	@printf '%s\n' "[INFO] Installing macOS wrapper/theme/app artifacts"
@@ -275,8 +279,9 @@ uninstall:
 uninstall-linux:
 	@rm -f "$(BINDIR)/subminer" "$(BINDIR)/SubMiner.AppImage"
 	@rm -f "$(LINUX_DATA_DIR)/themes/$(THEME_FILE)"
+	@rm -f "$(LINUX_DATA_DIR)/thumbnailers/$(THUMBNAILER_FILE)"
 	@rm -rf "$(LINUX_DATA_DIR)/plugin/subminer"
-	@printf '%s\n' "Removed:" "  $(BINDIR)/subminer" "  $(BINDIR)/SubMiner.AppImage" "  $(LINUX_DATA_DIR)/themes/$(THEME_FILE)" "  $(LINUX_DATA_DIR)/plugin/subminer"
+	@printf '%s\n' "Removed:" "  $(BINDIR)/subminer" "  $(BINDIR)/SubMiner.AppImage" "  $(LINUX_DATA_DIR)/themes/$(THEME_FILE)" "  $(LINUX_DATA_DIR)/thumbnailers/$(THUMBNAILER_FILE)" "  $(LINUX_DATA_DIR)/plugin/subminer"
 
 uninstall-macos:
 	@rm -f "$(BINDIR)/subminer"

README.md

@@ -15,7 +15,7 @@ Integrates Yomitan and mpv - on-screen lookups, mine to Anki, and track immersio
 [![License](https://img.shields.io/github/license/ksyasuda/SubMiner?style=flat-square&color=1a1a2e)](https://www.gnu.org/licenses/gpl-3.0)
 [![TypeScript](https://img.shields.io/badge/TypeScript-1a1a2e?style=flat-square&logo=typescript&logoColor=3178c6)](https://www.typescriptlang.org)
 
-[![SubMiner demo](./assets/minecard.webp)](https://github.com/user-attachments/assets/89e61895-e2b7-4b47-8d50-a35afe4132b2)
+[![SubMiner demo](./assets/minecard.webp)](https://github.com/user-attachments/assets/7abab8a9-4e4e-4f06-9f3c-9783e15a3807)
 
 </div>
 

assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer

@@ -0,0 +1,4 @@
+[Thumbnailer Entry]
+TryExec=ffmpegthumbnailer
+Exec=ffmpegthumbnailer -i %i -o %o -s %s -f
+MimeType=video/matroska;video/matroska-3d;video/x-matroska;video/x-matroska-3d;

changes/README.md

@@ -42,13 +42,14 @@ How fragments turn into a release:
 
 - At release time, `bun run changelog:build` (and `bun run changelog:prerelease-notes`) pipes every pending fragment through `claude -p` to merge related items, drop noise, and rewrite into a clean user-facing release body. Write fragments as raw, informative notes — don't worry about polished prose, deduping across PRs, or line-by-line phrasing. The polish step handles all of that.
 - The polish step treats pending fragments as the final release outcome, not prerelease history. If a feature is added and then renamed or fixed before the stable cut, ship the final feature bullet instead of separate prerelease-only breaking/fix entries.
-- GitHub release notes and prerelease notes use short top-level items with nested bullets for the change, user benefit, and any useful action note. The stable `CHANGELOG.md` can stay in compact single-line bullets.
+- `CHANGELOG.md`, GitHub release notes, and prerelease notes all use short top-level items with one nested bullet per distinct change, instead of packing a release's worth of detail into a single paragraph bullet. An item with only one thing to say stays inline on the top-level bullet. Release notes and prerelease notes additionally cover user benefit and any useful action note in their nested bullets.
 - `internal` fragments stay in `CHANGELOG.md` (inside a collapsed `<details>` block) but are dropped from the GitHub release notes entirely.
 - The polished `CHANGELOG.md` and `release/release-notes.md` are committed and reviewed before tagging — edit the Markdown by hand if Claude misses something.
 
 Prerelease notes:
 
 - prerelease tags like `v0.11.3-beta.1` and `v0.11.3-rc.1` reuse the current pending fragments to generate `release/prerelease-notes.md`
+- from the second prerelease of a base version onward, the notes also open with a `## Changes since <previous tag>` section generated from the fragment diff against the previous beta/RC tag; keep fragment edits meaningful. Editorial-only rewording is filtered out of that section, while genuinely changed behavior and deleted fragments (reverted changes) are reported
 - existing prerelease notes are a reviewed baseline; later prerelease runs should replace stale beta/RC wording with the current outcome instead of appending fix churn
 - prerelease note generation does not consume fragments and does not update `CHANGELOG.md` or `docs-site/changelog.md`
 - the final stable release is the point where `bun run changelog:build` consumes fragments into the stable changelog and release notes

changes/anki-overlay-update-progress.md

@@ -0,0 +1,4 @@
+type: fixed
+area: overlay
+
+- Kept the Anki card update spinner visible until audio and image updates finish.

changes/audio-generation-network-mounts.md

@@ -1,4 +0,0 @@
-type: fixed
-area: Anki media
-
-- Fixed sentence-audio generation timing out on slow network-mounted MKV files with many subtitle and font-attachment streams. Selected audio tracks now use bounded FFmpeg probing and a two-minute extraction budget, and missing output reports a clear FFmpeg error instead of raw `ENOENT`.

changes/docs-site-index-hygiene.md

@@ -1,5 +0,0 @@
-type: internal
-area: docs
-
-- Excluded the `/main/` and `/v/<version>/` docs trees from search indexing with a self-referential canonical, `noindex,follow`, and a matching `X-Robots-Tag` header, so crawlers spend their budget on the current docs instead of ~30 archived copies of every page.
-- Restored `<lastmod>` dates in the docs sitemap, which were silently dropped because production builds render from an untracked release snapshot.

changes/duplicate-line-stats-cleanup.md

@@ -1,5 +0,0 @@
-type: fixed
-area: stats
-
-- Typeset subtitles no longer flood the stats. Karaoke openings and animated signs are authored as one subtitle event per animation frame, and immersion tracking counted every frame, which was enough to put an OP lyric at the top of "Top Repeated Words" for good. Lines are now collapsed on the way in using the same rules the subtitle sidebar already applies: matching parsed timings record exactly the cues the sidebar shows, while shifted, changing, or unparsed sources use a strict fallback where identical, contiguous, sub-0.1s lines stop counting after a few frames. Ordinary repeated dialogue and rewatches are unaffected.
-- Added a cleanup for stats already affected. The Vocabulary tab has a **Duplicates** button that scans a chosen window (7 days through all time), shows the bursts it found and the word and kanji counts they added, and collapses each run to one line once confirmed. `subminer stats cleanup --duplicate-lines` does the same from the terminal, with `--dry-run` and `--lookback-days <n>`. Only subtitle lines and the vocabulary counts they feed are touched; watch time and lines-seen totals are left as recorded.

changes/fix-anki-configured-word-fields.md

@@ -0,0 +1,4 @@
+type: fixed
+area: anki
+
+- Word-card enrichment now always writes sentence text and audio to the configured AnkiConnect fields, while dedicated Lapis/Kiku sentence-card and audio-card actions keep their compatible field names.

changes/fix-macos-fullscreen-modal-spaces.md

@@ -1,5 +0,0 @@
-type: fixed
-area: overlay
-
-- Dedicated overlay modals are prewarmed on macOS and Windows so shortcuts open them promptly on the first press. Windows now refreshes the hidden modal renderer between sessions to keep later modals interactive. On macOS, reused modals and the in-app stats window also open above fullscreen mpv on its current Space instead of appearing on another desktop or forcing a Space change.
-- Updated subtitle ASS observation to mpv's current `sub-text/ass` property, removing its deprecation warning.

changes/fix-native-wayland-overlay-file-drop.md

@@ -1,4 +0,0 @@
-type: fixed
-area: overlay
-
-- Fixed native Wayland drag-and-drop from file managers such as Thunar so subtitle and video files dropped on the visible overlay are resolved and forwarded to mpv.

changes/fix-overlapping-subtitle-line-drop.md

@@ -0,0 +1,5 @@
+type: fixed
+area: overlay
+
+- Subtitle lines that start while another line is still on screen now appear alongside it instead of staying hidden until a track switch or seek (#220).
+- Two subtitles shown at the same time now always render on separate lines, stacked by their authored screen position (top signs and song lines above bottom dialogue), so lines no longer merge into one sentence or swap rows mid-display.

changes/fix-windows-system-mouse-lag.md

@@ -1,4 +0,0 @@
-type: fixed
-area: overlay
-
-- Fixed system-wide mouse lag on Windows while SubMiner is running: the overlay no longer installs Electron's global mouse hook for click-through forwarding, and the mpv window tracker no longer blocks the app on repeated PowerShell command-line lookups.

changes/hide-feature-demos.md

@@ -1,4 +0,0 @@
-type: docs
-area: documentation
-
-- Hid the unfinished feature demos page from the documentation sidebar while keeping its direct URL available.

changes/library-merge-and-move.md

@@ -1,6 +0,0 @@
-type: added
-area: stats
-
-- Library: duplicate cards for the same show can now be combined. Press "Select" above the library grid, tick the cards, and use "Merge Selected"; the dialog picks which entry to keep and moves every episode onto it. Sessions, mined cards, and watch time are preserved, the emptied entries disappear, and remembered title aliases keep future episodes on the merged card.
-- Library: episodes can be reassigned to another library entry from the "→" button on an episode row, which is the fix when one file lands under a stray title (e.g. an episode name parsed as the series). Manual assignments now survive later filename parsing, Jellyfin refreshes, and season repair. Local episodes in the same directory reuse a uniquely corrected destination unless they parse to a title that already has its own library entry, while conflicting seasons or manual destinations are not forced together. Emptying an entry this way removes it and returns to the grid.
-- Library: exact AniList title matches with compatible seasons fold duplicate cards automatically. Fuzzy same-AniList matches appear as dismissible "Possible duplicate" reviews instead of changing the library without confirmation; conflicting explicit seasons are left alone.

changes/linux-notification-in-place-replace.md

@@ -1,4 +0,0 @@
-type: fixed
-area: notifications
-
-- Character dictionary progress notifications on Linux now update in place instead of flickering off and reappearing on every status change.

changes/mining-clip-range-snapshot.md

@@ -1,4 +0,0 @@
-type: fixed
-area: anki
-
-- Mined audio and animated AVIF clips now capture the subtitle line that was actually mined. The clip range is snapshotted once at Yomitan lookup time (and reused for both audio and image), instead of each generator reading the live mpv subtitle when it starts — which clipped whatever line was on screen after slow audio extraction finished, producing too-short or misaligned AVIF clips.

changes/stats-delete-fast-incremental.md

@@ -1,9 +0,0 @@
-type: fixed
-area: stats
-
-- Stats deletes no longer freeze the stats dashboard: the delete worker module now resolves when running from source, so deletes actually run off the serving thread instead of silently falling back to it.
-- Deletes now subtract their exact contribution from lifetime summaries instead of rebuilding them from retained sessions, making delete cost proportional to what is deleted and preserving lifetime totals older than the session retention window.
-- If the delete worker crashes, the delete now retries on the current thread instead of failing.
-- Library merges, video moves, AniList reassignments, and `subminer stats cleanup -l` also stopped rebuilding lifetime summaries from retained sessions; they now recompute from per-episode history, so those operations are faster and no longer erase lifetime totals older than the session retention window.
-- Deleting content that contains very common words no longer rescans every occurrence of those words across the whole library; first/last-seen dates are refreshed with index seeks instead.
-- Session deletes on large databases dropped from minutes to milliseconds: an index on the subtitle-line event reference now prevents each deleted session event from scanning the whole subtitle-line table for foreign-key enforcement.

docs-site/anki-integration.md

@@ -136,6 +136,8 @@ SubMiner maps its data to your Anki note fields. Configure these under `ankiConn
 
 Field names are matched against your Anki note type case-insensitively (an exact match wins, then a lowercase comparison). If a configured field does not exist on the note type, SubMiner skips it without error.
 
+These mappings always control normal word-card enrichment, including Yomitan proxy/polling updates and manual clipboard updates. Enabling Lapis or Kiku does not replace the configured word-card sentence and audio fields with `Sentence` and `SentenceAudio`. The dedicated sentence-card and audio-card shortcuts still use those Lapis/Kiku field names.
+
 Two related options live alongside `fields`: `ankiConnect.deck` (target deck; empty falls back as described above) and `ankiConnect.tags` (tags added to mined cards, default `["SubMiner"]`; set `[]` to disable tagging). The `miscInfo` content is controlled by `ankiConnect.metadata.pattern` (default `[SubMiner] %f (%t)`; tokens: `%f` filename, `%F` filename with extension, `%t` timestamp, `%T` timestamp with milliseconds, `<br>` newline).
 
 ### Minimal Config
@@ -233,7 +235,7 @@ Animated AVIF requires an AV1 encoder (`libaom-av1`, `libsvtav1`, or `librav1e`)
 
 When media is available, mined-card overlay and system notifications include the same current-frame thumbnail.
 
-`overwriteAudio` applies to automatic card updates and duplicate-card enrichment. Manual clipboard subtitle updates (`Ctrl/Cmd+C`, then `Ctrl/Cmd+V`) always replace generated sentence audio, while leaving the word audio field unchanged.
+`overwriteAudio` applies to automatic card updates and duplicate-card enrichment. Manual clipboard subtitle updates (`Ctrl/Cmd+C`, then `Ctrl/Cmd+V`) always replace generated sentence audio in `ankiConnect.fields.audio`, even when `overwriteAudio` is disabled.
 
 ## AI Translation
 
@@ -287,6 +289,8 @@ Sentence card creation and audio card marking require a non-empty `ankiConnect.i
 
 Trigger with the mine sentence shortcut (`Ctrl/Cmd+S` by default). The card is created directly via AnkiConnect with the sentence, audio, and image filled in.
 
+The dedicated sentence-card and audio-card shortcuts use the Lapis/Kiku-compatible `Sentence` and `SentenceAudio` fields. This does not affect the configured fields used to enrich normal word cards.
+
 To mine multiple subtitle lines as one sentence card, use `Ctrl/Cmd+Shift+S` followed by a digit (1–9) to select how many recent lines to combine.
 
 ## Word Card Type (Kiku/Lapis)

docs-site/changelog.md

@@ -1,5 +1,46 @@
 # Changelog
 
+## v0.19.4 (2026-08-25)
+
+**Added**
+- **Library Merge & Move**: Duplicate library cards for the same show can now be combined. Select cards in the library grid and use "Merge Selected" to pick which entry to keep and move every episode onto it, preserving sessions, mined cards, and watch time. Episodes can also be reassigned individually via the "→" button, useful when a file lands under a stray title; manual assignments survive later filename parsing, Jellyfin refreshes, and season repair. Exact AniList title matches with compatible seasons now merge automatically, while fuzzy matches surface as dismissible "Possible duplicate" reviews instead of merging silently.
+- **Duplicate Line Cleanup Tool**: The Vocabulary tab's new "Duplicates" button scans a chosen time window (7 days through all time) for old karaoke/typeset duplicate-line bursts, shows what it found, and collapses each run to one line once confirmed; `subminer stats cleanup --duplicate-lines` does the same from the terminal, with `--dry-run` and `--lookback-days <n>` options. Watch time and lines-seen totals are left unchanged.
+
+**Changed**
+- **Prerelease Release Notes**: Prerelease notes now open with a "Changes since" section listing only what changed versus the previous beta/RC of the same version, above the cumulative highlights, and CI rejects prerelease tags whose committed notes were generated for a different beta/RC.
+
+**Fixed**
+- **Subtitle & Karaoke Duplication**:
+  - Karaoke and animated signs are reconstructed once from their authored text and shown only while actually sung, with original word spacing preserved, instead of flooding the overlay, subtitle sidebar, immersion history, mining, or stats with glyph fragments, per-frame color phases, and repeated animation events.
+  - Decorative layers (highlight sweeps, glow/shadow copies, symbol-font decoration, particle swarms, hidden or zero-scaled text) stay out of published text, while ordinary repeated dialogue, positioned signs, wrapped lyric rows, and multi-row CC-style blocks still display correctly.
+  - Embedded subtitle tracks on network-mounted (SMB/NFS) media are extracted and parsed again instead of falling back to live-text-only, restoring karaoke reconstruction, sidebar cues, and mining for releases that only ship subtitles inside the container.
+  - Secondary subtitles go through the same deduplication pipeline as primary subtitles and no longer clip display after about four lines.
+  - Event-heavy karaoke files that previously stalled subtitle loading for several seconds now parse in well under a second.
+- **Character Dictionary Reliability**:
+  - Generation, merged rebuilds, and imports no longer freeze the app on large dictionaries; snapshot I/O, archive building, and image/name lookup caches moved off the UI's critical path.
+  - Dictionaries are reused instead of regenerated when MeCab finds no name splits.
+  - Cached portraits restore correctly after the portrait index finishes loading post-tokenization.
+  - Desktop progress notifications on Linux AppImage installs update in place instead of flickering, fixing a bug where the AppImage's bundled libraries broke the system notification helper.
+- **Overlay Startup & Modals**:
+  - The macOS window-tracking helper targets macOS 12.0+ instead of requiring the build machine's exact macOS version, fixing crashes on older systems like Ventura that left the overlay stuck on "Overlay loading".
+  - mpv IPC connection attempts time out and retry, showing an actionable error if content still isn't ready after 30 seconds.
+  - Dedicated overlay modals are prewarmed on macOS and Windows so shortcuts open them promptly.
+  - On macOS, reused modals and the stats window open above fullscreen mpv on its current Space instead of jumping to another desktop.
+- **Wayland File Drop**: Fixed native Wayland drag-and-drop from file managers such as Thunar, so subtitle and video files dropped on the visible overlay are resolved and forwarded to mpv.
+- **Windows Mouse Lag**: Fixed system-wide mouse lag on Windows while SubMiner is running, caused by the overlay's global mouse hook for click-through forwarding and by the mpv window tracker blocking the app on repeated PowerShell lookups.
+- **Sentence Mining Audio & Clips**: Sentence-audio generation no longer times out on slow network-mounted media with many subtitle/font streams (bounded FFmpeg probing, two-minute extraction budget, clearer error reporting), and mined audio/animated AVIF clips now capture the subtitle line that was actually mined by snapshotting the clip range at lookup time instead of reading live mpv state later.
+- **Stats Performance & Reliability**: Immersion stats storage now sets its SQLite busy timeout before WAL setup, avoiding transient lock errors under concurrent writes. Deletes in the stats dashboard no longer freeze the UI, run proportional to what's deleted instead of rebuilding full lifetime summaries, retry safely if the delete worker crashes, and no longer rescan the whole library when deleting very common words; a new index also makes large session deletes drop from minutes to milliseconds. Library merges, video moves, and AniList reassignments got the same lifetime-summary fix.
+- **Vocabulary Stats Accuracy**: Vocabulary totals and charts now count all tracked vocabulary instead of only the first page, new-word history uses corrected daily rollups (fixing legacy timestamp and time-zone issues), summary cards refresh automatically after edits to the exclusion list, and rapid exclusion edits no longer race each other.
+- **Rofi MKV Thumbnails**: Fixed missing MKV thumbnails in the Linux rofi picker when system thumbnailer registrations only advertise legacy Matroska MIME aliases.
+
+<details>
+<summary>Internal changes</summary>
+
+**Internal**
+- Docs Site Indexing: Excluded the `/main/` and `/v/<version>/` docs trees from search indexing (self-referential canonical, `noindex,follow`, matching `X-Robots-Tag`) so crawlers focus on current docs instead of ~30 archived copies of every page, and restored `<lastmod>` dates in the docs sitemap that were silently dropped by production builds.
+
+</details>
+
 ## v0.19.3 (2026-08-13)
 
 **Added**

docs-site/demos.md

@@ -5,7 +5,7 @@ Short recordings of SubMiner's key features and integrations from real playback
 <script setup>
 import { withBase } from 'vitepress';
 
-const v = '20260301-1';
+const v = '20260819-1';
 </script>
 
 ## Anki Card Mining & Enrichment

docs-site/development.md

@@ -6,7 +6,7 @@ For internal architecture/workflow guidance, use `docs/README.md` at the repo ro
 
 - [Bun](https://bun.sh)
 - A system `lua` interpreter for `bun run test:launcher` / `bun run test:plugin:src`
-- macOS builds compile a Swift helper via `scripts/build-macos-helper.sh` (skip with `SUBMINER_SKIP_MACOS_HELPER_BUILD=1`)
+- macOS builds compile a Swift helper via `scripts/prepare-build-assets.mjs` (skip with `SUBMINER_SKIP_MACOS_HELPER_BUILD=1`)
 
 ## Setup
 

docs-site/immersion-tracking.md

@@ -82,7 +82,7 @@ Expandable session history with new-word activity, cumulative totals, and pause/
 
 #### Vocabulary
 
-Top repeated words (click a bar to open the word), new-word timeline, cross-title and frequency rank tables with Hide Known / Hide Kana filters, kanji breakdown, word exclusion list, and click-through occurrence drilldown with Mine Word / Mine Sentence / Mine Audio buttons.
+The summary cards show all unique vocabulary and kanji recorded in the local tracking database; **New This Week** is the only weekly figure and uses a rolling seven-day window. The word and kanji tables load first while those complete totals calculate separately. Top Repeated Words and New Words by Day use complete tracking history rather than the table's browsing page. New-word history is maintained as a permanent daily lexical rollup using the same token-visibility rules as the totals, including normalization of older timestamps stored in either seconds or milliseconds and retroactive corrections when tracked material is removed or reprocessed. On the first launch after an applicable upgrade, that history is version-rebuilt in the background and the chart refreshes when it is ready; if it remains unavailable, polling stops and an inline Retry control appears. The cards and charts also refresh automatically after the word exclusion list changes. The rest of the tab includes cross-title and frequency rank tables with Hide Known / Hide Kana filters, kanji breakdown, word exclusion list, and click-through occurrence drilldown with Mine Word / Mine Sentence / Mine Audio buttons.
 
 ![Stats Vocabulary](/screenshots/stats-vocabulary.png)
 
@@ -138,6 +138,8 @@ Karaoke openings and animated signs are authored as one subtitle event per anima
 
 Recording now collapses those runs as they happen, matching what the subtitle sidebar shows:
 
+- When a typeset ASS file stores a clean lyric or sign in a timed authoring comment, or in full-line events surrounding generated fragments, the matching complete line is recorded once. The repeated glyph or clip-animation frames are not recorded. Dialogue spoken while such an animation is on screen records as itself, without the fragment lines beside it.
+- When karaoke styling redraws the same complete lyric across consecutive color or highlight phases, those phases are combined into one line with their full timing. Repeated ordinary dialogue remains separate.
 - When the active subtitle source has been parsed, its cue list has already had duplicate events and animation bursts merged. A line landing inside a surviving cue but after that cue's start is a frame the sidebar merged away, and is not recorded.
 - When no parsed cue covers the live timing, including while a subtitle source is changing or shifted, the strict metadata-free rule applies: a run of identical, contiguous lines each shorter than 0.1s stops being recorded after a few frames. Runs are tracked per line of text, so dual-line karaoke (a kanji and a romaji line frame-flipped together) collapses both lines. Ordinary repeated dialogue, and lines held for a normal beat, always record.
 
@@ -180,6 +182,7 @@ In practice:
 - Anime and episode pages keep lifetime totals from summary tables while session drill-down still reads retained sessions directly. With the current defaults, both are kept forever.
 - Trends can read the full available history because daily/monthly rollups are also kept forever by default.
 - Vocabulary and kanji totals are cumulative and not bounded by the raw session retention knobs.
+- New-word charts use their own permanent lexical daily rollups, which are not pruned by activity-rollup retention.
 
 ## Storage / Performance Model
 
@@ -349,6 +352,7 @@ Rollup tables:
 
 - `imm_daily_rollups`
 - `imm_monthly_rollups`
+- `imm_lexical_daily_rollups` - permanent first-discovery counts for vocabulary and kanji chart history
 - `imm_rollup_state` - incremental rollup progress bookkeeping
 
 Vocabulary tables:

docs-site/index.md

@@ -88,7 +88,7 @@ features:
 <script setup>
 import { withBase } from 'vitepress';
 
-const demoAssetVersion = '20260223-2';
+const demoAssetVersion = '20260819-1';
 </script>
 
 <div class="landing-shell">

docs-site/installation.md

@@ -392,7 +392,7 @@ subminer -u
 subminer --update
 ```
 
-SubMiner verifies AppImage, launcher, and Linux support-asset downloads against `SHA256SUMS.txt`. On Linux those support assets include the launcher-managed runtime plugin copy under `SubMiner/plugin/subminer` plus the rofi theme at `SubMiner/themes/subminer.rasi`. If the binary is in a protected path, SubMiner shows the exact command to run rather than elevating itself.
+SubMiner verifies AppImage, launcher, and Linux support-asset downloads against `SHA256SUMS.txt`. On Linux those support assets include the launcher-managed runtime plugin copy under `SubMiner/plugin/subminer`, the rofi theme at `SubMiner/themes/subminer.rasi`, and the scoped Matroska thumbnailer registration under `SubMiner/thumbnailers`. If the binary is in a protected path, SubMiner shows the exact command to run rather than elevating itself.
 
 The tray "Check for Updates" entry installs the new app automatically on Linux, macOS, and Windows. On Linux it replaces the running `.AppImage` in place via `electron-updater` and refreshes the managed support assets from `subminer-assets.tar.gz`; AppImages managed by a system package (for example the AUR `/opt/SubMiner/SubMiner.AppImage`) are skipped so the package manager stays in charge.
 
@@ -404,7 +404,7 @@ SubMiner is an overlay that sits on top of mpv. It connects to mpv through an IP
 
 The `subminer` launcher handles mpv IPC socket setup automatically. If you launch mpv yourself or from another tool, you must pass `--input-ipc-server=/tmp/subminer-socket` (or `\\.\pipe\subminer-socket` on Windows) - without it the overlay starts but subtitles won't appear.
 
-The bundled mpv plugin is injected at runtime automatically - you don't need to install it separately. On Linux, the `subminer` launcher now checks for its managed runtime plugin copy and rofi theme before every mpv-managed launch and installs those support assets from the bundled app automatically if either one is missing. It provides in-player keybindings (the `y` chord) for controlling the overlay from within mpv. See [MPV Plugin](/mpv-plugin) for the full keybinding and configuration reference.
+The bundled mpv plugin is injected at runtime automatically - you don't need to install it separately. On Linux, the `subminer` launcher checks for its managed runtime plugin copy, rofi theme, and scoped thumbnailer registration before every mpv-managed launch and installs those support assets from the bundled app automatically if one is missing. For a rofi picker launch, this check runs before the picker opens. It provides in-player keybindings (the `y` chord) for controlling the overlay from within mpv. See [MPV Plugin](/mpv-plugin) for the full keybinding and configuration reference.
 
 ## Platform Notes
 
@@ -456,18 +456,20 @@ sudo chmod +x /usr/local/bin/subminer
 
 ### Linux Support Assets
 
-SubMiner ships the Linux rofi theme plus the launcher-managed runtime plugin copy in `subminer-assets.tar.gz`:
+SubMiner ships the Linux rofi theme, scoped Matroska thumbnailer registration, and launcher-managed runtime plugin copy in `subminer-assets.tar.gz`:
 
 ```bash
 wget https://github.com/ksyasuda/SubMiner/releases/latest/download/subminer-assets.tar.gz -O /tmp/subminer-assets.tar.gz
 tar -xzf /tmp/subminer-assets.tar.gz -C /tmp
 mkdir -p ~/.local/share/SubMiner/themes
 cp /tmp/assets/themes/subminer.rasi ~/.local/share/SubMiner/themes/subminer.rasi
+mkdir -p ~/.local/share/SubMiner/thumbnailers
+cp /tmp/assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer ~/.local/share/SubMiner/thumbnailers/
 mkdir -p ~/.local/share/SubMiner/plugin
 cp -R /tmp/plugin/subminer ~/.local/share/SubMiner/plugin/subminer
 ```
 
-`subminer -u` and the tray updater keep those Linux support assets in sync automatically once the `SubMiner` data dir exists. Normal Linux launcher playback also auto-installs the managed runtime plugin copy and rofi theme from the bundled app if either support asset is missing, so manual extraction is mainly useful for pre-seeding or custom setups.
+`subminer -u` and the tray updater keep those Linux support assets in sync automatically once the `SubMiner` data dir exists. Normal Linux launcher playback also auto-installs all three assets from the bundled app if one is missing, so manual extraction is mainly useful for pre-seeding or custom setups. Rofi receives the SubMiner data path through its process-local `XDG_DATA_DIRS`, so the thumbnailer registration does not change the desktop-wide configuration.
 
 Override the theme path with `SUBMINER_ROFI_THEME=/absolute/path/to/theme.rasi`.
 

docs-site/launcher-script.md

@@ -34,18 +34,22 @@ subminer -R -r -d ~/Anime              # rofi picker, recursive
 subminer -R /directory                 # rofi picker, directory shortcut
 ```
 
-rofi shows a GUI menu with icon thumbnails when available. SubMiner ships the rofi theme plus the Linux launcher-managed runtime plugin copy in the release assets tarball:
+rofi shows a GUI menu with icon thumbnails when available. SubMiner ships the rofi theme, a scoped `ffmpegthumbnailer` MIME registration, and the Linux launcher-managed runtime plugin copy in the release assets tarball:
 
 ```bash
 wget https://github.com/ksyasuda/SubMiner/releases/latest/download/subminer-assets.tar.gz -O /tmp/subminer-assets.tar.gz
 tar -xzf /tmp/subminer-assets.tar.gz -C /tmp
 mkdir -p ~/.local/share/SubMiner/themes
 cp /tmp/assets/themes/subminer.rasi ~/.local/share/SubMiner/themes/subminer.rasi
+mkdir -p ~/.local/share/SubMiner/thumbnailers
+cp /tmp/assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer ~/.local/share/SubMiner/thumbnailers/
 mkdir -p ~/.local/share/SubMiner/plugin
 cp -R /tmp/plugin/subminer ~/.local/share/SubMiner/plugin/subminer
 ```
 
-Once the `SubMiner` data dir exists, `subminer -u` refreshes both assets automatically. Normal Linux launcher playback also checks for the managed runtime plugin copy and rofi theme before mpv launch and installs them from the bundled app automatically if either one is missing.
+Once the `SubMiner` data dir exists, `subminer -u` refreshes these assets automatically. Normal Linux launcher playback checks for all three assets and installs them from the bundled app when one is missing. For `subminer -R`, this repair runs before rofi opens.
+
+When `ffmpegthumbnailer` is installed, SubMiner prepends its own data directory to `XDG_DATA_DIRS` for the rofi process only. This lets rofi recognize the canonical Matroska MIME types used by newer GLib versions without changing the desktop-wide MIME or thumbnailer configuration. An existing registration in your own `$XDG_DATA_HOME/thumbnailers` still takes priority.
 
 The theme is auto-detected from these paths (first match wins):
 

docs-site/mining-workflow.md

@@ -41,7 +41,7 @@ If you prefer a hands-on approach (animecards-style), you can copy the current s
    - For multiple lines: press `Ctrl/Cmd+Shift+C`, then a digit `1`–`9` to select how many recent subtitle lines to combine. The combined text is copied to the clipboard.
 3. Press `Ctrl/Cmd+V` to update the last-added card with the clipboard contents plus audio, image, and translation - the same fields auto-update would fill.
 
-Manual clipboard updates always replace generated sentence audio, even when `ankiConnect.behavior.overwriteAudio` is disabled. The word audio field is left unchanged because the word itself does not change in this flow.
+Manual clipboard updates always replace generated sentence audio in `ankiConnect.fields.audio`, even when `ankiConnect.behavior.overwriteAudio` is disabled. Normal word-card updates use the configured sentence and audio fields even when Lapis or Kiku support is enabled.
 
 This is useful when auto-update is disabled or when you want explicit control over which subtitle line gets attached to the card.
 
@@ -108,8 +108,12 @@ The secondary bar is a compact top-strip region in the same overlay window. It s
 - Quick comprehension checks without leaving the mining flow.
 - Auto-populating the translation field on mined cards - when a card is created, SubMiner uses the secondary subtitle text as the translation field value (unless AI translation is configured to override it).
 
+For local media, SubMiner can parse supported embedded secondary tracks into timed cues. For remote URLs and files on network mounts, it uses mpv's live secondary subtitle text instead of scanning the media with ffmpeg.
+
 It is controlled by `secondarySub` configuration and shares its lifecycle with the main overlay window. Cycle which track feeds it with `Shift+J`.
 
+SubMiner collapses duplicate ASS layers in parsed secondary tracks. Exact repeated lines collapse at any length, while distinct simultaneous short lines remain separate. Long dialogue and positioned-sign copies also collapse when they differ only in whitespace or terminal punctuation. Dense multi-row sign layouts, such as translated timetables, are excluded instead of being concatenated into the secondary bar.
+
 ### Display Modes
 
 Both the primary and secondary subtitle bars share the same three visibility modes, and each can be changed independently at runtime:

docs-site/subtitle-sidebar.md

@@ -9,9 +9,11 @@ The sidebar is enabled by default. Set `subtitleSidebar.enabled` to `false` if y
 When SubMiner parses the active subtitle source into a cue list, the sidebar becomes available. Toggle it with the `\` key (configurable via `subtitleSidebar.toggleKey`). While open:
 
 - The active cue is highlighted and kept in view as playback advances (when `autoScroll` is `true`).
-- Clicking any cue seeks mpv to that timestamp.
+- Clicking any cue seeks mpv into that line. For overlapping ASS karaoke, SubMiner moves past the previous line's exit animation when the selected cue has enough time remaining.
 - The sidebar stays synchronized with the overlay - media transitions and subtitle source changes update both simultaneously.
 
+For typeset ASS karaoke and animated signs, SubMiner collapses generated animation frames and repeated full-line color phases before they reach the sidebar. It recovers a clean complete line from a matching timed authoring comment or from full-line events surrounding generated fragments. Ordinary ASS comments, editor notes, alternate lines, repeated dialogue, and separately positioned signs remain distinct.
+
 The sidebar only appears when a parsed cue list is available. External subtitle sources that SubMiner cannot parse (for example, embedded ASS tracks rendered directly by mpv) will not populate the sidebar.
 
 ## Layout Modes

docs/RELEASING.md

@@ -58,12 +58,15 @@
    `latest*.yml` and `*.blockmap` files under `release/`.
 5. Commit the prerelease prep (package.json version bump + the generated
    `release/prerelease-notes.md`). CI does not regenerate notes — it uses the
-   committed file — so review it before committing. If you add more
-   `changes/*.md` fragments for a later beta/RC, rerun
-   `bun run changelog:prerelease-notes --version <version>`; the generator uses
-   the existing prerelease notes as the baseline only when their hidden
-   `prerelease-base-version` marker matches the current base version, and asks
-   Claude to merge only the new fragment material. Do not run
+   committed file — so review it before committing. Rerun
+   `bun run changelog:prerelease-notes --version <version>` for every later
+   beta/RC, even if no fragments changed: the notes carry a hidden
+   `prerelease-version` marker and CI rejects the tag when the marker does not
+   match it (verify locally with
+   `bun run changelog:check-prerelease-notes --version <version>`). The
+   generator reuses the existing notes as the cumulative baseline when their
+   marker (or legacy `prerelease-base-version` marker) matches the current base
+   version, and asks Claude to merge only the new fragment material. Do not run
    `bun run changelog:build`.
 6. Tag the commit: `git tag v<version>`.
 7. Push commit + tag.
@@ -78,6 +81,8 @@ Notes:
 - Pass `--date` explicitly when you want the release stamped with the local cut date; otherwise the generator uses the current ISO date, which can roll over to the next UTC day late at night.
 - `changelog:check` now rejects tag/package version mismatches.
 - `changelog:prerelease-notes` also rejects tag/package version mismatches and writes `release/prerelease-notes.md` without mutating tracked changelog files. When that file already exists, the generator includes it in the Claude prompt so later beta/RC notes reuse the reviewed text instead of starting over.
+- From the second prerelease of a base version onward, the notes open with a `## Changes since <previous tag>` section above the cumulative `## Highlights`. The generator locates the newest preceding beta/RC tag for the same base version (semver order: all betas before all RCs), diffs `changes/*.md` between that tag and the working tree, and asks Claude to describe only the behavioral beta-to-beta differences — added fragments as new changes, modified fragments by their before/after difference (editorial-only edits are dropped), deleted fragments as removed/reverted changes. If no fragments changed (for example a packaging-only rebuild), the section states that explicitly without a Claude call. The delta section carries no separate contributor attribution; `## What's Changed` stays cumulative like `## Highlights`.
+- `changelog:check-prerelease-notes --version <version>` verifies the committed notes' `prerelease-version` marker matches the version being tagged; the prerelease workflow runs it and fails the release on stale notes.
 - `changelog:build` generates `CHANGELOG.md` + `release/release-notes.md` (both polished by `claude -p`) and removes the released `changes/*.md` fragments. The CHANGELOG keeps internal notes inside a `<details><summary>Internal changes</summary>` collapse; the release notes drop them entirely.
 - `release/release-notes.md` (and `release/prerelease-notes.md`) include GitHub-style attribution after `## Highlights`: a `## What's Changed` list crediting each released fragment as `by @<author> in #<pr>`, plus a `## New Contributors` section for first-time authors. Attribution is resolved per fragment via `git log` (the commit that added the fragment) + `gh api .../commits/<sha>/pulls`, with one `gh` search per author for the first-contribution check. It needs `gh` installed and authenticated; if `gh` is unavailable or a lookup fails, the generator warns and emits notes without the attribution sections rather than failing. The CHANGELOG itself stays attribution-free.
 - The release workflow no longer auto-runs `changelog:build`. If pending `changes/*.md` fragments are present on a tag-based run, CI exits with a clear `::error::` pointing at the local fix. Run `bun run changelog:build --version <version>` locally, commit the polished output, then tag.

docs/architecture/2026-03-15-renderer-performance-design.md

@@ -70,18 +70,27 @@ interface SubtitleCue {
   startTime: number; // seconds
   endTime: number; // seconds
   text: string; // plain text, decoded from the source format
+  source?: 'canonical-ass'; // recovered authored text for generated ASS animation
+  animationStartTime?: number; // full generated-frame envelope; entrance/exit frames
+  animationEndTime?: number; //   run past the authored timing, live matching uses this
 }
 ```
 
 **Supported formats:**
 
 - SRT/VTT: Regex-based parsing of timing lines + text content between timing blocks.
-- ASS: Parse `[Events]` section, extract `Dialogue:` lines, read the field order from the `Format:` row, and take everything after the Text field index as the text (Text can itself contain commas).
+- ASS: Parse the `[Events]` section, read the field order from the `Format:` row, and extract timed `Dialogue:` lines. Timed `Comment:` lines are normally ignored, but can supply canonical authored text when they match a nearby generated animation from the same style and actor. Text can itself contain commas.
 
 **ASS decoding.** The parser is where ASS text is decoded, once, via `assToPlainText()` in `src/core/services/ass-text.ts`. That decoder mirrors mpv's `ass_to_plaintext` so a cue read from a file reads identically to the same line arriving live on `sub-text`: `{...}` override blocks are markup, `\pN … \p0` vector drawing runs are dropped rather than shown as text, `\N`/`\n`/`\h` are the only escapes (`\{`, `\}` and `\\` are not), and an unclosed `{` is rendered verbatim. Every layer downstream — renderer, timing tracker, tokenizer, tokenization cache keys — receives plain text and uses `normalizePlainSubtitleText()` for whitespace only, so nothing decodes the same string twice and one authored line always maps to one cache key.
 
 **Duplicate collapsing.** Typeset scripts emit one `Dialogue:` event per animation frame, plus layered copies of the same line. The parser collapses identical text over an identical span unconditionally, and collapses contiguous same-text runs of at least three events when the run looks like an animation. For ASS that means shared style and actor plus authoring evidence: a temporal tag (`\t`, `\move`, `\k`/`\kf`/`\ko`/`\K`, or anything wrapped in `\t(...)`), an animated `Effect` column (`Karaoke`, `Banner`, `Scroll`), or override values that change across the run. Static tags shared by every event (`\pos`, an identical `\clip`) are not evidence. SRT/VTT carry no such metadata, so there collapsing needs at least five contiguous events all under 0.1s — the frame timing left behind by ASS-to-SRT conversion. The parser keeps this authoring metadata (style, actor, layer, `Effect`, parsed override commands, source order) private; `parseSubtitleCues()` returns only `SubtitleCue`.
 
+ASS scripts can also redraw one complete lyric for two or more long color/highlight phases. Those flush-timed phases collapse separately from short animation frames when they share text, style, actor, and layer and carry direct animation evidence, such as temporal tags or changing non-spatial overrides. Spatial command changes do not prove a phase, so separately positioned signs remain distinct.
+
+**Canonical animation recovery.** Some ASS producers keep the readable lyric or sign as a timed `Comment:` and generate hundreds of `Dialogue:` frames containing repeated glyphs or changing clip regions. Others retain the complete line as brief `Dialogue:` events around the generated fragments. A complete event is promoted only when nearby dialogue from the same style and actor forms a proven animation cluster and reconstructs its entire text in source order. The generated frames are then replaced by one cue marked `source: 'canonical-ass'`. This source marker lets the live primary-subtitle path prefer the clean authored text and timing for display, sidebar history, immersion recording, and mining, while unmatched editor notes and alternative translations remain ignored. Secondary selection advances to an entering canonical cue at its generated animation start when the preceding authored cue ends before the new authored span. Unrelated simultaneous cues that continue through the new span remain visible.
+
+**Font texture cleanup.** A clipped repeated-glyph run or frequent changes to secondary alpha marks a texture seed. Clipped runs do not need a font override because some signs build their masks from ordinary `l` glyphs. The parser removes short clipped pieces that share a no-font seed's style and timing, or pieces that share a font seed's style, timing, and font even when the actor changes. It also removes positioned text layers with at least `E0` global alpha when they overlap a seed in the same style. Opaque authored sign text stays publishable when the texture switches fonts or actors around it.
+
 #### Prefetch Service Lifecycle
 
 1. **Activation trigger:** When a subtitle track is activated (or changes), check if it's external via MPV's `track-list` property. If `external === true`, read the file via `external-filename` using the existing `loadSubtitleSourceText` infrastructure.

docs/architecture/stats-trends-data-flow.md

@@ -23,7 +23,9 @@ Trend charts now consume one chart-oriented backend payload from `/api/stats/tre
   - lookup rate trends
   - watch-time by day-of-week/hour
 - vocabulary-backed:
-  - new-words trend
+  - new-words trend reads permanent daily lexical rollups
+  - rollup rows count only vocabulary-visible tokens and normalize mixed legacy timestamp units
+  - a persisted rollup version invalidates stale materializations and triggers an atomic background rebuild
 
 ## Metric Semantics
 

docs/architecture/subtitle-overlay-priming.md

@@ -3,7 +3,7 @@
 # Subtitle Overlay Priming
 
 Status: active
-Last verified: 2026-08-04
+Last verified: 2026-08-19
 Owner: Kyle Yasuda
 Read when: debugging subtitle state or blank Linux/X11 overlay windows when the visible overlay is shown or recreated
 
@@ -69,14 +69,82 @@ coming and prefetching would otherwise idle for the rest of the cue.
 
 ## Live Cue Delivery
 
+- Primary live text first resolves recovered canonical ASS animations. Otherwise, when
+  every live mpv line matches an active parsed cue, it uses the parsed cue text so exact
+  full-span style layers appear once instead of repeating for fill, border, blur, shadow,
+  or equivalent whitespace variants. Any unmatched live line keeps the complete live
+  stack, preserving dialogue or signs that overlap a lyric.
 - A tokenization cache miss emits the plain cue synchronously. Tokenization remains serialized so
   live work does not contend for Yomitan state.
+- The initial `time-pos`, explicit renderer seeks, and later seek-like jumps reprocess mpv's
+  current raw `sub-text` after the new playback time is stored. Explicit intent matters because
+  adjacent subtitle jumps can be shorter than the general seek-distance threshold. This corrects
+  ASS cleanup when mpv delivered the destination subtitle before the destination timestamp.
+- Renderer `sub-seek` commands use the active parsed cue list when available. Simultaneous cues
+  share one boundary, overlapping lyrics advance from the latest active boundary, and mpv's native
+  command remains the fallback when no parsed destination exists. This prevents generated karaoke
+  frames from consuming next/previous subtitle presses.
+- Subtitle sidebar selections seek past the preceding sanitized cue's overlapping exit span when
+  the selected cue has enough time remaining. This keeps direct row selection on the requested
+  karaoke line while clamping the seek inside that cue.
+- If startup paints raw text before embedded ASS parsing finishes, parsed cue arrival may replace
+  that provisional line. The one-prime-per-media guard still suppresses identical repeats.
 - If a newer cue arrives while an older line is still tokenizing, the newer plain cue or empty
   clear payload is emitted immediately. The older tokenization result is dropped before it can
   replace the current cue.
 - The current cue upgrades in place when its tokens and annotations are ready. This can reflow text
   or character images, but cue visibility does not wait for that work.
 
+## Secondary Subtitle Flow
+
+- `secondary-sub-text` remains the immediate fallback, so unreadable subtitle sources, remote URLs,
+  and still-extracting embedded tracks appear without waiting for file resolution. Embedded-track
+  extraction runs for local and network-mounted files alike (demuxing reads the whole container,
+  about 10 seconds per GB on gigabit, under a generous timeout); only true remote URLs skip it,
+  having no on-disk container to demux.
+- The live fallback also suppresses per-glyph typesetting walls: when many simultaneous
+  one-glyph lines are present (generated karaoke lettering flattened into live text), those
+  lines and their short syllable companions are dropped while concurrent dialogue lines stay.
+  This keeps the overlay clean while extraction is still in flight and for sources that never
+  produce parsed cues.
+- Parsed secondary text and the live fallback remove exact repeated lines at any length. A
+  flattened-line identity also removes long dialogue/sign repetitions that differ only in
+  whitespace or terminal punctuation, while distinct simultaneous short lines remain separate.
+- `secondary-subtitle-track.ts` resolves `secondary-sid` against mpv's track list. External tracks
+  are read directly; supported embedded text tracks are extracted through the same ffmpeg-backed
+  source resolver used by primary subtitle prefetching.
+- The selected source is parsed with `parseSubtitleCues()`, including metadata-aware ASS duplicate
+  and animation collapse. Playback `time-pos` selects the active parsed cue after applying
+  `secondary-sub-delay`.
+- Fragment reconstruction marks tall multi-row positioned parts as a grid only when they read
+  like tiling: a couple of texts repeated across many fragments, the same text re-shown at one
+  spot over time (countdown/animation frames), or scattered single glyphs. Secondary text omits
+  those grids instead of flattening a translated table or schedule into one synthetic line.
+  Wrapped lyric rows, CC-style dialogue blocks, and reconstructed single-line karaoke remain
+  eligible for display.
+- The resolved text is stored in `mpvClient.currentSecondarySubText` before it is broadcast. The
+  overlay, mining, timing tracker, and immersion statistics therefore consume the same secondary
+  text when a readable source is available.
+- Simultaneous parsed cues use whitespace-insensitive identity, so ASS layers that vary only
+  between ordinary, hard, or ideographic spaces appear once.
+- Simultaneous ASS lines are flattened in top-to-bottom positioned order, falling back to their
+  authored source order when no usable position exists.
+- Fragment-only ASS karaoke is reconstructed per style before publication. Explicit spaces
+  survive concatenation. Latin fragment typesetting with no literal spaces also recovers word
+  boundaries represented only by materially larger horizontal `\pos` or `\move` gaps within that
+  line. Unpositioned fragments stay compact instead of gaining guessed spaces between syllables.
+  Short runs qualify only when overlapping positioned events also show changing overrides or
+  repeated layer copies; an English or romaji style name alone never turns ordinary dialogue into
+  a lyric.
+- Recovered canonical ASS text remains active for the generated animation envelope. For
+  reconstructed lyric styles, the longest-lived active line wins over brief entrance and exit
+  fragments from the same style.
+- Media and `secondary-sid` changes clear the previous parsed state before refreshing the source;
+  track-list changes refresh without discarding an unchanged source. Observed
+  `secondary-sub-delay` changes retime the active parsed cue without rereading the file. If loading,
+  extraction, or parsing fails, the controller returns to live mpv text and the renderer's
+  conservative short stack heuristic remains the final display fallback.
+
 ## Emitted State
 
 - `emitSubtitle(payload)` maps to `emitSubtitlePayload(payload)`. Overlay windows and annotation
@@ -84,8 +152,8 @@ coming and prefetching would otherwise idle for the rest of the cue.
 - The basic subtitle websocket receives the immediate plain cue only. Because its serialized
   payload discards annotations, the later upgrade would be an identical duplicate and is skipped
   when text and cue timing match.
-- Secondary priming reads mpv `secondary-sub-text`, stores it in
-  `mpvClient.currentSecondarySubText`, and broadcasts `secondary-subtitle:set` to overlay windows.
+- Secondary priming reads mpv `secondary-sub-text` and routes it through the secondary track
+  controller. A parsed active cue replaces the live text when the selected source is readable.
 - If secondary `requestProperty` fails, the primary flow stays complete and only a debug line is
   written.
 

docs/knowledge-base/catalog.md

@@ -19,7 +19,7 @@ Read when: finding internal docs or checking verification status
 | Quality scorecard | `docs/knowledge-base/quality.md` | active | 2026-03-13 | quality grades and gaps |
 | Workflow index | `docs/workflow/README.md` | active | 2026-08-13 | execution map |
 | Planning guide | `docs/workflow/planning.md` | active | 2026-05-23 | lightweight vs execution plans |
-| Agent skills | `docs/workflow/agent-skills.md` | active | 2026-08-13 | repo-local workflow skill ownership |
+| Agent skills | `docs/workflow/agent-skills.md` | active | 2026-08-23 | repo-local workflow skill ownership |
 | Verification guide | `docs/workflow/verification.md` | active | 2026-08-13 | maintained verification lanes |
 | Release guide | `docs/RELEASING.md` | active | 2026-05-23 | release checklist |
 

docs/workflow/agent-skills.md

@@ -3,7 +3,7 @@
 # Agent Skills
 
 Status: active
-Last verified: 2026-08-13
+Last verified: 2026-08-23
 Owner: Kyle Yasuda
 Read when: using, adding, or changing a repo-local agent workflow skill
 
@@ -12,6 +12,9 @@ Read when: using, adding, or changing a repo-local agent workflow skill
 - `.agents/skills/subminer-change-verification/`
   - Selects the cheapest sufficient repo-native verification lane.
   - Defers command ownership to `package.json` and `docs/workflow/verification.md`.
+- `.agents/skills/subminer-release/`
+  - Prepares, cuts, publishes, or repairs stable and prerelease releases.
+  - Defers release procedure and policy to `docs/RELEASING.md`.
 
 Repo-local workflows stay as standalone skills. Do not add plugin packaging, marketplace metadata, or compatibility shims unless the workflow is intentionally being distributed beyond this repository.
 

launcher/commands/history-command.ts

@@ -23,6 +23,7 @@ import {
   type HistorySeriesEntry,
 } from '../history.js';
 import type { Args } from '../types.js';
+import { ensureLinuxRuntimePluginAvailable } from '../runtime-plugin-preflight.js';
 import type { LauncherCommandContext } from './context.js';
 
 export type HistorySessionAction = 'previous' | 'replay' | 'next' | 'browse' | 'quit';
@@ -333,6 +334,13 @@ export async function runHistoryCommand(
   const { args, scriptPath } = context;
 
   checkPickerDependencies(args);
+  if (args.useRofi) {
+    await ensureLinuxRuntimePluginAvailable({
+      appPath: context.appPath ?? undefined,
+      scriptPath,
+      logLevel: args.logLevel,
+    });
+  }
   const themePath = args.useRofi ? findRofiTheme(scriptPath) : null;
 
   const dbPath = resolveImmersionDbPath();

launcher/commands/jellyfin-command.ts

@@ -2,6 +2,7 @@ import { fail } from '../log.js';
 import { runAppCommandWithInherit } from '../mpv.js';
 import { commandExists } from '../util.js';
 import { runJellyfinPlayMenu } from '../jellyfin.js';
+import { ensureLinuxRuntimePluginAvailable } from '../runtime-plugin-preflight.js';
 import { shouldForwardLogLevel } from '../types.js';
 import type { LauncherCommandContext } from './context.js';
 
@@ -64,6 +65,13 @@ export async function runJellyfinCommand(context: LauncherCommandContext): Promi
     if (args.useRofi && !commandExists('rofi')) {
       fail('rofi not found. Install rofi or omit -R for fzf.');
     }
+    if (args.useRofi) {
+      await ensureLinuxRuntimePluginAvailable({
+        appPath,
+        scriptPath,
+        logLevel: args.logLevel,
+      });
+    }
     await runJellyfinPlayMenu(appPath, args, scriptPath, mpvSocketPath);
     return true;
   }

launcher/commands/playback-command.test.ts

@@ -496,3 +496,39 @@ test('playback command ensures Linux runtime plugin before mpv launch', async ()
 
   assert.deepEqual(calls, ['plugin', 'startMpv']);
 });
+
+test('rofi playback repairs support assets before opening the picker', async () => {
+  const context = createContext();
+  context.args = {
+    ...context.args,
+    target: '',
+    targetKind: '',
+    useRofi: true,
+  };
+  const calls: string[] = [];
+
+  await runPlaybackCommandWithDeps(context, {
+    ensurePlaybackSetupReady: async () => {},
+    ensureRuntimePluginReady: async () => {
+      calls.push('assets');
+    },
+    chooseTarget: async () => {
+      calls.push('picker');
+      return { target: '/tmp/movie.mkv', kind: 'file' };
+    },
+    checkPickerDependencies: () => {},
+    checkDependencies: () => {},
+    registerCleanup: () => {},
+    startMpv: async () => {
+      calls.push('startMpv');
+    },
+    waitForUnixSocketReady: async () => true,
+    startOverlay: async () => {},
+    launchAppCommandDetached: () => {},
+    log: () => {},
+    cleanupPlaybackSession: async () => {},
+    getMpvProc: () => null,
+  });
+
+  assert.deepEqual(calls, ['assets', 'picker', 'startMpv']);
+});

launcher/commands/playback-command.ts

@@ -157,6 +157,7 @@ export async function runPlaybackCommand(context: LauncherCommandContext): Promi
       });
     },
     chooseTarget,
+    checkPickerDependencies,
     checkDependencies,
     registerCleanup,
     startMpv,
@@ -177,6 +178,7 @@ type PlaybackCommandDeps = {
     args: Args,
     scriptPath: string,
   ) => Promise<{ target: string; kind: 'file' | 'url' } | null>;
+  checkPickerDependencies?: (args: Args) => void;
   checkDependencies: (args: Args) => void;
   registerCleanup: (context: LauncherCommandContext) => void;
   startMpv: typeof startMpv;
@@ -201,7 +203,18 @@ export async function runPlaybackCommandWithDeps(
   await deps.ensurePlaybackSetupReady(context);
 
   if (!args.target) {
-    checkPickerDependencies(args);
+    (deps.checkPickerDependencies ?? checkPickerDependencies)(args);
+  }
+
+  let runtimeAssetsReady = false;
+  const ensureRuntimeAssetsReady = async (): Promise<void> => {
+    if (runtimeAssetsReady) return;
+    await deps.ensureRuntimePluginReady(context);
+    runtimeAssetsReady = true;
+  };
+
+  if (!args.target && args.useRofi) {
+    await ensureRuntimeAssetsReady();
   }
 
   const targetChoice = await deps.chooseTarget(args, scriptPath);
@@ -266,7 +279,7 @@ export async function runPlaybackCommandWithDeps(
     );
   }
 
-  await deps.ensureRuntimePluginReady(context);
+  await ensureRuntimeAssetsReady();
 
   await deps.startMpv(
     selectedTarget.target,

launcher/commands/update-command.test.ts

@@ -36,6 +36,11 @@ test('runUpdateCommand updates directly on Linux without launching Electron', as
         launcher: { status: 'updated' },
         supportAssets: [
           { status: 'updated', component: 'theme', message: 'Installed theme.' },
+          {
+            status: 'updated',
+            component: 'thumbnailer',
+            message: 'Installed rofi thumbnailer.',
+          },
           { status: 'skipped', component: 'plugin', message: 'Plugin already up to date.' },
         ],
       };
@@ -52,6 +57,7 @@ test('runUpdateCommand updates directly on Linux without launching Electron', as
     'info:AppImage update: updated',
     'info:Launcher update: updated',
     'info:Support assets (theme) update: updated - Installed theme.',
+    'info:Support assets (thumbnailer) update: updated - Installed rofi thumbnailer.',
     'info:Support assets (plugin) update: skipped - Plugin already up to date.',
   ]);
 });

launcher/commands/update-command.ts

@@ -21,7 +21,10 @@ import {
   parseSha256Sums,
   type FetchLike,
 } from '../../src/main/runtime/update/release-assets.js';
-import { updateSupportAssetsFromRelease } from '../../src/main/runtime/update/support-assets.js';
+import {
+  updateSupportAssetsFromRelease,
+  type SupportAssetsUpdateResult,
+} from '../../src/main/runtime/update/support-assets.js';
 
 type UpdateCommandResponse = {
   ok: boolean;
@@ -36,15 +39,14 @@ type DirectReleaseUpdateRequest = {
   channel: UpdateChannel;
 };
 
+type DirectSupportAssetsUpdateResult = Omit<SupportAssetsUpdateResult, 'status'> & {
+  status: string;
+};
+
 type DirectReleaseUpdateResult = {
   appImage: { status: string; command?: string; message?: string };
   launcher: { status: string; command?: string; message?: string };
-  supportAssets: Array<{
-    status: string;
-    component?: 'theme' | 'plugin';
-    command?: string;
-    message?: string;
-  }>;
+  supportAssets: DirectSupportAssetsUpdateResult[];
 };
 
 type UpdateCommandDeps = {
@@ -129,12 +131,7 @@ function readUpdateChannel(root: Record<string, unknown> | null): UpdateChannel
 
 function logUpdateResult(
   label: string,
-  result: {
-    status: string;
-    component?: 'theme' | 'plugin';
-    command?: string;
-    message?: string;
-  },
+  result: DirectSupportAssetsUpdateResult,
   configuredLogLevel: NonNullable<LauncherCommandContext['args']['logLevel']>,
   deps: Pick<UpdateCommandDeps, 'log'>,
 ): void {

launcher/main.test.ts

@@ -73,20 +73,21 @@ function makeTestEnv(homeDir: string, xdgConfigHome: string): NodeJS.ProcessEnv
   };
 }
 
-// On Linux the playback path runs `ensureLinuxRuntimePluginAvailable`, which —
-// when the runtime plugin/theme are missing — spawns the app with
-// `--ensure-linux-runtime-plugin-assets` and polls up to 30s
+// On Linux the playback path runs `ensureLinuxRuntimePluginAvailable`, which
+// spawns the app with `--ensure-linux-runtime-plugin-assets` when managed
+// support assets are missing and polls up to 30s
 // (RESPONSE_TIMEOUT_MS) for an install response. A fake app that just exits
 // never writes that response, so the launcher hangs and the test times out on
 // Linux CI (the preflight is a no-op on macOS/Windows). This shell prelude makes
-// the fake app install the managed plugin/theme and write the response, matching
+// the fake app install the managed support assets and write the response, matching
 // launcher/smoke.e2e.test.ts. Prepend it to each fake app that reaches playback.
 const RUNTIME_PLUGIN_PREFLIGHT_SH = `if [ "$1" = "--ensure-linux-runtime-plugin-assets" ]; then
   data="\${XDG_DATA_HOME:-$HOME/.local/share}/SubMiner"
-  mkdir -p "$data/plugin/subminer" "$data/themes"
+  mkdir -p "$data/plugin/subminer" "$data/themes" "$data/thumbnailers"
   printf -- '-- test plugin\\n' > "$data/plugin/subminer/main.lua"
   printf 'test=true\\n' > "$data/plugin/subminer.conf"
   printf '/* test theme */\\n' > "$data/themes/subminer.rasi"
+  printf '[Thumbnailer Entry]\\n' > "$data/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer"
   if [ "$2" = "--ensure-linux-runtime-plugin-assets-response-path" ] && [ -n "$3" ]; then
     mkdir -p "$(dirname "$3")"
     printf '{"ok":true,"status":"installed","path":"%s"}' "$data/plugin/subminer/main.lua" > "$3"

launcher/picker.test.ts

@@ -3,7 +3,12 @@ import assert from 'node:assert/strict';
 import fs from 'node:fs';
 import path from 'node:path';
 import os from 'node:os';
-import { findRofiTheme, formatRofiPrompt } from './picker';
+import {
+  findRofiTheme,
+  findRofiThumbnailerDataRoot,
+  formatRofiPrompt,
+  prependXdgDataDir,
+} from './picker';
 
 // ── formatRofiPrompt: spacing between prompt and input field ──────────────────
 
@@ -23,6 +28,7 @@ test('formatRofiPrompt leaves an empty prompt empty', () => {
 // ── findRofiTheme: Linux packaged path discovery ──────────────────────────────
 
 const ROFI_THEME_FILE = 'subminer.rasi';
+const ROFI_THUMBNAILER_FILE = 'subminer-ffmpegthumbnailer.thumbnailer';
 
 function makeFile(filePath: string): void {
   fs.mkdirSync(path.dirname(filePath), { recursive: true });
@@ -121,3 +127,42 @@ test('findRofiTheme resolves ~/.local/share/SubMiner/themes/subminer.rasi when X
     fs.rmSync(baseDir, { recursive: true, force: true });
   }
 });
+
+test('findRofiThumbnailerDataRoot resolves the managed XDG data root', () => {
+  const xdgDataHome = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-test-xdg-'));
+  const originalXdgDataHome = process.env.XDG_DATA_HOME;
+  try {
+    process.env.XDG_DATA_HOME = xdgDataHome;
+    const dataRoot = path.join(xdgDataHome, 'SubMiner');
+    makeFile(path.join(dataRoot, 'thumbnailers', ROFI_THUMBNAILER_FILE));
+
+    const result = withPlatform('linux', () => findRofiThumbnailerDataRoot('/usr/bin/subminer'));
+    assert.equal(result, dataRoot);
+  } finally {
+    if (originalXdgDataHome === undefined) {
+      delete process.env.XDG_DATA_HOME;
+    } else {
+      process.env.XDG_DATA_HOME = originalXdgDataHome;
+    }
+    fs.rmSync(xdgDataHome, { recursive: true, force: true });
+  }
+});
+
+test('findRofiThumbnailerDataRoot is Linux-only', () => {
+  assert.equal(
+    withPlatform('darwin', () => findRofiThumbnailerDataRoot('/usr/bin/subminer')),
+    null,
+  );
+});
+
+test('prependXdgDataDir preserves existing roots and avoids duplicates', () => {
+  const root = '/tmp/subminer-data';
+  assert.equal(
+    prependXdgDataDir(root, `/opt/share${path.delimiter}${root}${path.delimiter}/usr/share`),
+    `${root}${path.delimiter}/opt/share${path.delimiter}/usr/share`,
+  );
+  assert.equal(
+    prependXdgDataDir(root),
+    `${root}${path.delimiter}/usr/local/share${path.delimiter}/usr/share`,
+  );
+});

launcher/picker.ts

@@ -159,6 +159,9 @@ interface RofiIconEntry {
   iconPath?: string;
 }
 
+const ROFI_THUMBNAILER_FILE = 'subminer-ffmpegthumbnailer.thumbnailer';
+const DEFAULT_XDG_DATA_DIRS = ['/usr/local/share', '/usr/share'];
+
 function showRofiIconMenu(
   entries: RofiIconEntry[],
   prompt: string,
@@ -389,6 +392,47 @@ export function findRofiTheme(scriptPath: string): string | null {
   return null;
 }
 
+export function findRofiThumbnailerDataRoot(scriptPath: string): string | null {
+  if (process.platform !== 'linux') return null;
+
+  const scriptDir = path.dirname(realpathMaybe(scriptPath));
+  const xdgDataHome = process.env.XDG_DATA_HOME || path.join(os.homedir(), '.local/share');
+  const roots = [
+    path.join(xdgDataHome, 'SubMiner'),
+    path.posix.join('/usr/local/share/SubMiner'),
+    path.posix.join('/usr/share/SubMiner'),
+    path.join(scriptDir, 'assets'),
+    path.join(scriptDir, '..', 'assets'),
+  ];
+
+  for (const root of roots) {
+    if (fs.existsSync(path.join(root, 'thumbnailers', ROFI_THUMBNAILER_FILE))) {
+      return root;
+    }
+  }
+
+  return null;
+}
+
+export function prependXdgDataDir(dataRoot: string, currentValue?: string): string {
+  const currentDirs = currentValue
+    ? currentValue.split(path.delimiter).filter(Boolean)
+    : DEFAULT_XDG_DATA_DIRS;
+  return [dataRoot, ...currentDirs.filter((candidate) => candidate !== dataRoot)].join(
+    path.delimiter,
+  );
+}
+
+function buildRofiThumbnailEnvironment(scriptPath: string): NodeJS.ProcessEnv {
+  if (!commandExists('ffmpegthumbnailer')) return process.env;
+  const dataRoot = findRofiThumbnailerDataRoot(scriptPath);
+  if (!dataRoot) return process.env;
+  return {
+    ...process.env,
+    XDG_DATA_DIRS: prependXdgDataDir(dataRoot, process.env.XDG_DATA_DIRS),
+  };
+}
+
 export function showRofiMenu(
   videos: string[],
   dir: string,
@@ -420,6 +464,7 @@ export function showRofiMenu(
   const result = spawnSync('rofi', args, {
     input: buildRofiMenu(videos, dir, recursive),
     encoding: 'utf8',
+    env: buildRofiThumbnailEnvironment(scriptPath),
     stdio: ['pipe', 'pipe', 'ignore'],
   });
   if (result.error) {

launcher/runtime-plugin-preflight.test.ts

@@ -1,6 +1,8 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
 import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
 import {
   ensureLinuxRuntimePluginAvailable,
   installManagedPluginAssetsViaApp,
@@ -31,7 +33,7 @@ test('ensureLinuxRuntimePluginAvailable is a no-op on non-Linux platforms', asyn
   assert.deepEqual(calls, []);
 });
 
-test('ensureLinuxRuntimePluginAvailable skips install when installed global plugin and managed theme exist', async () => {
+test('ensureLinuxRuntimePluginAvailable skips install when plugin, theme, and thumbnailer exist', async () => {
   const calls: string[] = [];
 
   await ensureLinuxRuntimePluginAvailable({
@@ -52,13 +54,17 @@ test('ensureLinuxRuntimePluginAvailable skips install when installed global plug
       calls.push('theme');
       return true;
     },
+    isManagedThumbnailerAvailable: () => {
+      calls.push('thumbnailer');
+      return true;
+    },
     log: () => {},
   });
 
-  assert.deepEqual(calls, ['detect', 'theme']);
+  assert.deepEqual(calls, ['detect', 'theme', 'thumbnailer']);
 });
 
-test('ensureLinuxRuntimePluginAvailable skips install when managed runtime path and theme already resolve', async () => {
+test('ensureLinuxRuntimePluginAvailable skips install when all managed assets resolve', async () => {
   const calls: string[] = [];
 
   await ensureLinuxRuntimePluginAvailable({
@@ -80,14 +86,19 @@ test('ensureLinuxRuntimePluginAvailable skips install when managed runtime path
       calls.push('theme');
       return true;
     },
+    isManagedThumbnailerAvailable: () => {
+      calls.push('thumbnailer');
+      return true;
+    },
     log: () => {},
   });
 
-  assert.deepEqual(calls, ['detect', 'resolve', 'theme']);
+  assert.deepEqual(calls, ['detect', 'resolve', 'theme', 'thumbnailer']);
 });
 
 test('ensureLinuxRuntimePluginAvailable installs managed assets when rofi theme is missing', async () => {
   const calls: string[] = [];
+  let themeAvailable = false;
 
   await ensureLinuxRuntimePluginAvailable({
     platform: 'linux',
@@ -102,10 +113,15 @@ test('ensureLinuxRuntimePluginAvailable installs managed assets when rofi theme
     },
     isManagedThemeAvailable: () => {
       calls.push('theme');
-      return false;
+      return themeAvailable;
+    },
+    isManagedThumbnailerAvailable: () => {
+      calls.push('thumbnailer');
+      return true;
     },
     installManagedPluginAssets: async () => {
       calls.push('install');
+      themeAvailable = true;
       return { ok: true, status: 'installed', path: '/tmp/plugin/main.lua' };
     },
     log: (level, _configured, message) => {
@@ -117,13 +133,68 @@ test('ensureLinuxRuntimePluginAvailable installs managed assets when rofi theme
     'detect',
     'resolve',
     'theme',
-    'info:Linux runtime support assets missing; installing managed plugin/theme assets.',
+    'info:Linux runtime support assets missing; installing managed plugin/theme/thumbnailer assets.',
     'install',
-    'info:Managed Linux runtime support assets installed: plugin=/tmp/plugin/main.lua theme=/tmp/xdg-data/SubMiner/themes/subminer.rasi',
+    'info:Managed Linux runtime support assets installed: plugin=/tmp/plugin/main.lua theme=/tmp/xdg-data/SubMiner/themes/subminer.rasi thumbnailer=/tmp/xdg-data/SubMiner/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer',
     'resolve',
+    'theme',
+    'thumbnailer',
   ]);
 });
 
+test('ensureLinuxRuntimePluginAvailable installs managed assets when thumbnailer is missing', async () => {
+  const calls: string[] = [];
+  let thumbnailerAvailable = false;
+
+  await ensureLinuxRuntimePluginAvailable({
+    platform: 'linux',
+    xdgDataHome: '/tmp/xdg-data',
+    detectInstalledPlugin: () => true,
+    resolveRuntimePluginPath: () => '/tmp/plugin/main.lua',
+    isManagedThemeAvailable: () => true,
+    isManagedThumbnailerAvailable: () => thumbnailerAvailable,
+    installManagedPluginAssets: async () => {
+      calls.push('install');
+      thumbnailerAvailable = true;
+      return { ok: true, status: 'installed', path: '/tmp/plugin/main.lua' };
+    },
+    log: (_level, _configured, message) => {
+      calls.push(message);
+    },
+  });
+
+  assert.deepEqual(calls, [
+    'Linux runtime support assets missing; installing managed plugin/theme/thumbnailer assets.',
+    'install',
+    'Managed Linux runtime support assets installed: plugin=/tmp/plugin/main.lua theme=/tmp/xdg-data/SubMiner/themes/subminer.rasi thumbnailer=/tmp/xdg-data/SubMiner/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer',
+  ]);
+});
+
+test('ensureLinuxRuntimePluginAvailable retains an installed plugin after installing support assets', async () => {
+  const calls: string[] = [];
+  let thumbnailerAvailable = false;
+
+  await ensureLinuxRuntimePluginAvailable({
+    platform: 'linux',
+    xdgDataHome: '/tmp/xdg-data',
+    detectInstalledPlugin: () => true,
+    resolveRuntimePluginPath: () => {
+      calls.push('resolve');
+      return null;
+    },
+    isManagedThemeAvailable: () => true,
+    isManagedThumbnailerAvailable: () => thumbnailerAvailable,
+    installManagedPluginAssets: async () => {
+      calls.push('install');
+      thumbnailerAvailable = true;
+      return { ok: true, status: 'installed', path: '/tmp/plugin/main.lua' };
+    },
+    log: () => {},
+  });
+
+  assert.deepEqual(calls, ['install']);
+});
+
 test('ensureLinuxRuntimePluginAvailable installs managed assets and re-resolves plugin path', async () => {
   const calls: string[] = [];
   let resolveCount = 0;
@@ -137,6 +208,8 @@ test('ensureLinuxRuntimePluginAvailable installs managed assets and re-resolves
       calls.push(`resolve:${resolveCount}`);
       return resolveCount === 1 ? null : '/tmp/plugin/main.lua';
     },
+    isManagedThemeAvailable: () => true,
+    isManagedThumbnailerAvailable: () => true,
     installManagedPluginAssets: async () => {
       calls.push('install');
       return { ok: true, status: 'installed', path: '/tmp/plugin/main.lua' };
@@ -148,9 +221,9 @@ test('ensureLinuxRuntimePluginAvailable installs managed assets and re-resolves
 
   assert.deepEqual(calls, [
     'resolve:1',
-    'info:Linux runtime support assets missing; installing managed plugin/theme assets.',
+    'info:Linux runtime support assets missing; installing managed plugin/theme/thumbnailer assets.',
     'install',
-    'info:Managed Linux runtime support assets installed: plugin=/tmp/plugin/main.lua theme=/tmp/xdg-data/SubMiner/themes/subminer.rasi',
+    'info:Managed Linux runtime support assets installed: plugin=/tmp/plugin/main.lua theme=/tmp/xdg-data/SubMiner/themes/subminer.rasi thumbnailer=/tmp/xdg-data/SubMiner/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer',
     'resolve:2',
   ]);
 });
@@ -191,6 +264,60 @@ test('ensureLinuxRuntimePluginAvailable fails when runtime path remains unresolv
   );
 });
 
+test('ensureLinuxRuntimePluginAvailable fails when thumbnailer remains missing after install', async () => {
+  await assert.rejects(
+    () =>
+      ensureLinuxRuntimePluginAvailable({
+        platform: 'linux',
+        xdgDataHome: '/tmp/xdg-data',
+        detectInstalledPlugin: () => true,
+        resolveRuntimePluginPath: () => '/tmp/plugin/main.lua',
+        isManagedThemeAvailable: () => true,
+        isManagedThumbnailerAvailable: () => false,
+        installManagedPluginAssets: async () => ({
+          ok: true,
+          status: 'installed',
+          path: '/tmp/plugin/main.lua',
+        }),
+        log: () => {},
+      }),
+    /thumbnailer=.*subminer-ffmpegthumbnailer\.thumbnailer/i,
+  );
+});
+
+test('ensureLinuxRuntimePluginAvailable rejects a thumbnailer directory before and after install', async () => {
+  const xdgDataHome = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-thumbnailer-directory-'));
+  const thumbnailerPath = path.join(
+    xdgDataHome,
+    'SubMiner',
+    'thumbnailers',
+    'subminer-ffmpegthumbnailer.thumbnailer',
+  );
+  fs.mkdirSync(thumbnailerPath, { recursive: true });
+  const calls: string[] = [];
+
+  try {
+    await assert.rejects(
+      () =>
+        ensureLinuxRuntimePluginAvailable({
+          platform: 'linux',
+          xdgDataHome,
+          detectInstalledPlugin: () => true,
+          isManagedThemeAvailable: () => true,
+          installManagedPluginAssets: async () => {
+            calls.push('install');
+            return { ok: true, status: 'installed', path: '/tmp/plugin/main.lua' };
+          },
+          log: () => {},
+        }),
+      /thumbnailer=.*subminer-ffmpegthumbnailer\.thumbnailer/i,
+    );
+    assert.deepEqual(calls, ['install']);
+  } finally {
+    fs.rmSync(xdgDataHome, { recursive: true, force: true });
+  }
+});
+
 test('installManagedPluginAssetsViaApp returns launch errors without waiting for a response file', async () => {
   let waited = false;
 

launcher/runtime-plugin-preflight.ts

@@ -31,6 +31,7 @@ type EnsureLinuxRuntimePluginAvailableOptions = {
   detectInstalledPlugin?: () => boolean;
   resolveRuntimePluginPath?: () => string | null;
   isManagedThemeAvailable?: () => boolean;
+  isManagedThumbnailerAvailable?: () => boolean;
   installManagedPluginAssets?: () => Promise<EnsureLinuxRuntimePluginAssetsResult>;
   log?: PreflightLog;
 };
@@ -48,6 +49,14 @@ function resolveConfiguredLogLevel(
   return logLevel ?? 'warn';
 }
 
+function isRegularFile(filePath: string): boolean {
+  try {
+    return fs.statSync(filePath).isFile();
+  } catch {
+    return false;
+  }
+}
+
 async function waitForInstallResponse(
   responsePath: string,
 ): Promise<RuntimePluginPreflightResponse | null> {
@@ -170,15 +179,17 @@ export async function ensureLinuxRuntimePluginAvailable(
     });
   const isManagedThemeAvailable =
     options.isManagedThemeAvailable ?? (() => fs.existsSync(managedPaths.themePath));
+  const isManagedThumbnailerAvailable =
+    options.isManagedThumbnailerAvailable ?? (() => isRegularFile(managedPaths.thumbnailerPath));
   const runtimePluginAvailable = installedPluginAvailable || Boolean(resolveRuntimePluginPath());
-  if (runtimePluginAvailable && isManagedThemeAvailable()) {
+  if (runtimePluginAvailable && isManagedThemeAvailable() && isManagedThumbnailerAvailable()) {
     return;
   }
 
   log(
     'info',
     configuredLogLevel,
-    'Linux runtime support assets missing; installing managed plugin/theme assets.',
+    'Linux runtime support assets missing; installing managed plugin/theme/thumbnailer assets.',
   );
   const installManagedPluginAssets =
     options.installManagedPluginAssets ??
@@ -207,16 +218,21 @@ export async function ensureLinuxRuntimePluginAvailable(
   log(
     'info',
     configuredLogLevel,
-    `Managed Linux runtime support assets installed: plugin=${installResult.path ?? 'unknown path'} theme=${managedPaths.themePath}`,
+    `Managed Linux runtime support assets installed: plugin=${installResult.path ?? 'unknown path'} theme=${managedPaths.themePath} thumbnailer=${managedPaths.thumbnailerPath}`,
   );
-  const runtimePluginPath = resolveRuntimePluginPath();
-  if (runtimePluginPath) {
+  const runtimePluginAvailableAfterInstall =
+    installedPluginAvailable || Boolean(resolveRuntimePluginPath());
+  if (
+    runtimePluginAvailableAfterInstall &&
+    isManagedThemeAvailable() &&
+    isManagedThumbnailerAvailable()
+  ) {
     return;
   }
 
   const message =
     `Linux managed runtime plugin assets could not be installed. ` +
-    `Checked path: ${managedPaths.pluginEntrypointPath}. ` +
+    `Checked paths: plugin=${managedPaths.pluginEntrypointPath} theme=${managedPaths.themePath} thumbnailer=${managedPaths.thumbnailerPath}. ` +
     'Launch aborted before starting mpv.';
   log('warn', configuredLogLevel, message);
   throw new Error(message);

launcher/smoke.e2e.test.ts

@@ -165,11 +165,14 @@ if (entry.argv.includes('--ensure-linux-runtime-plugin-assets')) {
   const pluginDir = path.join(dataDir, 'plugin', 'subminer');
   const pluginConfigPath = path.join(dataDir, 'plugin', 'subminer.conf');
   const themePath = path.join(dataDir, 'themes', 'subminer.rasi');
+  const thumbnailerPath = path.join(dataDir, 'thumbnailers', 'subminer-ffmpegthumbnailer.thumbnailer');
   fs.mkdirSync(pluginDir, { recursive: true });
   fs.mkdirSync(path.dirname(themePath), { recursive: true });
+  fs.mkdirSync(path.dirname(thumbnailerPath), { recursive: true });
   fs.writeFileSync(path.join(pluginDir, 'main.lua'), '-- smoke plugin\\n');
   fs.writeFileSync(pluginConfigPath, 'smoke=true\\n');
   fs.writeFileSync(themePath, '/* smoke theme */\\n');
+  fs.writeFileSync(thumbnailerPath, '[Thumbnailer Entry]\\n');
   if (responsePath) {
     fs.mkdirSync(path.dirname(responsePath), { recursive: true });
     fs.writeFileSync(responsePath, JSON.stringify({ ok: true, status: 'installed', path: path.join(pluginDir, 'main.lua') }));
@@ -620,11 +623,22 @@ test(
       );
       assert.match(result.stdout, /pause mpv until overlay and tokenization are ready/i);
       if (process.platform === 'linux') {
-        assert.match(result.stdout, /managed plugin\/theme assets/i);
+        assert.match(result.stdout, /managed plugin\/theme\/thumbnailer assets/i);
         assert.equal(
           fs.existsSync(path.join(smokeCase.xdgDataHome, 'SubMiner', 'themes', 'subminer.rasi')),
           true,
         );
+        assert.equal(
+          fs.existsSync(
+            path.join(
+              smokeCase.xdgDataHome,
+              'SubMiner',
+              'thumbnailers',
+              'subminer-ffmpegthumbnailer.thumbnailer',
+            ),
+          ),
+          true,
+        );
       }
     });
   },

launcher/test-support/immersion-db-fixture.ts

@@ -18,6 +18,9 @@ export function createImmersionDbFixture(dbPath: string): void {
     db.prepare(
       `INSERT INTO imm_rollup_state(state_key, state_value) VALUES ('last_rollup_sample_ms', 0)`,
     ).run();
+    db.prepare(
+      `INSERT INTO imm_rollup_state(state_key, state_value) VALUES ('lexical_daily_rollups_version', 0)`,
+    ).run();
     db.prepare(
       `INSERT INTO imm_lifetime_global(global_id, CREATED_DATE, LAST_UPDATE_DATE) VALUES (1, ?, ?)`,
     ).run(String(Date.now()), String(Date.now()));

launcher/test-support/immersion-db-schema.test.ts

@@ -108,6 +108,36 @@ test('fixture schema stays aligned with production sync-touched tables and index
   }
 });
 
+test('fixture leaves lexical rollups pending when their table is absent', () => {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-fixture-rollup-state-'));
+  const fixturePath = path.join(dir, 'fixture.sqlite');
+  try {
+    createImmersionDbFixture(fixturePath);
+    const db = new BunDatabase(fixturePath, { readonly: true });
+    try {
+      const state = db
+        .query<{ state_value: string }>(
+          `SELECT state_value FROM imm_rollup_state
+           WHERE state_key = 'lexical_daily_rollups_version'`,
+        )
+        .get();
+      const rollupTable = db
+        .query<{ name: string }>(
+          `SELECT name FROM sqlite_schema
+           WHERE type = 'table' AND name = 'imm_lexical_daily_rollups'`,
+        )
+        .get();
+
+      assert.equal(state?.state_value, '0');
+      assert.equal(rollupTable, null);
+    } finally {
+      db.close();
+    }
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true });
+  }
+});
+
 test('fixture session inserts enforce foreign keys', () => {
   const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-fixture-foreign-keys-'));
   const fixturePath = path.join(dir, 'fixture.sqlite');

launcher/test-support/immersion-db-schema.ts

@@ -154,6 +154,7 @@ export const IMMERSION_DB_FIXTURE_DDL = `
     last_seen REAL,
     frequency INTEGER,
     frequency_rank INTEGER,
+    vocabulary_visible INTEGER NOT NULL DEFAULT 1 CHECK(vocabulary_visible IN (0, 1)),
     UNIQUE(headword, word, reading)
   );
   CREATE TABLE imm_kanji(

package.json

@@ -2,7 +2,7 @@
   "name": "subminer",
   "productName": "SubMiner",
   "desktopName": "SubMiner.desktop",
-  "version": "0.19.4-beta.1",
+  "version": "0.19.4",
   "description": "All-in-one sentence mining overlay with AnkiConnect and dictionary integration",
   "packageManager": "bun@1.3.5",
   "main": "dist/main-entry.js",
@@ -32,6 +32,7 @@
     "changelog:pr-check": "bun run scripts/build-changelog.ts pr-check",
     "changelog:release-notes": "bun run scripts/build-changelog.ts release-notes",
     "changelog:prerelease-notes": "bun run scripts/build-changelog.ts prerelease-notes",
+    "changelog:check-prerelease-notes": "bun run scripts/build-changelog.ts check-prerelease-notes",
     "format": "prettier --write .",
     "format:check": "prettier --check .",
     "format:src": "bash scripts/prettier-scope.sh --write",

packaging/aur/subminer-bin/PKGBUILD

@@ -58,6 +58,8 @@ package() {
 		"${pkgdir}/usr/share/SubMiner/plugin/subminer.conf"
 	install -Dm644 "${srcdir}/assets/themes/subminer.rasi" \
 		"${pkgdir}/usr/share/SubMiner/themes/subminer.rasi"
+	install -Dm644 "${srcdir}/assets/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer" \
+		"${pkgdir}/usr/share/SubMiner/thumbnailers/subminer-ffmpegthumbnailer.thumbnailer"
 
 	install -dm755 "${pkgdir}/usr/share/SubMiner/plugin/subminer"
 	cp -a "${srcdir}/plugin/subminer/." "${pkgdir}/usr/share/SubMiner/plugin/subminer/"

plugin/subminer/process.lua

@@ -7,6 +7,8 @@ local OVERLAY_RESTART_PING_MAX_ATTEMPTS = 20
 local OVERLAY_LOADING_OSD_PREFIX = "Overlay loading "
 local OVERLAY_LOADING_OSD_FRAMES = { "|", "/", "-", "\\" }
 local OVERLAY_LOADING_OSD_REFRESH_SECONDS = 0.18
+local OVERLAY_LOADING_OSD_DEADLINE_SECONDS = 30
+local OVERLAY_LOADING_OSD_TIMEOUT_MESSAGE = "Overlay did not become ready; check SubMiner logs"
 local AUTO_PLAY_READY_LOADING_OSD = "Loading subtitle tokenization..."
 local AUTO_PLAY_READY_READY_OSD = "Subtitle tokenization ready"
 local DEFAULT_AUTO_PLAY_READY_TIMEOUT_SECONDS = 30
@@ -265,10 +267,19 @@ function M.create(ctx)
 		state.overlay_loading_osd_timer = nil
 	end
 
+	local function clear_overlay_loading_osd_deadline()
+		local timeout = state.overlay_loading_osd_deadline
+		if timeout and timeout.kill then
+			timeout:kill()
+		end
+		state.overlay_loading_osd_deadline = nil
+	end
+
 	local function stop_overlay_loading_osd()
 		state.overlay_loading_osd_active = false
 		state.overlay_loading_osd_frame = 1
 		clear_overlay_loading_osd_timer()
+		clear_overlay_loading_osd_deadline()
 	end
 
 	local function start_overlay_loading_osd()
@@ -291,6 +302,21 @@ function M.create(ctx)
 				end
 			end)
 		end
+		if type(mp.add_timeout) == "function" then
+			state.overlay_loading_osd_deadline = mp.add_timeout(OVERLAY_LOADING_OSD_DEADLINE_SECONDS, function()
+				if not state.overlay_loading_osd_active then
+					return
+				end
+				state.overlay_loading_osd_deadline = nil
+				stop_overlay_loading_osd()
+				subminer_log(
+					"warn",
+					"process",
+					"Overlay loading deadline expired before the app reported content ready"
+				)
+				show_osd(OVERLAY_LOADING_OSD_TIMEOUT_MESSAGE, { force = true })
+			end)
+		end
 	end
 
 	local function disarm_auto_play_ready_gate(options)

plugin/subminer/state.lua

@@ -26,6 +26,7 @@ function M.new()
 		auto_play_ready_initial_pause_ownership_consumed = false,
 		overlay_loading_osd_active = false,
 		overlay_loading_osd_timer = nil,
+		overlay_loading_osd_deadline = nil,
 		overlay_loading_osd_frame = 1,
 		pending_visible_overlay_hide_timer = nil,
 		pending_visible_overlay_hide_generation = 0,

release/prerelease-notes.md

@@ -4,37 +4,46 @@
 
 ## Highlights
 ### Added
-- **Library Merge and Move**
-  - Duplicate library cards for the same show can now be combined: select cards in the library grid, choose "Merge Selected," and pick which entry to keep. Sessions, mined cards, and watch time move over, and future episodes stay matched to the merged card.
-  - Episodes can be reassigned to a different library entry with a "→" button on the episode row, fixing cases where a file lands under the wrong title. Manual assignments survive later filename parsing, Jellyfin refreshes, and season repair.
-  - Exact AniList matches with compatible seasons now merge automatically, while fuzzy matches show up as a dismissible "Possible duplicate" prompt instead of merging without confirmation.
+
+- Library Merge & Reassignment
+  - Duplicate library cards for the same show can be combined: select entries in "Select" mode and use "Merge Selected" to combine their sessions, mined cards, and watch time onto one card.
+  - Episodes can be moved to a different entry with a per-episode "→" button, fixing stray files that split off their own entry; manual assignments now survive later filename parsing, Jellyfin refreshes, and season repair.
+  - Exact AniList matches with compatible seasons merge automatically, while likely (fuzzy) matches surface as a dismissible "Possible duplicate" suggestion instead of merging silently.
+
+- Duplicate Line Cleanup
+  - The Vocabulary tab's new **Duplicates** button scans a chosen time window for the repeated-line bursts described under Fixed below and collapses each burst to a single line once confirmed.
+  - A matching `subminer stats cleanup --duplicate-lines` command (with `--dry-run` and `--lookback-days <n>`) is available from the terminal.
+  - Only the affected subtitle lines and the vocabulary counts they inflated are touched; watch time and lines-seen totals are left as recorded.
 
 ### Fixed
-- **Anki Audio Generation on Network Drives**
-  - Fixed sentence-audio generation timing out on slow network-mounted video files with many subtitle and font-attachment streams.
-  - Extraction now uses bounded probing and a two-minute budget, and failures show a clear error instead of a cryptic one.
-- **Duplicate Subtitle Line Stats**
-  - Fixed karaoke openings and animated signs (which record one subtitle event per animation frame) inflating word and kanji counts and skewing "Top Repeated Words." Ordinary repeated dialogue and rewatches are unaffected.
-  - Already-inflated stats can be cleaned up with the new "Duplicates" button in the Vocabulary tab, or `subminer stats cleanup --duplicate-lines` (supports `--dry-run` and `--lookback-days`). Only the affected subtitle lines and vocabulary counts are touched; watch time and lines-seen totals are untouched.
-- **Overlay Modals on macOS and Windows**
-  - Fixed overlay modals and the stats window opening on the wrong macOS Space, or forcing a Space switch, when mpv is fullscreen. They now open above fullscreen mpv on its current Space.
-  - Modals are now prewarmed on macOS and Windows so shortcuts open them promptly, and Windows keeps the hidden modal responsive between sessions.
-- **Wayland File Drag-and-Drop**
-  - Fixed dragging subtitle and video files from file managers like Thunar onto the overlay on native Wayland; dropped files are now resolved and sent to mpv.
-- **Windows Mouse Lag**
-  - Fixed system-wide mouse lag while SubMiner is running on Windows, caused by a global mouse hook and blocking window lookups during click-through tracking.
-- **Mining Clip Accuracy**
-  - Fixed mined audio and animated image clips sometimes capturing the wrong subtitle line when audio extraction was slow. The clip range is now locked in at the moment of lookup, so audio and image clips always match.
-- **Linux Notifications**
-  - Character dictionary progress notifications on Linux now update in place instead of flickering off and back on with every status change.
-- **Stats Delete Performance**
-  - Fixed stats deletes freezing the dashboard; deletes now reliably run off the main thread, with automatic retry if the delete worker crashes.
-  - Deletes, library merges/moves, and AniList reassignments are now much faster because totals are updated incrementally instead of rebuilt from scratch, and no longer erase lifetime totals older than the recent session history.
-  - Session deletes on large libraries dropped from minutes to milliseconds.
 
-### Docs
-- **Feature Demos Page**
-  - Hidden the unfinished feature demos page from the documentation sidebar; it's still reachable by direct URL.
+- Subtitle Duplication from Karaoke & Animated Signs
+  - Typeset ASS karaoke and animated signs no longer flood the overlay, subtitle sidebar, immersion history, mined cards, or stats with repeated glyph fragments or per-frame duplicates; the complete authored line is recovered instead, without merging genuinely repeated dialogue or separately positioned signs.
+  - Fragmented karaoke now preserves the spaces the author placed between words instead of joining them together, and lyric transitions (including seeking into the middle of a line) resolve to the clean line instead of a stray entrance or exit frame.
+  - The secondary overlay shares the same deduplication logic as the primary overlay, including collapsing lines that differ only by whitespace or trailing punctuation, and sidebar navigation moves between clean lyric lines while keeping the right line selected.
+
+- Anki Media Generation
+  - Sentence-audio generation no longer times out on slow network-mounted video files with many subtitle and font streams, and a failed extraction now reports a clear error instead of a raw `ENOENT`.
+  - Mined audio and animated AVIF clips now capture the subtitle line you actually mined, instead of whatever line happened to be on screen once slow audio extraction finished.
+
+- Character Dictionary Performance & Notifications
+  - Character dictionary generation, merged rebuilds, and imports no longer freeze the app on large dictionaries, and cached results (including character portraits) are reused across launches instead of regenerating everything every time.
+  - Portraits also now display correctly if their cache finishes loading after subtitles have already started showing.
+  - Desktop progress notifications, including on Linux AppImage installs, now update in place instead of flickering closed and reopening.
+
+- Overlay Reliability
+  - Overlay modals (settings, stats, etc.) now open promptly on the first shortcut press, including on repeated sessions on Windows, and appear above fullscreen mpv on macOS instead of switching Spaces or opening off-screen.
+  - The macOS window-tracking helper is now built for macOS 12.0+, so the overlay attaches to mpv on older systems like Ventura instead of crashing and getting stuck on "Overlay loading."
+  - The overlay no longer gets stuck on "Overlay loading" indefinitely if mpv's connection stalls; it now retries and shows an actionable error after 30 seconds.
+  - Fixed native Wayland drag-and-drop from file managers like Thunar, and fixed system-wide mouse lag on Windows caused by the overlay's click-through handling.
+
+- Stats Dashboard
+  - Deletes, library merges, video moves, and AniList reassignments no longer freeze the stats dashboard or rebuild lifetime totals from scratch; large deletes that used to take minutes now finish in milliseconds.
+  - Vocabulary totals and charts now count all tracked vocabulary instead of just the first page, and new-word history uses corrected daily rollups.
+  - Calendar labels respect time zones west of UTC, and vocabulary cards refresh automatically after editing the word exclusion list (with a Retry option if a load fails).
+
+- Linux Launcher Thumbnails
+  - Fixed missing MKV thumbnails in the Linux rofi picker when the system thumbnailer only registers legacy Matroska MIME aliases.
 
 ## What's Changed
 
@@ -47,6 +56,13 @@
 - fix(overlay): support native Wayland file drag-and-drop by @ksyasuda in #199
 - fix(overlay): keep macOS modal windows on fullscreen Spaces by @ksyasuda in #200
 - fix(overlay): prevent Windows mouse lag during click-through tracking by @ksyasuda in #201
+- fix(stats): report complete vocabulary totals and new-word history by @ksyasuda in #202
+- fix(mpv): recover from stalled IPC connects by @ksyasuda in #204
+- fix(dictionary): prevent freezes and restore AppImage notifications by @ksyasuda in #205
+- fix(subtitles): recover canonical lines from ASS animation by @ksyasuda in #207
+- fix(overlay): deduplicate secondary subtitle rendering by @ksyasuda in #208
+- fix(launcher): restore Matroska thumbnails in Linux rofi picker by @ksyasuda in #210
+- fix(character-dictionary): cache completed MeCab refreshes by @ksyasuda in #212
 
 ## Installation
 

scripts/build-changelog.test.ts

@@ -1,4 +1,5 @@
 import assert from 'node:assert/strict';
+import { execFileSync } from 'node:child_process';
 import fs from 'node:fs';
 import path from 'node:path';
 import test from 'node:test';
@@ -43,14 +44,22 @@ function fragmentTypesInPrompt(input: string): string[] {
     .map((line) => line.slice('type: '.length).trim());
 }
 
-function assertReleaseNotesPromptRequestsNestedBullets(input: string): void {
-  assert.match(input, /In MODE: release-notes, use short top-level change bullets/);
-  assert.match(input, /Nested bullets should cover the change, user benefit, and any user action/);
-  assert.match(input, /Do not require the exact nested labels/);
+function assertPromptRequestsNestedBullets(input: string): void {
+  assert.match(input, /In both modes, split every item into one nested bullet per distinct change/);
+  assert.match(input, /Never stack several distinct changes into one long paragraph-shaped bullet/);
   assert.match(input, /Keep nested bullets short, concrete, and readable by non-technical users/);
   assert.match(input, /Avoid paragraph-style release-note bullets/);
 }
 
+function assertReleaseNotesPromptRequestsNestedBullets(input: string): void {
+  assertPromptRequestsNestedBullets(input);
+  assert.match(
+    input,
+    /In MODE: release-notes, nested bullets should also cover user benefit and any user action/,
+  );
+  assert.match(input, /Do not require the exact nested labels/);
+}
+
 function defaultPolishedBody(input: string): string {
   const mode = modeFromPrompt(input);
   const types = fragmentTypesInPrompt(input);
@@ -445,6 +454,7 @@ test('writeChangelogArtifacts prompts Claude to summarize the final stable outco
         prompt,
         /Multiple fixes within the same prerelease cycle should collapse into one current-state bullet/,
       );
+      assertPromptRequestsNestedBullets(prompt);
     }
 
     const releaseNotesPrompt = stub.calls.find(
@@ -583,7 +593,7 @@ test('writePrereleaseNotesForVersion writes cumulative beta notes without mutati
     const outputPath = writePrereleaseNotesForVersion({
       cwd: projectRoot,
       version: '0.11.3-beta.1',
-      deps: { runClaude: stub.runClaude },
+      deps: { runClaude: stub.runClaude, listPrereleaseTags: () => [] },
     });
 
     assert.equal(outputPath, path.join(projectRoot, 'release', 'prerelease-notes.md'));
@@ -605,7 +615,8 @@ test('writePrereleaseNotesForVersion writes cumulative beta notes without mutati
 
     const prereleaseNotes = fs.readFileSync(outputPath, 'utf8');
     assert.match(prereleaseNotes, /^> This is a prerelease build for testing\./m);
-    assert.match(prereleaseNotes, /<!-- prerelease-base-version: 0\.11\.3 -->/);
+    assert.match(prereleaseNotes, /<!-- prerelease-version: 0\.11\.3-beta\.1 -->/);
+    assert.doesNotMatch(prereleaseNotes, /## Changes since /);
     assert.match(prereleaseNotes, /## Highlights\n### Added\n- Polished: added entry\./);
     assert.match(prereleaseNotes, /### Fixed\n- Polished: fixed entry\./);
     assert.match(prereleaseNotes, /## Installation\n\nSee the README and docs\/installation guide/);
@@ -668,7 +679,7 @@ test('writePrereleaseNotesForVersion reuses existing prerelease notes when addin
     const outputPath = writePrereleaseNotesForVersion({
       cwd: projectRoot,
       version: '0.11.3-beta.2',
-      deps: { runClaude: stub.runClaude },
+      deps: { runClaude: stub.runClaude, listPrereleaseTags: () => [] },
     });
 
     assert.equal(stub.calls.length, 1, 'prerelease should issue exactly one Claude call');
@@ -723,7 +734,7 @@ test('writePrereleaseNotesForVersion ignores unmarked prerelease notes from an o
     const outputPath = writePrereleaseNotesForVersion({
       cwd: projectRoot,
       version: '0.17.0-beta.1',
-      deps: { runClaude: stub.runClaude },
+      deps: { runClaude: stub.runClaude, listPrereleaseTags: () => [] },
     });
 
     assert.equal(stub.calls.length, 1, 'prerelease should issue exactly one Claude call');
@@ -790,7 +801,7 @@ test('writePrereleaseNotesForVersion prompts Claude to revise stale prerelease b
     writePrereleaseNotesForVersion({
       cwd: projectRoot,
       version: '0.12.0-beta.2',
-      deps: { runClaude: stub.runClaude },
+      deps: { runClaude: stub.runClaude, listPrereleaseTags: () => [] },
     });
 
     assert.equal(stub.calls.length, 1, 'prerelease should issue exactly one Claude call');
@@ -830,7 +841,7 @@ test('writePrereleaseNotesForVersion supports rc prereleases', async () => {
     const outputPath = writePrereleaseNotesForVersion({
       cwd: projectRoot,
       version: '0.11.3-rc.1',
-      deps: { runClaude: stub.runClaude },
+      deps: { runClaude: stub.runClaude, listPrereleaseTags: () => [] },
     });
 
     const prereleaseNotes = fs.readFileSync(outputPath, 'utf8');
@@ -1447,3 +1458,373 @@ test('writeChangelogArtifacts strips <details> blocks from release notes when re
     fs.rmSync(workspace, { recursive: true, force: true });
   }
 });
+
+test('selectPreviousPrereleaseTag orders betas before rcs and filters other base versions', async () => {
+  const { selectPreviousPrereleaseTag } = await loadModule();
+
+  const tags = [
+    'v0.19.4-beta.1',
+    'v0.19.4-beta.3',
+    'v0.19.4-beta.2',
+    'v0.19.3-beta.9',
+    'v0.19.4-rc.1',
+    'not-a-tag',
+  ];
+
+  assert.equal(selectPreviousPrereleaseTag(tags, '0.19.4-beta.1'), null);
+  assert.equal(selectPreviousPrereleaseTag(tags, '0.19.4-beta.2'), 'v0.19.4-beta.1');
+  assert.equal(selectPreviousPrereleaseTag(tags, '0.19.4-beta.4'), 'v0.19.4-beta.3');
+  assert.equal(selectPreviousPrereleaseTag(tags, '0.19.4-rc.1'), 'v0.19.4-beta.3');
+  assert.equal(selectPreviousPrereleaseTag(tags, '0.19.4-rc.2'), 'v0.19.4-rc.1');
+  // Regenerating notes for an already-tagged version must not pick itself.
+  assert.equal(selectPreviousPrereleaseTag(tags, '0.19.4-beta.3'), 'v0.19.4-beta.2');
+  assert.equal(selectPreviousPrereleaseTag(['v0.19.3-beta.1'], '0.19.4-beta.2'), null);
+});
+
+test('writePrereleaseNotesForVersion adds a delta section generated from fragment diffs', async () => {
+  const { writePrereleaseNotesForVersion } = await loadModule();
+  const workspace = createWorkspace('prerelease-delta-section');
+  const projectRoot = path.join(workspace, 'SubMiner');
+
+  fs.mkdirSync(path.join(projectRoot, 'changes'), { recursive: true });
+  fs.writeFileSync(
+    path.join(projectRoot, 'package.json'),
+    JSON.stringify({ name: 'subminer', version: '0.12.0-beta.2' }, null, 2),
+    'utf8',
+  );
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', '001.md'),
+    ['type: fixed', 'area: overlay', '', '- Fixed overlay focus and macOS helper.'].join('\n'),
+    'utf8',
+  );
+
+  try {
+    const stub = recordingRunClaude((input) =>
+      input.includes('MODIFIED FRAGMENT')
+        ? '- Fixed the macOS helper deployment target for older systems.'
+        : '### Fixed\n- Overlay: cumulative fixed entry.',
+    );
+    const outputPath = writePrereleaseNotesForVersion({
+      cwd: projectRoot,
+      version: '0.12.0-beta.2',
+      deps: {
+        runClaude: stub.runClaude,
+        listPrereleaseTags: () => ['v0.12.0-beta.1'],
+        resolveFragmentDelta: (_cwd, previousTag) => {
+          assert.equal(previousTag, 'v0.12.0-beta.1');
+          return [
+            {
+              path: 'changes/002.md',
+              status: 'added',
+              after: 'type: fixed\narea: macos\n\n- Fixed helper deployment target.',
+            },
+            {
+              path: 'changes/001.md',
+              status: 'modified',
+              before: '- Fixed overlay focus.',
+              after: '- Fixed overlay focus and macOS helper.',
+            },
+            {
+              path: 'changes/003.md',
+              status: 'deleted',
+              before: 'type: added\narea: stats\n\n- Reverted experimental stats view.',
+            },
+          ];
+        },
+      },
+    });
+
+    assert.equal(stub.calls.length, 2, 'delta and cumulative polish are separate Claude calls');
+    const deltaPrompt = stub.calls[0]!.input;
+    assert.match(deltaPrompt, /ADDED FRAGMENT changes\/002\.md/);
+    assert.match(deltaPrompt, /MODIFIED FRAGMENT changes\/001\.md/);
+    assert.match(deltaPrompt, /BEFORE:\n- Fixed overlay focus\./);
+    assert.match(deltaPrompt, /AFTER:\n- Fixed overlay focus and macOS helper\./);
+    assert.match(deltaPrompt, /DELETED FRAGMENT changes\/003\.md/);
+    assert.match(deltaPrompt, /If the edit is editorial/);
+    assert.match(deltaPrompt, /removed or reverted/);
+    assert.match(deltaPrompt, /No user-facing changes since v0\.12\.0-beta\.1\./);
+    assert.equal(modeFromPrompt(stub.calls[1]!.input), 'release-notes');
+
+    const prereleaseNotes = fs.readFileSync(outputPath, 'utf8');
+    assert.match(
+      prereleaseNotes,
+      /<!-- prerelease-version: 0\.12\.0-beta\.2; since: v0\.12\.0-beta\.1 -->/,
+    );
+    const deltaIndex = prereleaseNotes.indexOf('## Changes since v0.12.0-beta.1');
+    const highlightsIndex = prereleaseNotes.indexOf('## Highlights');
+    assert.ok(deltaIndex !== -1, 'delta section heading should be present');
+    assert.ok(deltaIndex < highlightsIndex, 'delta section should precede Highlights');
+    assert.match(prereleaseNotes, /- Fixed the macOS helper deployment target for older systems\./);
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});
+
+test('writePrereleaseNotesForVersion renders a fallback delta line when no fragments changed', async () => {
+  const { writePrereleaseNotesForVersion } = await loadModule();
+  const workspace = createWorkspace('prerelease-empty-delta');
+  const projectRoot = path.join(workspace, 'SubMiner');
+
+  fs.mkdirSync(path.join(projectRoot, 'changes'), { recursive: true });
+  fs.writeFileSync(
+    path.join(projectRoot, 'package.json'),
+    JSON.stringify({ name: 'subminer', version: '0.12.0-beta.3' }, null, 2),
+    'utf8',
+  );
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', '001.md'),
+    ['type: fixed', 'area: overlay', '', '- Fixed overlay focus.'].join('\n'),
+    'utf8',
+  );
+
+  try {
+    const stub = defaultStubClaude();
+    const outputPath = writePrereleaseNotesForVersion({
+      cwd: projectRoot,
+      version: '0.12.0-beta.3',
+      deps: {
+        runClaude: stub.runClaude,
+        listPrereleaseTags: () => ['v0.12.0-beta.1', 'v0.12.0-beta.2'],
+        resolveFragmentDelta: () => [],
+      },
+    });
+
+    assert.equal(stub.calls.length, 1, 'empty delta must not spend a Claude call');
+    const prereleaseNotes = fs.readFileSync(outputPath, 'utf8');
+    assert.match(
+      prereleaseNotes,
+      /## Changes since v0\.12\.0-beta\.2\n\n- No changelog fragment changes since v0\.12\.0-beta\.2; this build contains packaging or internal-only updates\./,
+    );
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});
+
+test('writePrereleaseNotesForVersion rejects non-bullet delta output from Claude', async () => {
+  const { writePrereleaseNotesForVersion } = await loadModule();
+  const workspace = createWorkspace('prerelease-delta-invalid-output');
+  const projectRoot = path.join(workspace, 'SubMiner');
+
+  fs.mkdirSync(path.join(projectRoot, 'changes'), { recursive: true });
+  fs.writeFileSync(
+    path.join(projectRoot, 'package.json'),
+    JSON.stringify({ name: 'subminer', version: '0.12.0-beta.2' }, null, 2),
+    'utf8',
+  );
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', '001.md'),
+    ['type: fixed', 'area: overlay', '', '- Fixed overlay focus.'].join('\n'),
+    'utf8',
+  );
+
+  try {
+    const stub = recordingRunClaude(() => 'Here are the changes:\n- One change.');
+    assert.throws(
+      () =>
+        writePrereleaseNotesForVersion({
+          cwd: projectRoot,
+          version: '0.12.0-beta.2',
+          deps: {
+            runClaude: stub.runClaude,
+            listPrereleaseTags: () => ['v0.12.0-beta.1'],
+            resolveFragmentDelta: () => [
+              { path: 'changes/001.md', status: 'added', after: '- Fixed overlay focus.' },
+            ],
+          },
+        }),
+      /delta output must contain only Markdown bullets/,
+    );
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});
+
+test('writePrereleaseNotesForVersion strips the stale delta section from the reused baseline', async () => {
+  const { writePrereleaseNotesForVersion } = await loadModule();
+  const workspace = createWorkspace('prerelease-reuse-strips-delta');
+  const projectRoot = path.join(workspace, 'SubMiner');
+  const existingNotes = [
+    '> This is a prerelease build for testing. Stable changelog and docs-site updates remain pending until the final stable release.',
+    '',
+    '<!-- prerelease-version: 0.12.0-beta.2; since: v0.12.0-beta.1 -->',
+    '',
+    '## Changes since v0.12.0-beta.1',
+    '',
+    '- Stale beta-to-beta delta bullet.',
+    '',
+    '## Highlights',
+    '### Added',
+    '- Overlay: Previous beta entry.',
+    '',
+    '## Installation',
+    '',
+    'See the README and docs/installation guide for full setup steps.',
+    '',
+  ].join('\n');
+
+  fs.mkdirSync(path.join(projectRoot, 'changes'), { recursive: true });
+  fs.mkdirSync(path.join(projectRoot, 'release'), { recursive: true });
+  fs.writeFileSync(
+    path.join(projectRoot, 'package.json'),
+    JSON.stringify({ name: 'subminer', version: '0.12.0-beta.3' }, null, 2),
+    'utf8',
+  );
+  fs.writeFileSync(path.join(projectRoot, 'release', 'prerelease-notes.md'), existingNotes, 'utf8');
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', '001.md'),
+    ['type: added', 'area: overlay', '', '- Added overlay coverage.'].join('\n'),
+    'utf8',
+  );
+
+  try {
+    const stub = defaultStubClaude();
+    writePrereleaseNotesForVersion({
+      cwd: projectRoot,
+      version: '0.12.0-beta.3',
+      deps: {
+        runClaude: stub.runClaude,
+        listPrereleaseTags: () => [],
+        resolveFragmentDelta: () => [],
+      },
+    });
+
+    assert.equal(stub.calls.length, 1);
+    const prompt = stub.calls[0]!.input;
+    assert.match(prompt, /EXISTING PRERELEASE NOTES/);
+    assert.match(prompt, /Overlay: Previous beta entry\./);
+    assert.doesNotMatch(prompt, /Stale beta-to-beta delta bullet\./);
+    assert.doesNotMatch(prompt, /## Changes since /);
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});
+
+test('verifyPrereleaseNotesMatchVersion accepts matching notes and rejects stale or legacy markers', async () => {
+  const { verifyPrereleaseNotesMatchVersion } = await loadModule();
+  const workspace = createWorkspace('verify-prerelease-notes');
+  const projectRoot = path.join(workspace, 'SubMiner');
+  const notesPath = path.join(projectRoot, 'release', 'prerelease-notes.md');
+
+  fs.mkdirSync(path.join(projectRoot, 'release'), { recursive: true });
+  fs.writeFileSync(
+    path.join(projectRoot, 'package.json'),
+    JSON.stringify({ name: 'subminer', version: '0.12.0-beta.2' }, null, 2),
+    'utf8',
+  );
+
+  try {
+    assert.throws(
+      () => verifyPrereleaseNotesMatchVersion({ cwd: projectRoot, version: '0.12.0-beta.2' }),
+      /Missing .*prerelease-notes\.md/,
+    );
+
+    fs.writeFileSync(
+      notesPath,
+      '<!-- prerelease-version: 0.12.0-beta.2; since: v0.12.0-beta.1 -->\n\n## Highlights\n',
+      'utf8',
+    );
+    verifyPrereleaseNotesMatchVersion({ cwd: projectRoot, version: '0.12.0-beta.2' });
+    verifyPrereleaseNotesMatchVersion({ cwd: projectRoot, version: 'v0.12.0-beta.2' });
+
+    fs.writeFileSync(
+      notesPath,
+      '<!-- prerelease-version: 0.12.0-beta.1 -->\n\n## Highlights\n',
+      'utf8',
+    );
+    assert.throws(
+      () => verifyPrereleaseNotesMatchVersion({ cwd: projectRoot, version: '0.12.0-beta.2' }),
+      /generated for 0\.12\.0-beta\.1 but this release is 0\.12\.0-beta\.2/,
+    );
+
+    fs.writeFileSync(
+      notesPath,
+      '<!-- prerelease-base-version: 0.12.0 -->\n\n## Highlights\n',
+      'utf8',
+    );
+    assert.throws(
+      () => verifyPrereleaseNotesMatchVersion({ cwd: projectRoot, version: '0.12.0-beta.2' }),
+      /missing or legacy prerelease-version marker/,
+    );
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});
+
+test('default git tag listing and fragment delta resolution work against a real repository', async () => {
+  const { writePrereleaseNotesForVersion } = await loadModule();
+  const workspace = createWorkspace('prerelease-git-defaults');
+  const projectRoot = path.join(workspace, 'SubMiner');
+  const git = (...args: string[]): void => {
+    execFileSync('git', args, { cwd: projectRoot, stdio: 'ignore' });
+  };
+
+  fs.mkdirSync(path.join(projectRoot, 'changes'), { recursive: true });
+  fs.writeFileSync(
+    path.join(projectRoot, 'package.json'),
+    JSON.stringify({ name: 'subminer', version: '0.11.3-beta.1' }, null, 2),
+    'utf8',
+  );
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', 'kept.md'),
+    ['type: added', 'area: overlay', '', '- Kept change.'].join('\n'),
+    'utf8',
+  );
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', 'edited.md'),
+    ['type: fixed', 'area: launcher', '', '- Original launcher fix.'].join('\n'),
+    'utf8',
+  );
+  fs.writeFileSync(
+    path.join(projectRoot, 'changes', 'removed.md'),
+    ['type: added', 'area: stats', '', '- Reverted stats change.'].join('\n'),
+    'utf8',
+  );
+
+  try {
+    git('init', '--quiet');
+    git('-c', 'user.email=test@example.com', '-c', 'user.name=Test', 'add', '.');
+    git('-c', 'user.email=test@example.com', '-c', 'user.name=Test', 'commit', '-m', 'beta.1');
+    git('tag', 'v0.11.3-beta.1');
+
+    fs.writeFileSync(
+      path.join(projectRoot, 'changes', 'edited.md'),
+      ['type: fixed', 'area: launcher', '', '- Broader launcher fix.'].join('\n'),
+      'utf8',
+    );
+    fs.rmSync(path.join(projectRoot, 'changes', 'removed.md'));
+    fs.writeFileSync(
+      path.join(projectRoot, 'changes', 'new.md'),
+      ['type: added', 'area: anki', '', '- New anki change.'].join('\n'),
+      'utf8',
+    );
+    fs.writeFileSync(
+      path.join(projectRoot, 'package.json'),
+      JSON.stringify({ name: 'subminer', version: '0.11.3-beta.2' }, null, 2),
+      'utf8',
+    );
+
+    const stub = recordingRunClaude((input) =>
+      input.includes('PREVIOUS_TAG:') ? '- Delta bullet.' : defaultPolishedBody(input),
+    );
+    writePrereleaseNotesForVersion({
+      cwd: projectRoot,
+      version: '0.11.3-beta.2',
+      deps: { runClaude: stub.runClaude },
+    });
+
+    assert.equal(stub.calls.length, 2);
+    const deltaPrompt = stub.calls[0]!.input;
+    assert.match(deltaPrompt, /PREVIOUS_TAG: v0\.11\.3-beta\.1/);
+    assert.match(deltaPrompt, /ADDED FRAGMENT changes\/new\.md/);
+    assert.match(deltaPrompt, /- New anki change\./);
+    assert.match(deltaPrompt, /MODIFIED FRAGMENT changes\/edited\.md/);
+    assert.match(deltaPrompt, /- Original launcher fix\./);
+    assert.match(deltaPrompt, /- Broader launcher fix\./);
+    assert.match(deltaPrompt, /DELETED FRAGMENT changes\/removed\.md/);
+    assert.match(deltaPrompt, /- Reverted stats change\./);
+    assert.doesNotMatch(deltaPrompt, /kept\.md/);
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});

scripts/build-changelog.ts

@@ -18,6 +18,15 @@ type Contribution = {
 // and the GitHub API.
 type ResolveContributions = (fragmentPaths: string[], cwd: string) => Contribution[];
 
+// One changelog fragment's change between the previous prerelease tag and the
+// working tree. `before` is the content at the tag, `after` the current content.
+export type FragmentDeltaEntry = {
+  path: string;
+  status: 'added' | 'modified' | 'deleted';
+  before?: string;
+  after?: string;
+};
+
 type ChangelogFsDeps = {
   existsSync?: (candidate: string) => boolean;
   mkdirSync?: (candidate: string, options: { recursive: true }) => void;
@@ -28,6 +37,8 @@ type ChangelogFsDeps = {
   log?: (message: string) => void;
   runClaude?: RunClaude;
   resolveContributions?: ResolveContributions;
+  listPrereleaseTags?: (cwd: string, baseVersion: string) => string[];
+  resolveFragmentDelta?: (cwd: string, previousTag: string) => FragmentDeltaEntry[];
 };
 
 type PolishMode = 'changelog' | 'release-notes';
@@ -103,16 +114,57 @@ function resolvePrereleaseBaseVersion(version: string): string {
   return match[1]!;
 }
 
-function renderPrereleaseBaseVersionMarker(version: string): string {
-  return `<!-- prerelease-base-version: ${resolvePrereleaseBaseVersion(version)} -->`;
+// The marker records which exact prerelease the committed notes were generated
+// for (and which prior tag the delta section compares against), so CI can
+// reject notes that were prepared for a different beta/RC.
+function renderPrereleaseVersionMarker(version: string, previousTag: string | null): string {
+  const since = previousTag ? `; since: ${previousTag}` : '';
+  return `<!-- prerelease-version: ${normalizeVersion(version)}${since} -->`;
 }
 
+export function extractPrereleaseVersionMarker(notes: string): string | null {
+  return (
+    /<!--\s*prerelease-version:\s*(\d+\.\d+\.\d+-(?:beta|rc)\.\d+)(?:;\s*since:\s*\S+)?\s*-->/u.exec(
+      notes,
+    )?.[1] ?? null
+  );
+}
+
+// Legacy marker written before the per-version marker existed. Still accepted
+// when deciding whether existing notes can seed the cumulative baseline.
 function extractPrereleaseBaseVersionMarker(notes: string): string | null {
+  const fullVersion = extractPrereleaseVersionMarker(notes);
+  if (fullVersion) {
+    return resolvePrereleaseBaseVersion(fullVersion);
+  }
   return /<!--\s*prerelease-base-version:\s*(\d+\.\d+\.\d+)\s*-->/u.exec(notes)?.[1] ?? null;
 }
 
+const DELTA_SECTION_HEADING_PREFIX = '## Changes since ';
+
+// Removes the previous run's "Changes since" section so the cumulative baseline
+// fed back to Claude never carries a stale beta-to-beta delta.
+function stripDeltaSection(notes: string): string {
+  const lines = notes.split(/\r?\n/);
+  const start = lines.findIndex((line) => line.startsWith(DELTA_SECTION_HEADING_PREFIX));
+  if (start === -1) {
+    return notes;
+  }
+  let end = lines.length;
+  for (let index = start + 1; index < lines.length; index += 1) {
+    if (lines[index]!.startsWith('## ')) {
+      end = index;
+      break;
+    }
+  }
+  return [...lines.slice(0, start), ...lines.slice(end)].join('\n');
+}
+
 function stripPrereleaseMetadata(notes: string): string {
-  return notes.replace(/<!--\s*prerelease-base-version:\s*\d+\.\d+\.\d+\s*-->\s*/u, '').trim();
+  return notes
+    .replace(/<!--\s*prerelease-version:[^>]*-->\s*/u, '')
+    .replace(/<!--\s*prerelease-base-version:\s*\d+\.\d+\.\d+\s*-->\s*/u, '')
+    .trim();
 }
 
 function resolveReusablePrereleaseNotes(notes: string, version: string): string | undefined {
@@ -120,7 +172,124 @@ function resolveReusablePrereleaseNotes(notes: string, version: string): string
   if (existingBaseVersion !== resolvePrereleaseBaseVersion(version)) {
     return undefined;
   }
-  return stripPrereleaseMetadata(notes);
+  return stripPrereleaseMetadata(stripDeltaSection(notes));
+}
+
+type ParsedPrereleaseTag = {
+  tag: string;
+  base: string;
+  channel: 'beta' | 'rc';
+  iteration: number;
+};
+
+function parsePrereleaseTag(tag: string): ParsedPrereleaseTag | null {
+  const match = /^v?(\d+\.\d+\.\d+)-(beta|rc)\.(\d+)$/u.exec(tag.trim());
+  if (!match) {
+    return null;
+  }
+  return {
+    tag: tag.trim(),
+    base: match[1]!,
+    channel: match[2] as 'beta' | 'rc',
+    iteration: Number.parseInt(match[3]!, 10),
+  };
+}
+
+// Semver prerelease order: every beta sorts before every rc, then numerically.
+function comparePrereleaseTags(a: ParsedPrereleaseTag, b: ParsedPrereleaseTag): number {
+  if (a.channel !== b.channel) {
+    return a.channel === 'beta' ? -1 : 1;
+  }
+  return a.iteration - b.iteration;
+}
+
+// Picks the newest prerelease tag for the same base version that strictly
+// precedes the version being released. Returns null for the first prerelease.
+export function selectPreviousPrereleaseTag(tags: string[], version: string): string | null {
+  const current = parsePrereleaseTag(normalizeVersion(version));
+  if (!current) {
+    return null;
+  }
+
+  const candidates = tags
+    .map(parsePrereleaseTag)
+    .filter((parsed): parsed is ParsedPrereleaseTag => parsed !== null)
+    .filter((parsed) => parsed.base === current.base)
+    .filter((parsed) => comparePrereleaseTags(parsed, current) < 0)
+    .sort(comparePrereleaseTags);
+
+  return candidates[candidates.length - 1]?.tag ?? null;
+}
+
+function defaultListPrereleaseTags(cwd: string, baseVersion: string): string[] {
+  return execFileSync('git', ['tag', '--list', `v${baseVersion}-beta.*`, `v${baseVersion}-rc.*`], {
+    cwd,
+    encoding: 'utf8',
+  })
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter(Boolean);
+}
+
+// Diffs changes/*.md between the previous prerelease tag and the working tree.
+// Renamed fragments are treated as modifications of the new path.
+//
+// Like every other path in this script, git paths are resolved against `cwd`,
+// which is the project root and also the repository root. Callers that point
+// `cwd` elsewhere already fail earlier and loudly, when package.json and
+// changes/ come back missing.
+function defaultResolveFragmentDelta(cwd: string, previousTag: string): FragmentDeltaEntry[] {
+  const output = execFileSync(
+    'git',
+    ['diff', '--name-status', '--find-renames', previousTag, '--', 'changes'],
+    { cwd, encoding: 'utf8' },
+  );
+  const showAtTag = (fragmentPath: string): string =>
+    execFileSync('git', ['show', `${previousTag}:${fragmentPath}`], { cwd, encoding: 'utf8' });
+  const readCurrent = (fragmentPath: string): string =>
+    fs.readFileSync(path.join(cwd, fragmentPath), 'utf8');
+
+  const entries: FragmentDeltaEntry[] = [];
+  for (const line of output.split(/\r?\n/)) {
+    if (!line.trim()) {
+      continue;
+    }
+    const [status = '', ...paths] = line.split('\t');
+    const oldPath = paths[0] ?? '';
+    const newPath = paths[paths.length - 1] ?? '';
+    if (!isFragmentPath(newPath) && !isFragmentPath(oldPath)) {
+      continue;
+    }
+
+    if (status.startsWith('A')) {
+      entries.push({ path: newPath, status: 'added', after: readCurrent(newPath) });
+    } else if (status.startsWith('D')) {
+      entries.push({ path: oldPath, status: 'deleted', before: showAtTag(oldPath) });
+    } else if (status.startsWith('M') || status.startsWith('R')) {
+      entries.push({
+        path: newPath,
+        status: 'modified',
+        before: showAtTag(oldPath),
+        after: readCurrent(newPath),
+      });
+    }
+  }
+
+  // git diff misses fragments that exist only in the working tree; treat
+  // untracked fragments as additions so a pre-commit run still sees them.
+  const untracked = execFileSync(
+    'git',
+    ['ls-files', '--others', '--exclude-standard', '--', 'changes'],
+    { cwd, encoding: 'utf8' },
+  )
+    .split(/\r?\n/)
+    .map((line) => line.trim())
+    .filter((candidate) => candidate && isFragmentPath(candidate));
+  for (const fragmentPath of untracked) {
+    entries.push({ path: fragmentPath, status: 'added', after: readCurrent(fragmentPath) });
+  }
+
+  return entries;
 }
 
 function verifyRequestedVersionMatchesPackageVersion(
@@ -311,10 +480,15 @@ You will receive a list of FRAGMENT entries below. Each fragment has metadata (t
    - Be merged with related bullets when possible. If five fragments all touch Windows overlay z-order/focus/restore, write one or two bullets that summarize the overall improvement instead of five.
    - Drop bullets that only describe PR housekeeping, CodeRabbit follow-ups, or test-only changes that don't affect users.
    - Preserve the substance of breaking changes that remain breaking after applying the Release Outcome Rules. Do not soften or omit them.
-5. In MODE: changelog, each item may be a conventional single-level bullet, e.g. "- Playlist Browser: Adds faster saved-show browsing."
-6. In MODE: release-notes, use short top-level change bullets with two or three nested bullets when an item needs explanation.
-   Nested bullets should cover the change, user benefit, and any user action or compatibility note when useful. Do not require the exact nested labels; natural phrasing is fine. Omit the action bullet when no action is needed.
+5. In both modes, split every item into one nested bullet per distinct change. Write a short bold name on the top-level bullet, then indent the details two spaces:
+   - **Playlist Browser**:
+     - Saved shows now open without rescanning the library.
+     - The picker remembers the last folder you browsed between launches.
+   Each nested bullet covers exactly one change, behavior, or user-visible outcome. Never stack several distinct changes into one long paragraph-shaped bullet.
+   Aim for two to five nested bullets per item. When an item genuinely has only one thing to say, put it inline on the top-level bullet ("- **Playlist Browser**: Saved shows now open without rescanning the library.") instead of emitting a single nested bullet.
    Keep nested bullets short, concrete, and readable by non-technical users. Avoid paragraph-style release-note bullets.
+   Bullets inside the Internal section may stay single-level.
+6. In MODE: release-notes, nested bullets should also cover user benefit and any user action or compatibility note when useful. Do not require the exact nested labels; natural phrasing is fine. Omit the action bullet when no action is needed.
 7. Do not invent features. Every bullet must be grounded in the input fragments.
 8. Do not include the version heading (## v...) — that wrapper is added by the caller.
 
@@ -615,7 +789,7 @@ function polishFragmentsWithClaude(
     ? [
         '## Existing Prerelease Notes',
         '',
-        'The input includes EXISTING PRERELEASE NOTES before the fragment list. Existing prerelease notes are a baseline, not an immutable changelog. Reuse reviewed highlight bullets when they still describe the current outcome, but replace stale beta or RC wording when new fragments supersede it. Merge in only new or changed fragment material, and deduplicate instead of restating existing bullets. Output only the final highlights body using the section headings above; do not include the prerelease disclaimer, Installation, or Assets sections.',
+        'The input includes EXISTING PRERELEASE NOTES before the fragment list. Existing prerelease notes are a baseline, not an immutable changelog. Reuse reviewed highlight bullets when they still describe the current outcome, but replace stale beta or RC wording when new fragments supersede it. Merge in only new or changed fragment material, and deduplicate instead of restating existing bullets. Output only the final highlights body using the section headings above; do not include the prerelease disclaimer, any "Changes since" section, or the Installation or Assets sections.',
         '',
       ].join('\n')
     : '';
@@ -627,6 +801,75 @@ function polishFragmentsWithClaude(
   return validatePolishedOutput(output, mode, hasInternalFragments);
 }
 
+const DELTA_PROMPT_INSTRUCTIONS = `You are writing the "changes since the previous prerelease" section of a prerelease notes file for SubMiner, an Electron app for Japanese sentence mining.
+
+You will receive changelog fragment diffs between the previous prerelease tag and the current build. Fragments are engineer-written release-note sources; a fragment diff is a proxy for what changed, not proof of a behavior change.
+
+Rules:
+
+1. Output Markdown bullets ONLY. No headings, no preamble, no commentary. Every line must be a top-level "- " bullet or an indented nested bullet.
+2. Describe only what changed for users between the two prerelease builds, in user-facing language. Drop implementation jargon, file paths, and PR numbers.
+3. ADDED fragments describe changes that are new in this build; summarize them.
+4. MODIFIED fragments include BEFORE and AFTER content. Describe only the behavioral difference between them. If the edit is editorial (rewording, deduplication, reformatting, reconciling stale phrasing) with no user-visible behavior change, omit it entirely.
+5. DELETED fragments mean the described change was removed or reverted before this build; say so explicitly.
+6. Keep bullets short and concrete. Use nested bullets sparingly.
+7. Do not invent changes. Every bullet must be grounded in the diffs.
+8. If no bullet survives rules 2-5, output exactly this single line:
+   - No user-facing changes since PREVIOUS_TAG.
+
+The input begins below.
+
+`;
+
+function serializeFragmentDeltaForPrompt(
+  delta: FragmentDeltaEntry[],
+  version: string,
+  previousTag: string,
+): string {
+  const header = [`VERSION: ${version}`, `PREVIOUS_TAG: ${previousTag}`];
+  const blocks = delta.map((entry) => {
+    if (entry.status === 'added') {
+      return [`ADDED FRAGMENT ${entry.path}`, entry.after ?? ''].join('\n');
+    }
+    if (entry.status === 'deleted') {
+      return [`DELETED FRAGMENT ${entry.path}`, entry.before ?? ''].join('\n');
+    }
+    return [
+      `MODIFIED FRAGMENT ${entry.path}`,
+      'BEFORE:',
+      entry.before ?? '',
+      'AFTER:',
+      entry.after ?? '',
+    ].join('\n');
+  });
+  return [...header, '', ...blocks].join('\n\n');
+}
+
+function validateDeltaOutput(output: string): string {
+  const trimmed = output.trim();
+  if (!trimmed) {
+    throw new Error('claude returned empty output for the prerelease delta section.');
+  }
+  const invalidLine = trimmed.split(/\r?\n/).find((line) => line.trim() && !/^\s*- /.test(line));
+  if (invalidLine !== undefined) {
+    throw new Error(
+      `claude delta output must contain only Markdown bullets. Offending line:\n${invalidLine}`,
+    );
+  }
+  return trimmed;
+}
+
+function buildDeltaSectionWithClaude(
+  delta: FragmentDeltaEntry[],
+  options: { version: string; previousTag: string; deps?: ChangelogFsDeps },
+): string {
+  const runClaude = options.deps?.runClaude ?? defaultRunClaude;
+  const prompt =
+    DELTA_PROMPT_INSTRUCTIONS.replace('PREVIOUS_TAG', options.previousTag) +
+    serializeFragmentDeltaForPrompt(delta, options.version, options.previousTag);
+  return validateDeltaOutput(runClaude(prompt, CLAUDE_CLI_ARGS));
+}
+
 function stripDetailsBlocks(body: string): string {
   return body.replace(/<details>[\s\S]*?<\/details>\s*/gm, '').trim();
 }
@@ -709,15 +952,18 @@ function renderReleaseNotes(
     contributions?: Contribution[];
     contributorSections?: string[];
     metadata?: string[];
+    deltaSection?: string[];
   },
 ): string {
   const prefix = options?.disclaimer ? [options.disclaimer, ''] : [];
   const metadata = options?.metadata?.length ? [...options.metadata, ''] : [];
+  const deltaSection = options?.deltaSection?.length ? [...options.deltaSection, ''] : [];
   const contributorSections =
     options?.contributorSections ?? renderContributorsSections(options?.contributions ?? []);
   return [
     ...prefix,
     ...metadata,
+    ...deltaSection,
     '## Highlights',
     changes,
     '',
@@ -748,6 +994,7 @@ function writeReleaseNotesFile(
     contributions?: Contribution[];
     contributorSections?: string[];
     metadata?: string[];
+    deltaSection?: string[];
   },
 ): string {
   const mkdirSync = deps?.mkdirSync ?? fs.mkdirSync;
@@ -1079,6 +1326,26 @@ export function writePrereleaseNotesForVersion(options?: ChangelogOptions): stri
     throw new Error('No changelog fragments found in changes/.');
   }
 
+  const listPrereleaseTags = options?.deps?.listPrereleaseTags ?? defaultListPrereleaseTags;
+  const previousTag = selectPreviousPrereleaseTag(
+    listPrereleaseTags(cwd, resolvePrereleaseBaseVersion(version)),
+    version,
+  );
+
+  // Later betas/RCs get a "Changes since <previous tag>" section on top of the
+  // cumulative Highlights, generated from the fragment diff between the
+  // previous prerelease tag and the working tree.
+  let deltaSection: string[] = [];
+  if (previousTag) {
+    const resolveFragmentDelta = options?.deps?.resolveFragmentDelta ?? defaultResolveFragmentDelta;
+    const delta = resolveFragmentDelta(cwd, previousTag);
+    const deltaBody =
+      delta.length === 0
+        ? `- No changelog fragment changes since ${previousTag}; this build contains packaging or internal-only updates.`
+        : buildDeltaSectionWithClaude(delta, { version, previousTag, deps: options?.deps });
+    deltaSection = [`${DELTA_SECTION_HEADING_PREFIX}${previousTag}`, '', deltaBody];
+  }
+
   const prereleaseNotesPath = path.join(cwd, PRERELEASE_NOTES_PATH);
   const existingReleaseNotes = existsSync(prereleaseNotesPath)
     ? resolveReusablePrereleaseNotes(readFileSync(prereleaseNotesPath, 'utf8'), version)
@@ -1095,10 +1362,41 @@ export function writePrereleaseNotesForVersion(options?: ChangelogOptions): stri
       '> This is a prerelease build for testing. Stable changelog and docs-site updates remain pending until the final stable release.',
     outputPath: PRERELEASE_NOTES_PATH,
     contributions,
-    metadata: [renderPrereleaseBaseVersionMarker(version)],
+    metadata: [renderPrereleaseVersionMarker(version, previousTag)],
+    deltaSection,
   });
 }
 
+// CI gate: the committed prerelease notes must carry a marker generated for
+// exactly the version being tagged, so stale beta.N-1 notes can't ship.
+export function verifyPrereleaseNotesMatchVersion(options?: ChangelogOptions): void {
+  verifyRequestedVersionMatchesPackageVersion(options ?? {});
+
+  const cwd = options?.cwd ?? process.cwd();
+  const existsSync = options?.deps?.existsSync ?? fs.existsSync;
+  const readFileSync = options?.deps?.readFileSync ?? fs.readFileSync;
+  const version = resolveVersion(options ?? {});
+  if (!isSupportedPrereleaseVersion(version)) {
+    throw new Error(
+      `Unsupported prerelease version (${version}). Expected x.y.z-beta.N or x.y.z-rc.N.`,
+    );
+  }
+
+  const prereleaseNotesPath = path.join(cwd, PRERELEASE_NOTES_PATH);
+  if (!existsSync(prereleaseNotesPath)) {
+    throw new Error(
+      `Missing ${prereleaseNotesPath}. Run 'bun run changelog:prerelease-notes --version ${version}' and commit the file before tagging.`,
+    );
+  }
+
+  const markerVersion = extractPrereleaseVersionMarker(readFileSync(prereleaseNotesPath, 'utf8'));
+  if (markerVersion !== version) {
+    throw new Error(
+      `release/prerelease-notes.md was generated for ${markerVersion ?? 'an unknown version (missing or legacy prerelease-version marker)'} but this release is ${version}. Rerun 'bun run changelog:prerelease-notes --version ${version}' and commit the result.`,
+    );
+  }
+}
+
 function parseCliArgs(argv: string[]): {
   baseRef?: string;
   cwd?: string;
@@ -1206,6 +1504,11 @@ function main(): void {
     return;
   }
 
+  if (command === 'check-prerelease-notes') {
+    verifyPrereleaseNotesMatchVersion(options);
+    return;
+  }
+
   if (command === 'docs') {
     generateDocsChangelog(options);
     return;

scripts/build-macos-helper.sh

@@ -1,54 +0,0 @@
-#!/bin/bash
-# Build macOS window tracking helper binary
-
-set -e
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-SWIFT_SOURCE="$SCRIPT_DIR/get-mpv-window-macos.swift"
-OUTPUT_DIR="$SCRIPT_DIR/../dist/scripts"
-OUTPUT_BINARY="$OUTPUT_DIR/get-mpv-window-macos"
-OUTPUT_SOURCE_COPY="$OUTPUT_DIR/get-mpv-window-macos.swift"
-
-fallback_to_source() {
-  echo "Falling back to source fallback: $OUTPUT_SOURCE_COPY"
-  mkdir -p "$OUTPUT_DIR"
-  cp "$SWIFT_SOURCE" "$OUTPUT_SOURCE_COPY"
-}
-
-build_swift_helper() {
-  echo "Compiling macOS window tracking helper..."
-  if ! command -v swiftc >/dev/null 2>&1; then
-    echo "swiftc not found in PATH; skipping compilation."
-    return 1
-  fi
-
-  if ! swiftc -O "$SWIFT_SOURCE" -o "$OUTPUT_BINARY"; then
-    return 1
-  fi
-
-  chmod +x "$OUTPUT_BINARY"
-  echo "✓ Built $OUTPUT_BINARY"
-  return 0
-}
-
-# Optional skip flag for non-macOS CI/dev environments
-if [[ "${SUBMINER_SKIP_MACOS_HELPER_BUILD:-}" == "1" ]]; then
-  echo "Skipping macOS helper build (SUBMINER_SKIP_MACOS_HELPER_BUILD=1)"
-  fallback_to_source
-  exit 0
-fi
-
-# Only build on macOS
-if [[ "$(uname)" != "Darwin" ]]; then
-  echo "Skipping macOS helper build (not on macOS)"
-  fallback_to_source
-  exit 0
-fi
-
-# Create output directory
-mkdir -p "$OUTPUT_DIR"
-
-# Compile Swift script to binary, fallback to source if unavailable or compilation fails
-if ! build_swift_helper; then
-  fallback_to_source
-fi

scripts/dev.sh

@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+FILE="${1:-}"
+
+if [[ ! -f "$FILE" ]]; then
+	printf 'Not a file: %s\n' "${FILE:-<missing>}" >&2
+	exit 1
+fi
+
+if ! mpv --no-config --no-terminal --msg-level=all=no --vo=null --ao=null --frames=1 -- "$FILE"; then
+	printf 'Not playable by mpv: %s\n' "$FILE" >&2
+	exit 1
+fi
+
+exec subminer app --dev --launch-mpv "$FILE"

scripts/mkv-to-readme-video.sh

@@ -19,7 +19,8 @@ Options:
   -w, --webp   Generate animated WebP preview
 
 Encoding profile:
-  - Crop: 1920x1080 at x=760 y=200
+  - Crop: mpv region at 1920x1080, x=760 y=205 on a 3440x1440 canvas
+  - Output size: 1920x1080
   - MP4:  H.264 + AAC
   - WebM: AV1/VP9 + Opus at 30 fps
 USAGE
@@ -148,7 +149,8 @@ pick_webp_encoder() {
 	return 1
 }
 
-crop_vf="crop=1920:1080:760:205"
+# OBS may resize the 3440x1440 canvas, so scale the mpv bounds with the input.
+crop_vf="crop=1920*iw/3440:1080*ih/1440:760*iw/3440:205*ih/1440,scale=1920:1080:flags=lanczos"
 webm_vf="${crop_vf},fps=30"
 
 echo "Generating MP4: $mp4_out"

scripts/mkv-to-readme-video.test.ts

@@ -40,7 +40,7 @@ function toBashPath(filePath: string): string {
   return `${drive.toUpperCase()}:/${rest}`;
 }
 
-test('mkv-to-readme-video accepts libwebp_anim when libwebp is unavailable', () => {
+test('mkv-to-readme-video builds every output with the scaled mpv crop', () => {
   withTempDir((root) => {
     const binDir = path.join(root, 'bin');
     const inputPath = path.join(root, 'sample.mkv');
@@ -104,5 +104,9 @@ touch "$output"
 
     const ffmpegLog = fs.readFileSync(ffmpegLogPath, 'utf8');
     assert.match(ffmpegLog, /-c:v libwebp_anim/);
+    const scaledCropUses = ffmpegLog.match(
+      /-vf crop=1920\*iw\/3440:1080\*ih\/1440:760\*iw\/3440:205\*ih\/1440,scale=1920:1080:flags=lanczos/g,
+    );
+    assert.equal(scaledCropUses?.length, 4);
   });
 });

scripts/prepare-build-assets.mjs

@@ -1,4 +1,5 @@
 import fs from 'node:fs';
+import os from 'node:os';
 import path from 'node:path';
 import { execFileSync } from 'node:child_process';
 import { fileURLToPath } from 'node:url';
@@ -52,6 +53,16 @@ function fallbackToMacosSource() {
   process.stdout.write(`Staged macOS helper source fallback: ${macosHelperSourceCopyPath}\n`);
 }
 
+// Pin the minimum macOS to the app's own floor (Electron's `minos`). Without an
+// explicit target, swiftc stamps the build machine's OS version as the binary's
+// minimum and the helper fails to load on older systems (#213). The arch stays
+// the host's, matching the single-arch app electron-builder packages here.
+const MACOS_HELPER_DEPLOYMENT_TARGET = '12.0';
+
+function macosHelperTarget() {
+  return `${os.arch() === 'x64' ? 'x86_64' : 'arm64'}-apple-macos${MACOS_HELPER_DEPLOYMENT_TARGET}`;
+}
+
 function shouldSkipMacosHelperBuild() {
   return process.env.SUBMINER_SKIP_MACOS_HELPER_BUILD === '1';
 }
@@ -72,9 +83,13 @@ function buildMacosHelper() {
   ensureDir(scriptsOutputDir);
 
   try {
-    execFileSync('swiftc', ['-O', macosHelperSourcePath, '-o', macosHelperBinaryPath], {
+    execFileSync(
+      'swiftc',
+      ['-O', '-target', macosHelperTarget(), macosHelperSourcePath, '-o', macosHelperBinaryPath],
+      {
         stdio: 'inherit',
-    });
+      },
+    );
     fs.chmodSync(macosHelperBinaryPath, 0o755);
     process.stdout.write(`Built macOS helper: ${macosHelperBinaryPath}\n`);
   } catch (error) {

scripts/prepare-build-assets.test.ts

@@ -8,7 +8,7 @@ test('macOS helper build creates dist scripts directory before swiftc output', (
   const buildFunctionIndex = source.indexOf('function buildMacosHelper()');
   assert.notEqual(buildFunctionIndex, -1);
 
-  const swiftcIndex = source.indexOf("execFileSync('swiftc'", buildFunctionIndex);
+  const swiftcIndex = source.indexOf("'swiftc'", buildFunctionIndex);
   assert.notEqual(swiftcIndex, -1);
 
   const ensureDirIndex = source.lastIndexOf('ensureDir(scriptsOutputDir)', swiftcIndex);
@@ -18,3 +18,10 @@ test('macOS helper build creates dist scripts directory before swiftc output', (
     'buildMacosHelper must create dist/scripts before swiftc writes the helper binary',
   );
 });
+
+// Regression guard for #213: an untargeted swiftc stamps the build machine's OS
+// version as the helper's minimum, so released builds refuse to load on older macOS.
+test('macOS helper is compiled with an explicit deployment target', () => {
+  assert.match(source, /-target/);
+  assert.match(source, /apple-macos\$\{MACOS_HELPER_DEPLOYMENT_TARGET\}/);
+});

scripts/test-plugin-start-gate.lua

@@ -130,7 +130,9 @@ local function run_plugin_scenario(config)
 
 		function mp.add_timeout(seconds, callback)
 			recorded.timeouts[#recorded.timeouts + 1] = seconds
+			local delay = tonumber(seconds) or 0
 			local timeout = {
+				seconds = delay,
 				killed = false,
 				callback = callback,
 			}
@@ -138,7 +140,6 @@ local function run_plugin_scenario(config)
 				self.killed = true
 			end
 
-			local delay = tonumber(seconds) or 0
 			if callback and delay < 5 and not config.defer_timeouts then
 				callback()
 			end
@@ -514,6 +515,15 @@ local function has_timeout(timeouts, target)
 	return false
 end
 
+local function find_timeout_handle(recorded, target)
+	for _, timeout in ipairs(recorded.timeout_handles) do
+		if math.abs(timeout.seconds - target) < 0.0001 then
+			return timeout
+		end
+	end
+	return nil
+end
+
 local function env_has(call, target)
 	local env = (call and call.env) or {}
 	for _, value in ipairs(env) do
@@ -1636,6 +1646,8 @@ do
 		#recorded.periodic_timers == 1,
 		"auto-start visible overlay should refresh the early overlay loading OSD"
 	)
+	local overlay_loading_deadline = find_timeout_handle(recorded, 30)
+	assert_true(overlay_loading_deadline ~= nil, "overlay loading OSD should have a bounded deadline")
 	local overlay_loading_timer = recorded.periodic_timers[1]
 	recorded.periodic_timers[1].callback()
 	assert_true(
@@ -1670,6 +1682,46 @@ do
 		recorded.periodic_timers[1].killed == true,
 		"overlay loading ready should stop the early overlay loading OSD refresher"
 	)
+	assert_true(
+		overlay_loading_deadline.killed == true,
+		"overlay loading ready should cancel the bounded loading deadline"
+	)
+end
+
+do
+	local recorded, err = run_plugin_scenario({
+		defer_timeouts = true,
+		process_list = "",
+		option_overrides = {
+			binary_path = binary_path,
+			auto_start = "yes",
+			auto_start_visible_overlay = "yes",
+			osd_messages = false,
+			socket_path = "/tmp/subminer-socket",
+		},
+		input_ipc_server = "/tmp/subminer-socket",
+		media_title = "Random Movie",
+		files = {
+			[binary_path] = true,
+		},
+	})
+	assert_true(recorded ~= nil, "plugin failed to load for overlay loading deadline scenario: " .. tostring(err))
+	fire_event(recorded, "start-file")
+	local overlay_loading_deadline = find_timeout_handle(recorded, 30)
+	assert_true(overlay_loading_deadline ~= nil, "overlay loading deadline should be scheduled")
+	overlay_loading_deadline.callback()
+	assert_true(
+		recorded.periodic_timers[1].killed == true,
+		"overlay loading deadline should stop the loading spinner"
+	)
+	assert_true(
+		has_osd_message(recorded.osd, "SubMiner: Overlay did not become ready; check SubMiner logs"),
+		"overlay loading deadline should replace the spinner with actionable feedback"
+	)
+	assert_true(
+		has_log_containing(recorded.logs, "Overlay loading deadline expired"),
+		"overlay loading deadline should leave a diagnostic log entry"
+	)
 end
 
 do

scripts/update-aur-package.test.ts

@@ -82,6 +82,7 @@ test('update-aur-package updates PKGBUILD and .SRCINFO without makepkg', () => {
       pkgbuild,
       /^\s*install -Dm755 "\$\{srcdir\}\/subminer-\$\{pkgver\}" "\$\{pkgdir\}\/usr\/bin\/subminer"$/m,
     );
+    assert.match(pkgbuild, /assets\/thumbnailers\/subminer-ffmpegthumbnailer\.thumbnailer/);
     assert.match(srcinfo, /^\tpkgver = 0\.6\.3$/m);
     assert.match(srcinfo, /^\tprovides = subminer=0\.6\.3$/m);
     assert.match(

src/anki-integration.test.ts

@@ -11,10 +11,12 @@ import type { MediaInput } from './media-input';
 import { AnkiConnectConfig } from './types';
 
 type TestOverlayNotificationPayload = {
+  id?: string;
   title: string;
   body?: string;
   image?: string;
   variant?: string;
+  persistent?: boolean;
   actions?: Array<{ id: string; label: string; noteId?: number }>;
 };
 
@@ -606,6 +608,7 @@ test('AnkiIntegration applies ready YouTube cache media to every queued note id'
   const integration = new AnkiIntegration(
     {
       fields: {
+        audio: 'ExpressionAudio',
         image: 'Picture',
       },
       media: {
@@ -659,7 +662,7 @@ test('AnkiIntegration applies ready YouTube cache media to every queued note id'
       noteIds.map((noteId) => ({
         noteId,
         fields: {
-          SentenceAudio: { value: '' },
+          ExpressionAudio: { value: '' },
           Picture: { value: '' },
         },
       })),
@@ -944,7 +947,7 @@ test('AnkiIntegration queues YouTube media updates against recovered source URLs
     noteInfo: {
       noteId: 404,
       fields: {
-        SentenceAudio: { value: '' },
+        ExpressionAudio: { value: '' },
         Picture: { value: '' },
       },
     },
@@ -956,7 +959,8 @@ test('AnkiIntegration queues YouTube media updates against recovered source URLs
   assert.equal(queued, true);
   assert.equal(updatedNotes.length, 1);
   assert.equal(updatedNotes[0]?.noteId, 404);
-  assert.match(updatedNotes[0]?.fields.SentenceAudio ?? '', /^\[sound:audio_/);
+  assert.match(updatedNotes[0]?.fields.ExpressionAudio ?? '', /^\[sound:audio_/);
+  assert.equal(updatedNotes[0]?.fields.SentenceAudio, undefined);
   assert.match(updatedNotes[0]?.fields.Picture ?? '', /^<img src="image_/);
   assert.equal(storedMedia.length, 2);
   assert.deepEqual(audioVolumeScales, [0.3 ** 3]);
@@ -1182,6 +1186,82 @@ test('AnkiIntegration embeds generated notification image on overlay mined-card
   assert.deepEqual(cleanupPaths, [notificationIconPath]);
 });
 
+test('AnkiIntegration keeps overlay card-update progress visible until the terminal notification', async () => {
+  const overlayNotifications: TestOverlayNotificationPayload[] = [];
+  const integration = new AnkiIntegration(
+    {
+      behavior: {
+        notificationType: 'overlay',
+      },
+    },
+    {} as never,
+    {} as never,
+    undefined,
+    undefined,
+    undefined,
+    undefined,
+    {},
+    undefined,
+    (payload) => {
+      overlayNotifications.push(payload);
+    },
+  );
+  const updateNotifications = integration as unknown as {
+    beginUpdateProgress: (message: string) => void;
+    showNotification: (noteId: number, label: string | number) => Promise<void>;
+  };
+
+  updateNotifications.beginUpdateProgress('Updating card');
+  await updateNotifications.showNotification(42, '食べる');
+
+  assert.deepEqual(
+    overlayNotifications.map(({ id, variant, persistent }) => ({ id, variant, persistent })),
+    [
+      { id: 'anki-update-progress', variant: 'progress', persistent: true },
+      { id: 'anki-update-progress', variant: 'success', persistent: false },
+    ],
+  );
+});
+
+test('AnkiIntegration dismisses persistent overlay update progress when no terminal notification replaces it', () => {
+  const overlayNotifications: TestOverlayNotificationPayload[] = [];
+  const dismissedIds: string[] = [];
+  const integration = new AnkiIntegration(
+    {
+      behavior: {
+        notificationType: 'overlay',
+      },
+    },
+    {} as never,
+    {} as never,
+    undefined,
+    undefined,
+    undefined,
+    undefined,
+    {},
+    undefined,
+    (payload) => {
+      overlayNotifications.push(payload);
+    },
+    undefined,
+    undefined,
+    undefined,
+    (id) => {
+      dismissedIds.push(id);
+    },
+  );
+  const updateNotifications = integration as unknown as {
+    beginUpdateProgress: (message: string) => void;
+    endUpdateProgress: () => void;
+  };
+
+  updateNotifications.beginUpdateProgress('Updating card');
+  updateNotifications.endUpdateProgress();
+
+  assert.equal(overlayNotifications[0]?.persistent, true);
+  assert.deepEqual(dismissedIds, ['anki-update-progress']);
+});
+
 test('AnkiIntegration keeps overlay notification image when temp icon write fails', async () => {
   const desktopNotifications: Array<{ title: string; body?: string; icon?: string }> = [];
   const overlayNotifications: TestOverlayNotificationPayload[] = [];

src/anki-integration.ts

@@ -218,6 +218,8 @@ export class AnkiIntegration {
     null;
   private overlayNotificationCallback: ((payload: OverlayNotificationPayload) => void) | null =
     null;
+  private overlayNotificationDismissCallback: ((id: string) => void) | null = null;
+  private overlayUpdateProgressActive = false;
   private updateInProgress = false;
   private uiFeedbackState: UiFeedbackState = createUiFeedbackState();
   private parseWarningKeys = new Set<string>();
@@ -265,6 +267,7 @@ export class AnkiIntegration {
     getCachedMediaPath?: MediaGenerationInputResolverOptions['getCachedMediaPath'],
     shouldRequireRemoteMediaCache?: () => boolean,
     getYoutubeMediaSourceUrl?: () => Promise<string | null | undefined> | string | null | undefined,
+    overlayNotificationDismissCallback?: (id: string) => void,
   ) {
     this.config = normalizeAnkiIntegrationConfig(config);
     this.aiConfig = { ...aiConfig };
@@ -280,6 +283,7 @@ export class AnkiIntegration {
     this.getCachedMediaPath = getCachedMediaPath ?? null;
     this.shouldRequireRemoteMediaCache = shouldRequireRemoteMediaCache ?? null;
     this.getYoutubeMediaSourceUrl = getYoutubeMediaSourceUrl ?? null;
+    this.overlayNotificationDismissCallback = overlayNotificationDismissCallback ?? null;
     this.pendingYoutubeMediaQueue = this.createPendingYoutubeMediaQueue();
     this.knownWordCache = this.createKnownWordCache(knownWordCacheStatePath);
     this.pollingRunner = this.createPollingRunner();
@@ -379,8 +383,6 @@ export class AnkiIntegration {
       getCachedMediaPath: this.getCachedMediaPath,
       shouldRequireRemoteMediaCache: () => this.shouldRequireRemoteMediaCache?.() === true,
       getSubtitleMediaRange: (context) => this.getSubtitleMediaRange(context),
-      getResolvedSentenceAudioFieldName: (noteInfo) =>
-        this.getResolvedSentenceAudioFieldName(noteInfo),
       resolveConfiguredFieldName: (noteInfo, ...preferredNames) =>
         this.resolveConfiguredFieldName(noteInfo, ...preferredNames),
       mergeFieldValue: (existing, newValue, overwrite) =>
@@ -657,8 +659,6 @@ export class AnkiIntegration {
         this.setCardTypeFields(updatedFields, availableFieldNames, cardKind),
       resolveConfiguredFieldName: (noteInfo, ...preferredNames) =>
         this.resolveConfiguredFieldName(noteInfo, ...preferredNames),
-      getResolvedSentenceAudioFieldName: (noteInfo) =>
-        this.getResolvedSentenceAudioFieldName(noteInfo),
       getAnimatedImageLeadInSeconds: (noteInfo) => this.getAnimatedImageLeadInSeconds(noteInfo),
       mergeFieldValue: (existing, newValue, overwrite) =>
         this.mergeFieldValue(existing, newValue, overwrite),
@@ -1203,12 +1203,13 @@ export class AnkiIntegration {
   private beginUpdateProgress(initialMessage: string): void {
     if (!this.shouldUseOsdNotifications()) {
       if (this.shouldUseOverlayNotifications()) {
+        this.overlayUpdateProgressActive = true;
         this.overlayNotificationCallback?.({
           id: 'anki-update-progress',
           title: 'Anki update',
           body: initialMessage,
           variant: 'progress',
-          persistent: false,
+          persistent: true,
         });
       }
       return;
@@ -1220,6 +1221,10 @@ export class AnkiIntegration {
 
   private endUpdateProgress(): void {
     if (!this.shouldUseOsdNotifications()) {
+      if (this.overlayUpdateProgressActive) {
+        this.overlayUpdateProgressActive = false;
+        this.overlayNotificationDismissCallback?.('anki-update-progress');
+      }
       return;
     }
     endUpdateProgress(this.uiFeedbackState, (timer) => {
@@ -1243,18 +1248,20 @@ export class AnkiIntegration {
     if (!this.shouldUseOsdNotifications()) {
       this.updateInProgress = true;
       if (this.shouldUseOverlayNotifications()) {
+        this.overlayUpdateProgressActive = true;
         this.overlayNotificationCallback?.({
           id: 'anki-update-progress',
           title: 'Anki update',
           body: initialMessage,
           variant: 'progress',
-          persistent: false,
+          persistent: true,
         });
       }
       try {
         return await action();
       } finally {
         this.updateInProgress = false;
+        this.endUpdateProgress();
       }
     }
     return withUpdateProgress(
@@ -1353,6 +1360,7 @@ export class AnkiIntegration {
         : undefined;
 
     if (shouldShowOverlayNotification && this.overlayNotificationCallback) {
+      this.overlayUpdateProgressActive = false;
       this.overlayNotificationCallback({
         id: 'anki-update-progress',
         title: 'Anki Card Updated',
@@ -1555,13 +1563,6 @@ export class AnkiIntegration {
     return sentenceCardConfig.audioField || 'SentenceAudio';
   }
 
-  private getResolvedSentenceAudioFieldName(noteInfo: NoteInfo): string | null {
-    return (
-      this.resolveNoteFieldName(noteInfo, this.getPreferredSentenceAudioFieldName()) ||
-      this.resolveConfiguredFieldName(noteInfo, this.config.fields?.audio)
-    );
-  }
-
   private getConfiguredWordFieldName(): string {
     return getConfiguredWordFieldName(this.config);
   }

src/anki-integration/card-creation-manual-update.test.ts

@@ -143,7 +143,7 @@ function createManualUpdateService(overrides: Partial<CardCreationDeps> = {}): {
   };
 }
 
-test('manual clipboard subtitle update replaces sentence audio without touching expression audio', async () => {
+test('manual clipboard subtitle update replaces audio in the configured field', async () => {
   const { service, updatedFields, mergeCalls, storedMedia } = createManualUpdateService();
 
   await service.updateLastAddedFromClipboard('字幕');
@@ -151,14 +151,144 @@ test('manual clipboard subtitle update replaces sentence audio without touching
   assert.equal(updatedFields.length, 1);
   assert.equal(storedMedia.length, 1);
   const audioValue = `[sound:${storedMedia[0]}]`;
-  assert.equal(updatedFields[0]?.SentenceAudio, audioValue);
-  assert.equal('ExpressionAudio' in updatedFields[0]!, false);
+  assert.equal(updatedFields[0]?.ExpressionAudio, audioValue);
+  assert.equal('SentenceAudio' in updatedFields[0]!, false);
   assert.deepEqual(
     mergeCalls.map((call) => call.overwrite),
     [true],
   );
 });
 
+test('manual clipboard word-card update uses configured fields with Lapis and Kiku enabled', async () => {
+  const { service, updatedFields } = createManualUpdateService({
+    getConfig: () =>
+      ({
+        deck: 'Mining',
+        fields: {
+          word: 'Expression',
+          sentence: 'Context',
+          audio: 'ContextAudio',
+        },
+        media: {
+          generateAudio: true,
+          generateImage: false,
+          maxMediaDuration: 30,
+        },
+        behavior: {
+          overwriteAudio: false,
+          overwriteImage: false,
+        },
+        ai: false,
+      }) as AnkiConnectConfig,
+    client: {
+      addNote: async () => 0,
+      addTags: async () => undefined,
+      notesInfo: async () => [
+        {
+          noteId: 42,
+          fields: {
+            Expression: { value: '単語' },
+            Sentence: { value: '' },
+            SentenceAudio: { value: '' },
+            Context: { value: '' },
+            ContextAudio: { value: '' },
+          },
+        },
+      ],
+      updateNoteFields: async (_noteId, fields) => {
+        updatedFields.push(fields);
+      },
+      storeMediaFile: async () => undefined,
+      findNotes: async () => [42],
+      retrieveMediaFile: async () => '',
+    },
+    getEffectiveSentenceCardConfig: () => ({
+      model: 'Sentence',
+      sentenceField: 'Sentence',
+      audioField: 'SentenceAudio',
+      lapisEnabled: true,
+      kikuEnabled: true,
+      kikuFieldGrouping: 'disabled',
+      kikuDeleteDuplicateInAuto: false,
+    }),
+  });
+
+  await service.updateLastAddedFromClipboard('字幕');
+
+  assert.equal(updatedFields.length, 1);
+  assert.match(updatedFields[0]?.ContextAudio ?? '', /^\[sound:audio_\d+\.mp3\]$/);
+  assert.deepEqual(Object.keys(updatedFields[0] ?? {}).sort(), ['Context', 'ContextAudio']);
+  assert.equal(updatedFields[0]?.Context, '字幕');
+});
+
+test('audio-card action keeps Lapis and Kiku sentence fields', async () => {
+  const { service, updatedFields } = createManualUpdateService({
+    getConfig: () =>
+      ({
+        deck: 'Mining',
+        fields: {
+          word: 'Expression',
+          sentence: 'Context',
+          audio: 'ContextAudio',
+        },
+        media: {
+          generateAudio: true,
+          generateImage: false,
+          maxMediaDuration: 30,
+        },
+        behavior: {},
+        ai: false,
+      }) as AnkiConnectConfig,
+    getMpvClient: () =>
+      ({
+        currentVideoPath: '/video.mp4',
+        currentAudioStreamIndex: 0,
+        currentSubText: '字幕',
+        currentSubStart: 12,
+        currentSubEnd: 14,
+      }) as never,
+    client: {
+      addNote: async () => 0,
+      addTags: async () => undefined,
+      notesInfo: async () => [
+        {
+          noteId: 42,
+          fields: {
+            Expression: { value: '単語' },
+            Sentence: { value: '' },
+            SentenceAudio: { value: '' },
+            Context: { value: '' },
+            ContextAudio: { value: '' },
+          },
+        },
+      ],
+      updateNoteFields: async (_noteId, fields) => {
+        updatedFields.push(fields);
+      },
+      storeMediaFile: async () => undefined,
+      findNotes: async () => [42],
+      retrieveMediaFile: async () => '',
+    },
+    getEffectiveSentenceCardConfig: () => ({
+      model: 'Sentence',
+      sentenceField: 'Sentence',
+      audioField: 'SentenceAudio',
+      lapisEnabled: true,
+      kikuEnabled: true,
+      kikuFieldGrouping: 'disabled',
+      kikuDeleteDuplicateInAuto: false,
+    }),
+  });
+
+  await service.markLastCardAsAudioCard();
+
+  assert.equal(updatedFields.length, 1);
+  assert.equal(updatedFields[0]?.Sentence, '字幕');
+  assert.match(updatedFields[0]?.SentenceAudio ?? '', /^\[sound:audio_\d+\.mp3\]$/);
+  assert.equal('Context' in (updatedFields[0] ?? {}), false);
+  assert.equal('ContextAudio' in (updatedFields[0] ?? {}), false);
+});
+
 test('manual clipboard subtitle update marks Kiku word cards as word-and-sentence cards when enabled', async () => {
   const { service, updatedFields } = createManualUpdateService({
     getConfig: () =>
@@ -225,7 +355,7 @@ test('manual clipboard subtitle update marks Kiku word cards as word-and-sentenc
   });
 });
 
-test('manual clipboard subtitle update skips audio when sentence audio field is missing', async () => {
+test('manual clipboard subtitle update uses configured audio when SentenceAudio is missing', async () => {
   const { service, updatedFields, mergeCalls, storedMedia } = createManualUpdateService({
     client: {
       addNote: async () => 0,
@@ -255,8 +385,9 @@ test('manual clipboard subtitle update skips audio when sentence audio field is
 
   assert.equal(storedMedia.length, 1);
   assert.equal(updatedFields.length, 1);
-  assert.deepEqual(updatedFields[0], { Sentence: '字幕' });
-  assert.equal(mergeCalls.length, 0);
+  assert.match(updatedFields[0]?.ExpressionAudio ?? '', /^\[sound:audio_\d+\.mp3\]$/);
+  assert.equal(updatedFields[0]?.Sentence, '字幕');
+  assert.equal(mergeCalls.length, 1);
 });
 
 test('manual clipboard subtitle update uses resolved mpv stream URLs for remote media', async () => {

src/anki-integration/card-creation.ts

@@ -260,9 +260,16 @@ export class CardCreationService {
           fields,
           this.deps.getConfig(),
         );
-        const sentenceAudioField = this.getResolvedSentenceOnlyAudioFieldName(noteInfo);
+        const config = this.deps.getConfig();
+        const sentenceAudioField = this.deps.resolveConfiguredFieldName(
+          noteInfo,
+          config.fields?.audio ?? DEFAULT_ANKI_CONNECT_CONFIG.fields.audio,
+        );
         const sentenceCardConfig = this.deps.getEffectiveSentenceCardConfig();
-        const sentenceField = sentenceCardConfig.sentenceField;
+        const sentenceField = this.deps.resolveConfiguredFieldName(
+          noteInfo,
+          config.fields?.sentence ?? DEFAULT_ANKI_CONNECT_CONFIG.fields.sentence,
+        );
 
         const sentence = blocks.join(' ');
         const updatedFields: Record<string, string> = {};
@@ -284,7 +291,6 @@ export class CardCreationService {
           `Clipboard update: timing range ${rangeStart.toFixed(2)}s - ${rangeEnd.toFixed(2)}s`,
         );
 
-        const config = this.deps.getConfig();
         const generateAudio = shouldGenerateAudio(config);
         const generateImage = shouldGenerateImage(config);
         const mediaResolverOptions = this.getMediaResolverOptions();
@@ -457,13 +463,13 @@ export class CardCreationService {
 
         this.deps.setCardTypeFields(updatedFields, Object.keys(noteInfo.fields), 'audio');
 
-        const sentenceField = this.deps.getConfig().fields?.sentence;
+        const sentenceCardConfig = this.deps.getEffectiveSentenceCardConfig();
+        const sentenceField = sentenceCardConfig.sentenceField;
         if (sentenceField) {
           const processedSentence = this.deps.processSentence(mpvClient.currentSubText, fields);
           updatedFields[sentenceField] = processedSentence;
         }
 
-        const sentenceCardConfig = this.deps.getEffectiveSentenceCardConfig();
         const audioFieldName = sentenceCardConfig.audioField;
         try {
           const audioFilename = this.generateAudioFilename();
@@ -806,22 +812,6 @@ export class CardCreationService {
     }
   }
 
-  private getResolvedSentenceAudioFieldName(noteInfo: CardCreationNoteInfo): string | null {
-    return (
-      this.deps.resolveNoteFieldName(
-        noteInfo,
-        this.deps.getEffectiveSentenceCardConfig().audioField || 'SentenceAudio',
-      ) || this.deps.resolveConfiguredFieldName(noteInfo, this.deps.getConfig().fields?.audio)
-    );
-  }
-
-  private getResolvedSentenceOnlyAudioFieldName(noteInfo: CardCreationNoteInfo): string | null {
-    return this.deps.resolveNoteFieldName(
-      noteInfo,
-      this.deps.getEffectiveSentenceCardConfig().audioField || 'SentenceAudio',
-    );
-  }
-
   private createPendingNoteInfo(fields: Record<string, string>): CardCreationNoteInfo {
     return {
       noteId: -1,

src/anki-integration/note-update-workflow.test.ts

@@ -80,7 +80,6 @@ function createWorkflowHarness() {
       const names = Object.keys(noteInfo.fields);
       return names.find((name) => name.toLowerCase() === preferred.toLowerCase()) ?? null;
     },
-    getResolvedSentenceAudioFieldName: () => null,
     getAnimatedImageLeadInSeconds: async () => 0,
     mergeFieldValue: (_existing: string, next: string, _overwrite: boolean) => next,
     generateAudioFilename: () => 'audio_1.mp3',
@@ -120,6 +119,49 @@ test('NoteUpdateWorkflow updates sentence field and emits notification', async (
   assert.equal(harness.notifications.length, 1);
 });
 
+test('NoteUpdateWorkflow uses configured fields for word-card enrichment with Lapis and Kiku enabled', async () => {
+  const harness = createWorkflowHarness();
+  harness.deps.getConfig = () => ({
+    fields: {
+      sentence: 'Context',
+      audio: 'ContextAudio',
+    },
+    media: {
+      generateAudio: true,
+      generateImage: false,
+    },
+    behavior: {},
+  });
+  harness.deps.getEffectiveSentenceCardConfig = () => ({
+    sentenceField: 'Sentence',
+    lapisEnabled: true,
+    kikuEnabled: true,
+    kikuFieldGrouping: 'disabled',
+  });
+  harness.deps.client.notesInfo = async () =>
+    [
+      {
+        noteId: 42,
+        fields: {
+          Expression: { value: 'taberu' },
+          Sentence: { value: '' },
+          SentenceAudio: { value: '' },
+          Context: { value: '' },
+          ContextAudio: { value: '' },
+        },
+      },
+    ] satisfies NoteUpdateWorkflowNoteInfo[];
+  harness.deps.generateAudio = async () => Buffer.from('audio');
+
+  await harness.workflow.execute(42);
+
+  assert.equal(harness.updates.length, 1);
+  assert.deepEqual(harness.updates[0]?.fields, {
+    Context: 'subtitle-text',
+    ContextAudio: '[sound:audio_1.mp3]',
+  });
+});
+
 test('NoteUpdateWorkflow updates sentence furigana when highlight processor changes it', async () => {
   const harness = createWorkflowHarness();
   harness.deps.client.notesInfo = async () =>
@@ -432,6 +474,7 @@ test('NoteUpdateWorkflow uses subtitle sidebar context for sentence media timing
   harness.deps.getConfig = () => ({
     fields: {
       sentence: 'Sentence',
+      audio: 'SentenceAudio',
       image: 'Picture',
       miscInfo: 'MiscInfo',
     },
@@ -444,7 +487,6 @@ test('NoteUpdateWorkflow uses subtitle sidebar context for sentence media timing
   });
   harness.deps.getCurrentSubtitleText = () => 'current primary line';
   harness.deps.getCurrentSubtitleStart = () => 20;
-  harness.deps.getResolvedSentenceAudioFieldName = () => 'SentenceAudio';
   harness.deps.generateAudio = async (context?: SubtitleMiningContext) => {
     audioContext = context ?? null;
     return Buffer.from('audio');
@@ -501,6 +543,7 @@ test('NoteUpdateWorkflow snapshots one media range for audio and image without a
   harness.deps.getConfig = () => ({
     fields: {
       sentence: 'Sentence',
+      audio: 'SentenceAudio',
       image: 'Picture',
       miscInfo: 'MiscInfo',
     },
@@ -511,7 +554,6 @@ test('NoteUpdateWorkflow snapshots one media range for audio and image without a
     },
     behavior: {},
   });
-  harness.deps.getResolvedSentenceAudioFieldName = () => 'SentenceAudio';
   harness.deps.captureSubtitleMediaContext = () => {
     captureCalls += 1;
     return capturedContext;

src/anki-integration/note-update-workflow.ts

@@ -19,6 +19,7 @@ export interface NoteUpdateWorkflowDeps {
     fields?: {
       word?: string;
       sentence?: string;
+      audio?: string;
       image?: string;
       miscInfo?: string;
     };
@@ -75,7 +76,6 @@ export interface NoteUpdateWorkflowDeps {
     noteInfo: NoteUpdateWorkflowNoteInfo,
     ...preferredNames: (string | undefined)[]
   ) => string | null;
-  getResolvedSentenceAudioFieldName: (noteInfo: NoteUpdateWorkflowNoteInfo) => string | null;
   getAnimatedImageLeadInSeconds: (noteInfo: NoteUpdateWorkflowNoteInfo) => Promise<number>;
   mergeFieldValue: (existing: string, newValue: string, overwrite: boolean) => string;
   generateAudioFilename: () => string;
@@ -198,11 +198,13 @@ export class NoteUpdateWorkflow {
       const updatedFields: Record<string, string> = {};
       let updatePerformed = false;
       let miscInfoFilename: string | null = null;
-      const sentenceField = sentenceCardConfig.sentenceField;
+      const configuredSentenceField =
+        config.fields?.sentence ?? DEFAULT_ANKI_CONNECT_CONFIG.fields.sentence;
+      const sentenceField = this.deps.resolveConfiguredFieldName(noteInfo, configuredSentenceField);
       const subtitleMiningContext = this.consumeMatchingSubtitleMiningContext(
         fields,
-        sentenceField,
-        config.fields?.sentence,
+        sentenceField ?? configuredSentenceField,
+        configuredSentenceField,
       );
       // Audio and image generation run sequentially and audio extraction can take tens of
       // seconds, so resolve the clip range exactly once up front; reading live mpv sub
@@ -258,7 +260,10 @@ export class NoteUpdateWorkflow {
 
           if (audioBuffer) {
             await this.deps.client.storeMediaFile(audioFilename, audioBuffer);
-            const sentenceAudioField = this.deps.getResolvedSentenceAudioFieldName(noteInfo);
+            const sentenceAudioField = this.deps.resolveConfiguredFieldName(
+              noteInfo,
+              config.fields?.audio ?? DEFAULT_ANKI_CONNECT_CONFIG.fields.audio,
+            );
             if (sentenceAudioField) {
               const existingAudio = noteInfo.fields[sentenceAudioField]?.value || '';
               updatedFields[sentenceAudioField] = this.deps.mergeFieldValue(

src/anki-integration/pending-youtube-media-queue.test.ts

@@ -31,7 +31,6 @@ function createDeps(
     getCachedMediaPath: async () => null,
     shouldRequireRemoteMediaCache: () => true,
     getSubtitleMediaRange: () => ({ startTime: 1, endTime: 2 }),
-    getResolvedSentenceAudioFieldName: () => 'SentenceAudio',
     resolveConfiguredFieldName: () => 'Picture',
     mergeFieldValue: (_existing, newValue) => newValue,
     getAnimatedImageLeadInSeconds: async () => 0,
@@ -133,7 +132,7 @@ test('PendingYoutubeMediaQueue defaults missing media flags to enabled when queu
         noteIds.map((noteId) => ({
           noteId,
           fields: {
-            SentenceAudio: { value: '' },
+            ExpressionAudio: { value: '' },
             Picture: { value: '' },
           },
         })),
@@ -144,13 +143,16 @@ test('PendingYoutubeMediaQueue defaults missing media flags to enabled when queu
         storedMedia.push(filename);
       },
     },
-    getConfig: () => ({ media: {}, fields: { image: 'Picture' } }) as AnkiConnectConfig,
+    getConfig: () =>
+      ({ media: {}, fields: { audio: 'ExpressionAudio', image: 'Picture' } }) as AnkiConnectConfig,
+    resolveConfiguredFieldName: (noteInfo, ...preferredNames) =>
+      preferredNames.find((name) => name && name in noteInfo.fields) ?? null,
   });
   const queue = new PendingYoutubeMediaQueue(deps);
 
   const queued = await queue.queueFromNote({
     noteId: 42,
-    noteInfo: { noteId: 42, fields: {} },
+    noteInfo: { noteId: 42, fields: { ExpressionAudio: { value: '' } } },
     label: 'demo',
   });
   await queue.handleReady('https://youtu.be/abc123', '/tmp/media.mkv');
@@ -158,7 +160,8 @@ test('PendingYoutubeMediaQueue defaults missing media flags to enabled when queu
   assert.equal(queued, true);
   assert.equal(updatedNotes.length, 1);
   assert.equal(storedMedia.length, 2);
-  assert.match(updatedNotes[0]?.fields.SentenceAudio ?? '', /^\[sound:audio\.mp3\]$/);
+  assert.match(updatedNotes[0]?.fields.ExpressionAudio ?? '', /^\[sound:audio\.mp3\]$/);
+  assert.equal('SentenceAudio' in (updatedNotes[0]?.fields ?? {}), false);
   assert.match(updatedNotes[0]?.fields.Picture ?? '', /^<img src="image\.webp">$/);
 });
 

src/anki-integration/pending-youtube-media-queue.ts

@@ -39,7 +39,6 @@ export interface PendingYoutubeMediaQueueDeps {
     startTime: number;
     endTime: number;
   };
-  getResolvedSentenceAudioFieldName: (noteInfo: PendingYoutubeMediaNoteInfo) => string | null;
   resolveConfiguredFieldName: (
     noteInfo: PendingYoutubeMediaNoteInfo,
     ...preferredNames: (string | undefined)[]
@@ -136,7 +135,7 @@ export class PendingYoutubeMediaQueue {
       startTime: mediaRange.startTime,
       endTime: mediaRange.endTime,
       label: job.label,
-      audioFieldName: this.deps.getResolvedSentenceAudioFieldName(job.noteInfo) ?? undefined,
+      audioFieldName: this.resolveConfiguredAudioFieldName(job.noteInfo) ?? undefined,
       imageFieldName:
         this.deps.resolveConfiguredFieldName(
           job.noteInfo,
@@ -247,6 +246,14 @@ export class PendingYoutubeMediaQueue {
     return matched;
   }
 
+  private resolveConfiguredAudioFieldName(noteInfo: PendingYoutubeMediaNoteInfo): string | null {
+    const config = this.deps.getConfig();
+    return this.deps.resolveConfiguredFieldName(
+      noteInfo,
+      config.fields?.audio ?? DEFAULT_ANKI_CONNECT_CONFIG.fields.audio,
+    );
+  }
+
   private async applyUpdate(
     job: PendingYoutubeMediaUpdate,
     cachedPath: string,
@@ -283,7 +290,7 @@ export class PendingYoutubeMediaQueue {
         if (audioBuffer) {
           await this.deps.client.storeMediaFile(audioFilename, audioBuffer);
           const audioField =
-            job.audioFieldName || this.deps.getResolvedSentenceAudioFieldName(noteInfo) || null;
+            job.audioFieldName || this.resolveConfiguredAudioFieldName(noteInfo) || null;
           if (audioField) {
             const existingAudio = noteInfo.fields[audioField]?.value || '';
             mediaFields[audioField] = this.deps.mergeFieldValue(

src/core/services/__tests__/stats-server.test.ts

@@ -284,6 +284,22 @@ function createMockTracker(
     getSessionTimeline: async () => [],
     getSessionEvents: async () => [],
     getVocabularyStats: async () => VOCABULARY_STATS,
+    getVocabularySummary: async () => ({
+      uniqueWords: 501,
+      uniqueWordsWithoutNames: 500,
+      uniqueKanji: 201,
+      newThisWeek: 7,
+      newThisWeekWithoutNames: 6,
+      knownWordCount: 250,
+      knownWordCountWithoutNames: 249,
+    }),
+    getVocabularyChartData: async () => ({
+      ready: true,
+      topWords: [{ wordId: 1, headword: 'する', frequency: 50 }],
+      topWordsWithoutNames: [{ wordId: 1, headword: 'する', frequency: 50 }],
+      newWordsTimeline: [{ epochDay: 20_000, wordCount: 3 }],
+      newWordsTimelineWithoutNames: [{ epochDay: 20_000, wordCount: 3 }],
+    }),
     getStatsExcludedWords: async () => [],
     replaceStatsExcludedWords: async () => {},
     getKanjiStats: async () => KANJI_STATS,
@@ -711,6 +727,38 @@ describe('stats server API routes', () => {
     assert.equal(body[0].headword, 'する');
   });
 
+  it('GET /api/stats/vocabulary/summary returns database-wide card totals', async () => {
+    const app = createStatsApp(createMockTracker());
+
+    const res = await app.request('/api/stats/vocabulary/summary');
+
+    assert.equal(res.status, 200);
+    assert.deepEqual(await res.json(), {
+      uniqueWords: 501,
+      uniqueWordsWithoutNames: 500,
+      uniqueKanji: 201,
+      newThisWeek: 7,
+      newThisWeekWithoutNames: 6,
+      knownWordCount: 250,
+      knownWordCountWithoutNames: 249,
+    });
+  });
+
+  it('GET /api/stats/vocabulary/charts returns complete chart datasets', async () => {
+    const app = createStatsApp(createMockTracker());
+
+    const res = await app.request('/api/stats/vocabulary/charts');
+
+    assert.equal(res.status, 200);
+    assert.deepEqual(await res.json(), {
+      ready: true,
+      topWords: [{ wordId: 1, headword: 'する', frequency: 50 }],
+      topWordsWithoutNames: [{ wordId: 1, headword: 'する', frequency: 50 }],
+      newWordsTimeline: [{ epochDay: 20_000, wordCount: 3 }],
+      newWordsTimelineWithoutNames: [{ epochDay: 20_000, wordCount: 3 }],
+    });
+  });
+
   it('GET /api/stats/kanji returns kanji frequency data', async () => {
     const app = createStatsApp(createMockTracker());
     const res = await app.request('/api/stats/kanji');

src/core/services/anki-jimaku.ts

@@ -65,6 +65,7 @@ export interface AnkiJimakuIpcRuntimeOptions {
   getYoutubeMediaSourceUrl?: () => Promise<string | null | undefined> | string | null | undefined;
   showDesktopNotification: (title: string, options: { body?: string; icon?: string }) => void;
   showOverlayNotification?: (payload: OverlayNotificationPayload) => void;
+  dismissOverlayNotification?: (id: string) => void;
   createFieldGroupingCallback: () => (
     data: KikuFieldGroupingRequestData,
   ) => Promise<KikuFieldGroupingChoice>;
@@ -166,6 +167,7 @@ export function registerAnkiJimakuIpcRuntime(
           options.getCachedMediaPath,
           options.shouldRequireRemoteMediaCache,
           options.getYoutubeMediaSourceUrl,
+          options.dismissOverlayNotification,
         );
         integration.start();
         options.setAnkiIntegration(integration);

src/core/services/ass-text.test.ts

@@ -10,6 +10,8 @@ import {
   isAssTemporalCommand,
   normalizePlainSubtitleText,
   parseAssEffectField,
+  removeLiveGlyphFragmentLines,
+  removeAssControlDebrisLines,
 } from './ass-text';
 
 test('assToPlainText drops vector drawing runs', () => {
@@ -74,6 +76,14 @@ test('assToPlainText normalizes CRLF before converting', () => {
   assert.equal(assToPlainText('一行目\r\n二行目'), '一行目\n二行目');
 });
 
+test('removeAssControlDebrisLines drops malformed spacer resets without eating dialogue', () => {
+  assert.equal(
+    removeAssControlDebrisLines('Visible line\n\\\n{\\fr0\n\\{\\frz287.5'),
+    'Visible line',
+  );
+  assert.equal(removeAssControlDebrisLines('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
+});
+
 test('normalizePlainSubtitleText settles whitespace without decoding ASS', () => {
   // A brace reaching this layer is literal text mpv chose to show, not markup.
   assert.equal(normalizePlainSubtitleText('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
@@ -193,3 +203,33 @@ test('isAnimatedAssEffectKind covers the stock animated effects only', () => {
   assert.equal(isAnimatedAssEffectKind('other'), false);
   assert.equal(isAnimatedAssEffectKind('none'), false);
 });
+
+test('removeLiveGlyphFragmentLines drops a per-glyph typesetting wall and its syllable', () => {
+  const wall = [...'wansdumretoikhI'].join('\n');
+  assert.equal(removeLiveGlyphFragmentLines(`${wall}\ntai`), '');
+});
+
+test('removeLiveGlyphFragmentLines keeps concurrent dialogue beside a glyph wall', () => {
+  const wall = [...'wansdumretoikhI'].join('\n');
+  assert.equal(removeLiveGlyphFragmentLines(`${wall}\nそれよりも　ノート…`), 'それよりも　ノート…');
+});
+
+test('removeLiveGlyphFragmentLines leaves ordinary short lines alone', () => {
+  const text = 'え\nはい。\nそうだな';
+  assert.equal(removeLiveGlyphFragmentLines(text), text);
+});
+
+test('normalizePlainSubtitleText folds cue-boundary blank lines for text consumers', () => {
+  // The display layer splits on the blank line before normalizing; everyone else --
+  // tokenizer, cache key, dedup gate, mined sentence -- wants the plain line form.
+  assert.equal(
+    normalizePlainSubtitleText('\u4e00\u884c\u76ee\n\n\u4e8c\u884c\u76ee'),
+    '\u4e00\u884c\u76ee\n\u4e8c\u884c\u76ee',
+  );
+  assert.equal(
+    normalizePlainSubtitleText('\u4e00\u884c\u76ee\n\n\u4e8c\u884c\u76ee', {
+      collapseLineBreaks: true,
+    }),
+    '\u4e00\u884c\u76ee \u4e8c\u884c\u76ee',
+  );
+});

src/core/services/ass-text.ts

@@ -91,6 +91,41 @@ export function assToPlainText(text: string, lineBreak: AssLineBreak = '\n'): st
   return resolveWhitespaceEscapes(stripAssMarkup(text.replace(/\r\n/g, '\n')), lineBreak);
 }
 
+const MALFORMED_ASS_ROTATION_RESET = /^\\?\{\\(?:fr|frx|fry|frz|fax|fay)[-+.0-9]*$/u;
+
+/**
+ * Drop non-rendering spacer events left as literal text by a malformed, unclosed ASS
+ * rotation reset. These events otherwise become repeated `\\` or `{\\fr0` subtitle
+ * lines after mpv-compatible decoding.
+ */
+export function removeAssControlDebrisLines(text: string): string {
+  return text
+    .split('\n')
+    .filter((line) => {
+      const compact = line.replace(/\s+/gu, '');
+      return compact !== '\\' && !MALFORMED_ASS_ROTATION_RESET.test(compact);
+    })
+    .join('\n');
+}
+
+const MIN_GLYPH_BURST_LINES = 6;
+const MAX_GLYPH_BURST_COMPANION_GLYPHS = 3;
+
+/**
+ * Per-glyph karaoke typesetting flattened into live text becomes a wall of
+ * single-character lines plus the short syllable currently being typed. No authored
+ * subtitle stacks this many one-glyph lines at once, so when the wall is present drop
+ * it and its short companion fragments while keeping any concurrent dialogue line.
+ */
+export function removeLiveGlyphFragmentLines(text: string): string {
+  const lines = text.split('\n');
+  const singleGlyphLines = lines.filter((line) => [...line.trim()].length === 1).length;
+  if (singleGlyphLines < MIN_GLYPH_BURST_LINES) return text;
+  return lines
+    .filter((line) => [...line.trim()].length > MAX_GLYPH_BURST_COMPANION_GLYPHS)
+    .join('\n');
+}
+
 export interface NormalizePlainSubtitleTextOptions {
   /** Fold every line break into a single space. */
   collapseLineBreaks?: boolean;
@@ -118,6 +153,10 @@ export function normalizePlainSubtitleText(
   );
   if (collapseLineBreaks) {
     normalized = normalized.replace(/\n/g, ' ').replace(/\s+/g, ' ');
+  } else {
+    // Simultaneous cues reach the display layer separated by a blank line; every other
+    // consumer wants the plain one-break-per-line form.
+    normalized = normalized.replace(/\n{2,}/g, '\n');
   }
 
   return trim ? normalized.trim() : normalized;

src/core/services/immersion-tracker-service.test.ts

@@ -559,6 +559,241 @@ test('fresh tracker DB creates lifetime summary tables', async () => {
   }
 });
 
+test('fresh tracker DB skips lexical rollup backfill work', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let backfillRuns = 0;
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor({ dbPath }, {
+      runLexicalRollupBackfillTask: async () => {
+        backfillRuns += 1;
+      },
+    } as never);
+
+    assert.equal(backfillRuns, 0);
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('tracker starts the injected lexical rollup backfill when it is pending', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let backfillRuns = 0;
+
+  try {
+    const setupDb = new Database(dbPath);
+    const { ensureSchema } = await import('./immersion-tracker/storage');
+    ensureSchema(setupDb);
+    setupDb
+      .prepare(
+        `UPDATE imm_rollup_state SET state_value = '0' WHERE state_key = 'lexical_daily_rollups_version'`,
+      )
+      .run();
+    setupDb.close();
+
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor({ dbPath }, {
+      runLexicalRollupBackfillTask: async () => {
+        backfillRuns += 1;
+      },
+    } as never);
+
+    assert.equal(backfillRuns, 1);
+    await waitForCondition(
+      () => !(tracker as unknown as { writeLock: { locked: boolean } }).writeLock.locked,
+    );
+    assert.equal(
+      (tracker as unknown as { preserveWriteQueueUntilDrained: boolean })
+        .preserveWriteQueueUntilDrained,
+      false,
+    );
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('tracker runs startup session-rollup maintenance before lexical backfill locks writes', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let releaseBackfill = (): void => {};
+  const heldBackfill = new Promise<void>((resolve) => {
+    releaseBackfill = resolve;
+  });
+
+  try {
+    const startedAtMs = trackerNowMs() - 60_000;
+    const endedAtMs = trackerNowMs();
+    const setupDb = new Database(dbPath);
+    const { ensureSchema } = await import('./immersion-tracker/storage');
+    ensureSchema(setupDb);
+    setupDb.exec(`
+      INSERT INTO imm_videos (
+        video_id, video_key, canonical_title, source_type, duration_ms, CREATED_DATE, LAST_UPDATE_DATE
+      ) VALUES (1, 'local:/tmp/rollup-recovery.mkv', 'Rollup Recovery', 1, 0, '1', '1');
+      INSERT INTO imm_sessions (
+        session_id, session_uuid, video_id, started_at_ms, ended_at_ms, status,
+        active_watched_ms, lines_seen, tokens_seen, cards_mined, CREATED_DATE, LAST_UPDATE_DATE
+      ) VALUES (
+        1, 'rollup-recovery', 1, '${startedAtMs}', '${endedAtMs}', 2,
+        60000, 10, 20, 2, '${startedAtMs}', '${endedAtMs}'
+      );
+      INSERT INTO imm_session_telemetry (
+        session_id, sample_ms, total_watched_ms, active_watched_ms, lines_seen,
+        tokens_seen, cards_mined, lookup_count, lookup_hits, CREATED_DATE, LAST_UPDATE_DATE
+      ) VALUES (
+        1, '${endedAtMs}', 60000, 60000, 10, 20, 2, 0, 0,
+        '${endedAtMs}', '${endedAtMs}'
+      );
+      DELETE FROM imm_daily_rollups;
+      DELETE FROM imm_monthly_rollups;
+      UPDATE imm_rollup_state SET state_value = '0';
+    `);
+    setupDb.close();
+
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor({ dbPath }, {
+      runLexicalRollupBackfillTask: async () => heldBackfill,
+    } as never);
+
+    const privateApi = tracker as unknown as {
+      db: DatabaseSync;
+      writeLock: { locked: boolean };
+    };
+    assert.equal(privateApi.writeLock.locked, true);
+    assert.equal(
+      (
+        privateApi.db.prepare('SELECT COUNT(*) AS total FROM imm_daily_rollups').get() as {
+          total: number;
+        }
+      ).total,
+      1,
+    );
+    assert.equal(
+      (
+        privateApi.db.prepare('SELECT COUNT(*) AS total FROM imm_monthly_rollups').get() as {
+          total: number;
+        }
+      ).total,
+      1,
+    );
+  } finally {
+    releaseBackfill();
+    if (tracker) {
+      await waitForCondition(
+        () => !(tracker as unknown as { writeLock: { locked: boolean } }).writeLock.locked,
+      );
+    }
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('tracker queues playback writes until lexical rollup backfill settles', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let startBackfill = (): void => {};
+  let releaseBackfill = (): void => {};
+  let markBackfillStarted = (): void => {};
+  const backfillStartGate = new Promise<void>((resolve) => {
+    startBackfill = resolve;
+  });
+  const heldBackfill = new Promise<void>((resolve) => {
+    releaseBackfill = resolve;
+  });
+  const backfillStarted = new Promise<void>((resolve) => {
+    markBackfillStarted = resolve;
+  });
+
+  try {
+    const setupDb = new Database(dbPath);
+    const { ensureSchema } = await import('./immersion-tracker/storage');
+    ensureSchema(setupDb);
+    setupDb
+      .prepare(
+        `UPDATE imm_rollup_state SET state_value = '0' WHERE state_key = 'lexical_daily_rollups_version'`,
+      )
+      .run();
+    setupDb.close();
+
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath, policy: { queueCap: 100 } },
+      {
+        runLexicalRollupBackfillTask: async (workerDbPath) => {
+          await backfillStartGate;
+          const workerDb = new Database(workerDbPath);
+          try {
+            workerDb.exec('BEGIN IMMEDIATE');
+            markBackfillStarted();
+            await heldBackfill;
+            workerDb.exec('COMMIT');
+          } catch (error) {
+            try {
+              workerDb.exec('ROLLBACK');
+            } catch {
+              // Preserve the original worker failure.
+            }
+            throw error;
+          } finally {
+            workerDb.close();
+          }
+        },
+      },
+    );
+    tracker.handleMediaChange('https://example.com/backfill-test.mp4', 'Backfill Test');
+    startBackfill();
+    await backfillStarted;
+    for (let index = 0; index < 125; index += 1) tracker.recordCardsMined(1);
+
+    const privateApi = tracker as unknown as {
+      db: DatabaseSync;
+      queue: unknown[];
+      droppedWriteCount: number;
+      flushNow: () => void;
+      writeLock: { locked: boolean };
+    };
+    assert.equal(privateApi.writeLock.locked, true);
+    privateApi.flushNow();
+
+    assert.ok(privateApi.queue.length > 100, 'the protected queue may grow past its normal cap');
+    assert.equal(privateApi.droppedWriteCount, 0, 'backfill must not discard playback writes');
+    assert.equal(
+      (
+        privateApi.db.prepare('SELECT COUNT(*) AS total FROM imm_session_events').get() as {
+          total: number;
+        }
+      ).total,
+      0,
+    );
+
+    releaseBackfill();
+    await waitForCondition(() => privateApi.queue.length === 0, 5_000);
+    assert.equal(
+      (
+        privateApi.db.prepare('SELECT COUNT(*) AS total FROM imm_session_events').get() as {
+          total: number;
+        }
+      ).total,
+      125,
+    );
+  } finally {
+    releaseBackfill();
+    if (tracker) {
+      await waitForCondition(
+        () => !(tracker as unknown as { writeLock: { locked: boolean } }).writeLock.locked,
+        5_000,
+      );
+    }
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
 test('startup backfills lifetime summaries when retained sessions exist but summary tables are empty', async () => {
   const dbPath = makeDbPath();
   let tracker: ImmersionTrackerService | null = null;
@@ -4909,3 +5144,149 @@ test('ensureAnimeCoverArt fetches art via the latest video of the anime', async
     cleanupDbPath(dbPath);
   }
 });
+
+test('getVocabularySummary coalesces concurrent requests into one worker task', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let taskRuns = 0;
+  let releaseTask: (() => void) | null = null;
+  const seenKnownWords: Array<ReadonlySet<string> | null> = [];
+  const summary = {
+    uniqueWords: 1,
+    uniqueWordsWithoutNames: 1,
+    uniqueKanji: 0,
+    newThisWeek: 0,
+    newThisWeekWithoutNames: 0,
+    knownWordCount: null,
+    knownWordCountWithoutNames: null,
+  };
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runVocabularySummaryTask: async (_dbPath, knownWords) => {
+          taskRuns += 1;
+          seenKnownWords.push(knownWords);
+          await new Promise<void>((resolve) => {
+            releaseTask = resolve;
+          });
+          return summary;
+        },
+        destroyVocabularySummaryRunner: () => {},
+      },
+    );
+
+    const knownWordsSnapshot = new Set(['猫']);
+    const first = tracker.getVocabularySummary(knownWordsSnapshot);
+    const second = tracker.getVocabularySummary(knownWordsSnapshot);
+    await waitForCondition(() => releaseTask !== null);
+    let release = releaseTask as (() => void) | null;
+    assert.ok(release);
+    release();
+    assert.deepEqual(await first, summary);
+    assert.equal(await second, await first);
+    assert.equal(taskRuns, 1);
+    assert.deepEqual(seenKnownWords, [knownWordsSnapshot]);
+
+    releaseTask = null;
+    const third = tracker.getVocabularySummary(null);
+    await waitForCondition(() => releaseTask !== null);
+    release = releaseTask as (() => void) | null;
+    assert.ok(release);
+    release();
+    assert.deepEqual(await third, summary);
+    assert.equal(taskRuns, 2);
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('getVocabularySummary coalesces equivalent known-word snapshots by value', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let taskRuns = 0;
+  const releases: Array<() => void> = [];
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runVocabularySummaryTask: async () => {
+          taskRuns += 1;
+          await new Promise<void>((resolve) => releases.push(resolve));
+          return {
+            uniqueWords: 2,
+            uniqueWordsWithoutNames: 2,
+            uniqueKanji: 2,
+            newThisWeek: 0,
+            newThisWeekWithoutNames: 0,
+            knownWordCount: 2,
+            knownWordCountWithoutNames: 2,
+          };
+        },
+        destroyVocabularySummaryRunner: () => {},
+      },
+    );
+
+    const first = tracker.getVocabularySummary(new Set(['猫', '犬']));
+    const second = tracker.getVocabularySummary(new Set(['犬', '猫']));
+    await waitForCondition(() => releases.length > 0);
+    const observedTaskRuns = taskRuns;
+    for (const release of releases) release();
+    await Promise.all([first, second]);
+
+    assert.equal(observedTaskRuns, 1);
+
+    const third = tracker.getVocabularySummary(new Set(['猫', '犬']));
+    await waitForCondition(() => releases.length === 2);
+    releases[1]!();
+    await third;
+    assert.equal(taskRuns, 2, 'a settled snapshot must be evicted from the in-flight map');
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('getVocabularySummary keeps different known-word snapshots independent', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const releases: Array<() => void> = [];
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runVocabularySummaryTask: async (_dbPath, knownWords) => {
+          await new Promise<void>((resolve) => releases.push(resolve));
+          return {
+            uniqueWords: 1,
+            uniqueWordsWithoutNames: 1,
+            uniqueKanji: 0,
+            newThisWeek: 0,
+            newThisWeekWithoutNames: 0,
+            knownWordCount: knownWords?.size ?? null,
+            knownWordCountWithoutNames: knownWords?.size ?? null,
+          };
+        },
+        destroyVocabularySummaryRunner: () => {},
+      },
+    );
+
+    const withoutKnownWords = tracker.getVocabularySummary(null);
+    const withKnownWords = tracker.getVocabularySummary(new Set(['猫']));
+    await waitForCondition(() => releases.length === 2);
+    for (const release of releases) release();
+
+    assert.equal((await withoutKnownWords).knownWordCount, null);
+    assert.equal((await withKnownWords).knownWordCount, 1);
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});

src/core/services/immersion-tracker-service.ts

@@ -58,6 +58,7 @@ import {
   getSessionEvents,
   getSimilarWords,
   getStatsExcludedWords,
+  getVocabularyChartData,
   getVocabularyStats,
   replaceStatsExcludedWords,
   searchSubtitleSentences,
@@ -96,6 +97,12 @@ import {
   DeleteMaintenanceWorkerRuntime,
   type RunDeleteMaintenanceTask,
 } from './immersion-tracker/delete-maintenance-worker-runtime';
+import {
+  VocabularySummaryWorkerRuntime,
+  type RunVocabularySummaryTask,
+} from './immersion-tracker/vocabulary-summary-worker-runtime';
+import { LexicalRollupWorkerRuntime } from './immersion-tracker/lexical-rollup-worker-runtime';
+import { areLexicalDailyRollupsReady } from './immersion-tracker/lexical-rollups';
 import { DeleteMaintenanceScheduler } from './immersion-tracker/delete-maintenance-scheduler';
 import {
   cleanupDuplicateSubtitleLines,
@@ -185,6 +192,7 @@ import {
   type StatsExcludedWordRow,
   type StreakCalendarRow,
   type VocabularyCleanupSummary,
+  type VocabularyStatsSummary,
   type WatchTimePerAnimeRow,
   type WordAnimeAppearanceRow,
   type WordDetailRow,
@@ -405,13 +413,24 @@ export class ImmersionTrackerService {
   private readonly monthlyRollupRetentionMs: number;
   private readonly vacuumIntervalMs: number;
   private readonly dbPath: string;
-  private readonly writeLock = { locked: false };
+  private readonly writeLock = {
+    locked: false,
+    reasons: new Set<'flush' | 'delete-maintenance' | 'lexical-rollup-backfill'>(),
+  };
   private readonly destroyDeleteMaintenanceRunner: () => void;
+  private readonly runVocabularySummaryTask: (
+    knownWords: ReadonlySet<string> | null,
+  ) => Promise<VocabularyStatsSummary>;
+  private readonly vocabularySummariesInFlight = new Map<string, Promise<VocabularyStatsSummary>>();
+  private readonly destroyVocabularySummaryRunner: () => void;
+  private readonly runLexicalRollupBackfillTask: () => Promise<void>;
+  private readonly destroyLexicalRollupBackfillRunner: () => void;
   private readonly deleteMaintenanceScheduler: DeleteMaintenanceScheduler;
   private flushTimer: ReturnType<typeof setTimeout> | null = null;
   private maintenanceTimer: ReturnType<typeof setInterval> | null = null;
   private flushScheduled = false;
   private droppedWriteCount = 0;
+  private preserveWriteQueueUntilDrained = false;
   private lastVacuumMs = 0;
   private isDestroyed = false;
   private sessionState: SessionState | null = null;
@@ -434,6 +453,10 @@ export class ImmersionTrackerService {
     dependencies: {
       runDeleteMaintenanceTask?: RunDeleteMaintenanceTask;
       destroyDeleteMaintenanceRunner?: () => void;
+      runVocabularySummaryTask?: RunVocabularySummaryTask;
+      destroyVocabularySummaryRunner?: () => void;
+      runLexicalRollupBackfillTask?: (dbPath: string) => Promise<void>;
+      destroyLexicalRollupBackfillRunner?: () => void;
     } = {},
   ) {
     this.dbPath = options.dbPath;
@@ -453,13 +476,34 @@ export class ImmersionTrackerService {
       runTask: (task) => runDeleteMaintenanceTask(this.dbPath, task),
       onBusy: () => {
         this.requireWriteQueueDrained('delete maintenance');
-        this.writeLock.locked = true;
+        this.setWriteLock('delete-maintenance', true);
       },
       onIdle: () => {
-        this.writeLock.locked = false;
+        this.setWriteLock('delete-maintenance', false);
         if (!this.isDestroyed && this.queue.length > 0) this.scheduleFlush(0);
       },
     });
+    if (dependencies.runVocabularySummaryTask) {
+      this.runVocabularySummaryTask = (knownWords) =>
+        dependencies.runVocabularySummaryTask!(this.dbPath, knownWords);
+      this.destroyVocabularySummaryRunner =
+        dependencies.destroyVocabularySummaryRunner ?? (() => {});
+    } else {
+      const vocabularySummaryRuntime = new VocabularySummaryWorkerRuntime();
+      this.runVocabularySummaryTask = (knownWords) =>
+        vocabularySummaryRuntime.run(this.dbPath, knownWords);
+      this.destroyVocabularySummaryRunner = () => vocabularySummaryRuntime.destroy();
+    }
+    if (dependencies.runLexicalRollupBackfillTask) {
+      this.runLexicalRollupBackfillTask = () =>
+        dependencies.runLexicalRollupBackfillTask!(this.dbPath);
+      this.destroyLexicalRollupBackfillRunner =
+        dependencies.destroyLexicalRollupBackfillRunner ?? (() => {});
+    } else {
+      const lexicalRollupRuntime = new LexicalRollupWorkerRuntime();
+      this.runLexicalRollupBackfillTask = () => lexicalRollupRuntime.run(this.dbPath);
+      this.destroyLexicalRollupBackfillRunner = () => lexicalRollupRuntime.destroy();
+    }
     const parentDir = path.dirname(this.dbPath);
     if (!fs.existsSync(parentDir)) {
       fs.mkdirSync(parentDir, { recursive: true });
@@ -548,6 +592,7 @@ export class ImmersionTrackerService {
     }
     this.preparedStatements = createTrackerPreparedStatements(this.db);
     this.scheduleMaintenance();
+    if (!areLexicalDailyRollupsReady(this.db)) this.startLexicalRollupBackfill();
     this.scheduleFlush();
   }
 
@@ -565,6 +610,8 @@ export class ImmersionTrackerService {
     this.isDestroyed = true;
     this.deleteMaintenanceScheduler.destroy();
     this.destroyDeleteMaintenanceRunner();
+    this.destroyVocabularySummaryRunner();
+    this.destroyLexicalRollupBackfillRunner();
     this.db.close();
   }
 
@@ -634,6 +681,25 @@ export class ImmersionTrackerService {
     return getVocabularyStats(this.db, limit, excludePos);
   }
 
+  async getVocabularySummary(knownWords: ReadonlySet<string> | null) {
+    const key = knownWords ? JSON.stringify([...knownWords].sort()) : 'null';
+    const inFlight = this.vocabularySummariesInFlight.get(key);
+    if (inFlight) return inFlight;
+    const task = this.runVocabularySummaryTask(knownWords);
+    this.vocabularySummariesInFlight.set(key, task);
+    try {
+      return await task;
+    } finally {
+      if (this.vocabularySummariesInFlight.get(key) === task) {
+        this.vocabularySummariesInFlight.delete(key);
+      }
+    }
+  }
+
+  async getVocabularyChartData() {
+    return getVocabularyChartData(this.db);
+  }
+
   async getStatsExcludedWords(): Promise<StatsExcludedWordRow[]> {
     return getStatsExcludedWords(this.db);
   }
@@ -910,6 +976,33 @@ export class ImmersionTrackerService {
     }
   }
 
+  private setWriteLock(
+    reason: 'flush' | 'delete-maintenance' | 'lexical-rollup-backfill',
+    active: boolean,
+  ): void {
+    if (active) this.writeLock.reasons.add(reason);
+    else this.writeLock.reasons.delete(reason);
+    this.writeLock.locked = this.writeLock.reasons.size > 0;
+  }
+
+  private startLexicalRollupBackfill(): void {
+    this.requireWriteQueueDrained('lexical rollup backfill');
+    this.preserveWriteQueueUntilDrained = true;
+    this.setWriteLock('lexical-rollup-backfill', true);
+    void this.runLexicalRollupBackfillTask()
+      .catch((error: unknown) => {
+        this.logger.warn(
+          'Lexical daily rollup backfill failed; it will retry on next startup',
+          error,
+        );
+      })
+      .finally(() => {
+        this.setWriteLock('lexical-rollup-backfill', false);
+        if (this.queue.length === 0) this.preserveWriteQueueUntilDrained = false;
+        else if (!this.isDestroyed) this.scheduleFlush(0);
+      });
+  }
+
   async reassignAnimeAnilist(
     animeId: number,
     info: {
@@ -1906,7 +1999,12 @@ export class ImmersionTrackerService {
 
   private recordWrite(write: QueuedWrite): void {
     if (this.isDestroyed) return;
-    const { dropped } = enqueueWrite(this.queue, write, this.queueCap);
+    // A lexical migration owns the database write lock, so dropping the oldest
+    // entry cannot relieve pressure: nothing can flush until the worker exits.
+    // Preserve that finite startup burst and drain it as soon as the lock lifts.
+    const { dropped } = this.preserveWriteQueueUntilDrained
+      ? (this.queue.push(write), { dropped: 0 })
+      : enqueueWrite(this.queue, write, this.queueCap);
     if (dropped > 0) {
       this.droppedWriteCount += dropped;
       this.logger.warn(`Immersion tracker queue overflow; dropped ${dropped} oldest writes`);
@@ -1954,6 +2052,7 @@ export class ImmersionTrackerService {
   private flushNow(): void {
     if (this.writeLock.locked || this.isDestroyed) return;
     if (this.queue.length === 0) {
+      this.preserveWriteQueueUntilDrained = false;
       this.flushScheduled = false;
       return;
     }
@@ -1965,7 +2064,7 @@ export class ImmersionTrackerService {
     }
 
     const batch = this.queue.splice(0, Math.min(this.batchSize, this.queue.length));
-    this.writeLock.locked = true;
+    this.setWriteLock('flush', true);
     try {
       this.db.exec('BEGIN IMMEDIATE');
       for (const write of batch) {
@@ -1977,8 +2076,9 @@ export class ImmersionTrackerService {
       this.queue.unshift(...batch);
       this.logger.warn('Immersion tracker flush failed, retrying later', error as Error);
     } finally {
-      this.writeLock.locked = false;
+      this.setWriteLock('flush', false);
       this.flushScheduled = false;
+      if (this.queue.length === 0) this.preserveWriteQueueUntilDrained = false;
       if (this.queue.length > 0) {
         this.scheduleFlush(this.flushIntervalMs);
       }

src/core/services/immersion-tracker/__tests__/duplicate-line-cleanup.test.ts

@@ -1,7 +1,4 @@
 import assert from 'node:assert/strict';
-import fs from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
 import test from 'node:test';
 import { Database } from '../sqlite.js';
 import type { DatabaseSync } from '../sqlite.js';
@@ -21,17 +18,6 @@ interface SeedLine {
   createdMs?: number;
 }
 
-function makeDbPath(): string {
-  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-duplicate-line-test-'));
-  return path.join(dir, 'immersion.sqlite');
-}
-
-function cleanupDbPath(dbPath: string): void {
-  const dir = path.dirname(dbPath);
-  if (!fs.existsSync(dir)) return;
-  fs.rmSync(dir, { recursive: true, force: true });
-}
-
 /** One episode, two sessions of it, and one word occurrence per seeded line. */
 function seed(db: DatabaseSync, lines: SeedLine[]): void {
   db.exec(`
@@ -82,12 +68,16 @@ function seed(db: DatabaseSync, lines: SeedLine[]): void {
   `);
 }
 
-function createDb(lines: SeedLine[]): { db: DatabaseSync; dbPath: string } {
-  const dbPath = makeDbPath();
-  const db = new Database(dbPath);
+/**
+ * These tests exercise the cleanup SQL, not durability. A fresh on-disk database per
+ * test pays a schema-creation fsync that is cheap on a local NVMe but slow enough on CI
+ * runners to blow the 5s per-test timeout, so the database stays in memory.
+ */
+function createDb(lines: SeedLine[]): { db: DatabaseSync } {
+  const db = new Database(':memory:');
   ensureSchema(db);
   seed(db, lines);
-  return { db, dbPath };
+  return { db };
 }
 
 /** A typeset line mpv reported once per animation frame. */
@@ -119,7 +109,7 @@ function wordFrequency(db: DatabaseSync): number {
 }
 
 test('a karaoke burst collapses to one line and gives back its word counts', () => {
-  const { db, dbPath } = createDb([
+  const { db } = createDb([
     ...karaokeFrames(1, '飛び上がる', 10_000, 40, 40),
     { session: 1, text: 'おはよう', startMs: 20_000, endMs: 22_000 },
   ]);
@@ -148,7 +138,6 @@ test('a karaoke burst collapses to one line and gives back its word counts', ()
     assert.equal(summary.samples[0]!.videoTitle, 'Ep 1');
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
@@ -160,7 +149,7 @@ test('ordinary repeated dialogue survives', () => {
     startMs: 5_000 + index * 800,
     endMs: 5_000 + (index + 1) * 800,
   }));
-  const { db, dbPath } = createDb(lines);
+  const { db } = createDb(lines);
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -171,14 +160,13 @@ test('ordinary repeated dialogue survives', () => {
     assert.equal(wordFrequency(db), 6);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a long run of quarter-second frames is still a burst', () => {
   // Between the timing-only bound (0.1s) and the animation-frame bound (0.3s): heavier
   // typesetting lands here, and the run length is what makes it conclusive.
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 250));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 250));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -189,12 +177,11 @@ test('a long run of quarter-second frames is still a burst', () => {
     assert.equal(wordFrequency(db), 1);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a qualifying short-frame burst may end with one long hold frame', () => {
-  const { db, dbPath } = createDb([
+  const { db } = createDb([
     ...karaokeFrames(1, '飛び上がる', 10_000, 8, 40),
     { session: 1, text: '飛び上がる', startMs: 10_320, endMs: 12_320 },
   ]);
@@ -208,12 +195,11 @@ test('a qualifying short-frame burst may end with one long hold frame', () => {
     assert.equal(wordFrequency(db), 1);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a long event before the final frame prevents burst cleanup', () => {
-  const { db, dbPath } = createDb([
+  const { db } = createDb([
     ...karaokeFrames(1, '飛び上がる', 10_000, 5, 40),
     { session: 1, text: '飛び上がる', startMs: 10_200, endMs: 12_200 },
     { session: 1, text: '飛び上がる', startMs: 12_200, endMs: 12_240 },
@@ -226,12 +212,11 @@ test('a long event before the final frame prevents burst cleanup', () => {
     assert.equal(countLines(db), 7);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a run of frames longer than the animation bound survives', () => {
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 400));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 400));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -240,7 +225,6 @@ test('a run of frames longer than the animation bound survives', () => {
     assert.equal(countLines(db), 6);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
@@ -248,7 +232,7 @@ test('the four-frame residue the live gate stores is cleaned up', () => {
   // The streaming gate records the first four frames of a burst before the run is long
   // enough to recognise. Four contiguous identical events under the strict timing-only
   // bound are that residue, and no real dialogue.
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 4, 40));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 4, 40));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -259,14 +243,13 @@ test('the four-frame residue the live gate stores is cleaned up', () => {
     assert.equal(wordFrequency(db), 1);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a four-frame run above the strict frame bound survives', () => {
   // Long enough per event to be plausible dialogue; only a five-event run may use the
   // looser animation-frame bound.
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 4, 250));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 4, 250));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -275,14 +258,13 @@ test('a four-frame run above the strict frame bound survives', () => {
     assert.equal(countLines(db), 4);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('an explicit minRunLength raises the bar', () => {
   // Five quarter-second frames qualify under the defaults; a cautious run asking for six
   // leaves them alone. Above the strict bound, so the residue rule stays out of it.
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 5, 250));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 5, 250));
 
   try {
     const preview = cleanupDuplicateSubtitleLines(db, { dryRun: true });
@@ -293,12 +275,11 @@ test('an explicit minRunLength raises the bar', () => {
     assert.equal(countLines(db), 5);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('an explicit maxFrameSeconds tightens the frame bound', () => {
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 250));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 250));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db, { maxFrameSeconds: 0.2 });
@@ -307,13 +288,12 @@ test('an explicit maxFrameSeconds tightens the frame bound', () => {
     assert.equal(countLines(db), 6);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a non-finite maxFrameSeconds falls back to the default bound', () => {
   // Six normal-beat lines: Infinity must not turn every event into a "short frame".
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 800));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 6, 800));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db, { maxFrameSeconds: Infinity });
@@ -322,12 +302,11 @@ test('a non-finite maxFrameSeconds falls back to the default bound', () => {
     assert.equal(countLines(db), 6);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('sampleLimit zero removes bursts but reports no samples', () => {
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 40, 40));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 40, 40));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db, { sampleLimit: 0 });
@@ -337,12 +316,11 @@ test('sampleLimit zero removes bursts but reports no samples', () => {
     assert.equal(countLines(db), 1);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a short run below every threshold survives', () => {
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 3, 40));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 3, 40));
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -351,7 +329,6 @@ test('a short run below every threshold survives', () => {
     assert.equal(countLines(db), 3);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
@@ -361,7 +338,7 @@ test('interleaved dual-line karaoke collapses each line to one row', () => {
   const kanji = karaokeFrames(1, '飛び上がる', 10_000, 20, 60);
   const romaji = karaokeFrames(1, 'tobiagaru', 10_001, 20, 60);
   const interleaved = [...kanji, ...romaji].sort((a, b) => a.startMs - b.startMs);
-  const { db, dbPath } = createDb(interleaved);
+  const { db } = createDb(interleaved);
 
   try {
     const summary = cleanupDuplicateSubtitleLines(db);
@@ -372,12 +349,11 @@ test('interleaved dual-line karaoke collapses each line to one row', () => {
     assert.equal(wordFrequency(db), 2);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('the same line in a rewatch session is never merged into the first watch', () => {
-  const { db, dbPath } = createDb([
+  const { db } = createDb([
     ...karaokeFrames(1, '飛び上がる', 10_000, 6, 40),
     ...karaokeFrames(2, '飛び上がる', 10_000, 6, 40),
   ]);
@@ -392,12 +368,11 @@ test('the same line in a rewatch session is never merged into the first watch',
     assert.equal(wordFrequency(db), 2);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a gap between runs splits them', () => {
-  const { db, dbPath } = createDb([
+  const { db } = createDb([
     ...karaokeFrames(1, '飛び上がる', 10_000, 6, 40),
     ...karaokeFrames(1, '飛び上がる', 60_000, 6, 40),
   ]);
@@ -409,12 +384,11 @@ test('a gap between runs splits them', () => {
     assert.equal(countLines(db), 2);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('a dry run reports what an apply would do and writes nothing', () => {
-  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 40, 40));
+  const { db } = createDb(karaokeFrames(1, '飛び上がる', 10_000, 40, 40));
 
   try {
     const preview = cleanupDuplicateSubtitleLines(db, { dryRun: true });
@@ -430,14 +404,13 @@ test('a dry run reports what an apply would do and writes nothing', () => {
     assert.equal(countLines(db), 1);
   } finally {
     db.close();
-    cleanupDbPath(dbPath);
   }
 });
 
 test('the lookback window leaves older bursts alone', () => {
   const recentMs = BASE_MS;
   const oldMs = BASE_MS - 40 * DAY_MS;
-  const { db, dbPath } = createDb([
+  const { db } = createDb([
     ...karaokeFrames(1, '飛び上がる', 10_000, 6, 40).map((line) => ({
       ...line,
       createdMs: oldMs,
@@ -462,6 +435,5 @@ test('the lookback window leaves older bursts alone', () => {
   } finally {
     globalThis.__subminerTestNowMs = undefined;
     db.close();
-    cleanupDbPath(dbPath);
   }
 });

src/core/services/immersion-tracker/__tests__/query.test.ts

@@ -31,6 +31,7 @@ import {
   getKanjiOccurrences,
   getSessionSummaries,
   getVocabularyStats,
+  getVocabularySummary,
   getKanjiStats,
   getSessionEvents,
   getSessionTimeline,
@@ -1875,6 +1876,115 @@ test('getVocabularyStats returns rows ordered by frequency descending', () => {
   }
 });
 
+test('getVocabularySummary counts every tracked vocabulary row instead of a display page', () => {
+  const dbPath = makeDbPath();
+  const db = openTestDb(dbPath);
+
+  try {
+    ensureSchema(db);
+    const nowSec = Math.floor(Date.now() / 1000);
+    const insertWord = db.prepare(`
+      INSERT INTO imm_words (
+        headword, word, reading, part_of_speech, pos1, pos2, pos3,
+        first_seen, last_seen, frequency
+      ) VALUES (?, ?, '', 'noun', '名詞', '一般', '', ?, ?, 1)
+    `);
+    const insertKanji = db.prepare(`
+      INSERT INTO imm_kanji (kanji, first_seen, last_seen, frequency)
+      VALUES (?, ?, ?, 1)
+    `);
+
+    for (let index = 0; index < 501; index += 1) {
+      insertWord.run(`単語${index}`, `単語${index}`, nowSec - 8 * 86_400, nowSec - 8 * 86_400);
+    }
+    for (let index = 0; index < 201; index += 1) {
+      insertKanji.run(
+        String.fromCodePoint(0x4e00 + index),
+        nowSec - 8 * 86_400,
+        nowSec - 8 * 86_400,
+      );
+    }
+    insertWord.run('今週', '今週', nowSec - 86_400, nowSec - 86_400);
+
+    assert.deepEqual(getVocabularySummary(db, new Set(['単語0', '今週']), nowSec * 1000), {
+      uniqueWords: 502,
+      uniqueWordsWithoutNames: 502,
+      uniqueKanji: 201,
+      newThisWeek: 1,
+      newThisWeekWithoutNames: 1,
+      knownWordCount: 2,
+      knownWordCountWithoutNames: 2,
+    });
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('getVocabularySummary applies vocabulary exclusions and Hide Names totals', () => {
+  const dbPath = makeDbPath();
+  const db = openTestDb(dbPath);
+
+  try {
+    ensureSchema(db);
+    const insertWord = db.prepare(`
+      INSERT INTO imm_words (
+        headword, word, reading, part_of_speech, pos1, pos2, pos3,
+        first_seen, last_seen, frequency
+      ) VALUES (?, ?, '', 'noun', '名詞', ?, '', 1, 1, 1)
+    `);
+    insertWord.run('猫', '猫', '一般');
+    insertWord.run('太郎', '太郎', '固有名詞');
+    insertWord.run('東京', '東京都', '一般');
+    db.prepare(
+      `
+      INSERT INTO imm_stats_excluded_words (headword, word, reading)
+      VALUES ('東京', '東京', '')
+    `,
+    ).run();
+
+    assert.deepEqual(getVocabularySummary(db, new Set(['猫', '太郎', '東京']), 9 * 86_400_000), {
+      uniqueWords: 2,
+      uniqueWordsWithoutNames: 1,
+      uniqueKanji: 0,
+      newThisWeek: 0,
+      newThisWeekWithoutNames: 0,
+      knownWordCount: 2,
+      knownWordCountWithoutNames: 1,
+    });
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('getVocabularySummary counts identically across id-keyed scan batches', () => {
+  const dbPath = makeDbPath();
+  const db = openTestDb(dbPath);
+
+  try {
+    ensureSchema(db);
+    const insertWord = db.prepare(`
+      INSERT INTO imm_words (
+        headword, word, reading, part_of_speech, pos1, pos2, pos3,
+        first_seen, last_seen, frequency
+      ) VALUES (?, ?, '', 'noun', '名詞', '一般', '', 1, 1, 1)
+    `);
+    for (let index = 0; index < 5; index += 1) {
+      insertWord.run(`単語${index}`, `単語${index}`);
+    }
+
+    const fullScan = getVocabularySummary(db, new Set(['単語0']), 9 * 86_400_000);
+    const batchedScan = getVocabularySummary(db, new Set(['単語0']), 9 * 86_400_000, 2);
+
+    assert.equal(fullScan.uniqueWords, 5);
+    assert.deepEqual(batchedScan, fullScan);
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
 test('getVocabularyStats filters rows that fail tokenizer vocabulary rules', () => {
   const dbPath = makeDbPath();
   const db = openTestDb(dbPath);