fix(stats): stop counting duplicate typeset subtitle lines (#191 )

chore(workflow): simplify agent tooling and docs
fix(media): tolerate slow MKV audio extraction (#195 )
2026-08-14 01:55:58 -07:00 · 2026-08-14 00:12:56 -07:00 · 2026-08-13 23:34:04 -07:00 · 2026-08-13 23:19:22 -07:00 · 2026-08-13 23:02:57 -07:00 · 2026-08-13 22:31:52 -07:00
201 changed files with 14943 additions and 7136 deletions
@@ -1,20 +0,0 @@
-{
-  "name": "subminer-local",
-  "interface": {
-    "displayName": "SubMiner Local"
-  },
-  "plugins": [
-    {
-      "name": "subminer-workflow",
-      "source": {
-        "source": "local",
-        "path": "./plugins/subminer-workflow"
-      },
-      "policy": {
-        "installation": "AVAILABLE",
-        "authentication": "ON_INSTALL"
-      },
-      "category": "Productivity"
-    }
-  ]
-}
@@ -1,22 +1,45 @@
 ---
-name: 'subminer-change-verification'
-description: 'Compatibility shim. Canonical SubMiner change verification workflow now lives in the repo-local subminer-workflow plugin.'
+name: subminer-change-verification
+description: Verify SubMiner changes with repo-native cheap-first test lanes. Use after code, config, launcher, plugin, runtime, stats, documentation, or workflow changes; do not use for read-only questions.
 ---

-# Compatibility Shim
+# SubMiner Change Verification

-Canonical source:
+Verify the behavior claimed by a change without running unrelated expensive checks by default.

- `plugins/subminer-workflow/skills/subminer-change-verification/SKILL.md`
+## Workflow

-Canonical helper scripts:
+1. Inspect the requested scope and changed paths with `git status --short` and `git diff`.
+2. Read `docs/workflow/verification.md` as the source of truth for maintained lanes.
+3. Run the cheapest lane or lanes that cover the changed behavior.
+4. Escalate to the full handoff gate only for substantial or cross-boundary changes.
+5. Report exact commands, results, skipped checks, blockers, and remaining risk.

- `plugins/subminer-workflow/skills/subminer-change-verification/scripts/classify_subminer_diff.sh`
- `plugins/subminer-workflow/skills/subminer-change-verification/scripts/verify_subminer_change.sh`
+Do not use hidden wrapper commands. Verification commands are owned by `package.json` and the workflow documentation.

-When this shim is invoked:
+## Lane Selection

-1. Read the canonical plugin-owned skill.
-2. Follow the plugin-owned skill as the source of truth.
-3. Use the wrapper scripts in this shim directory only for compatibility with existing commands and docs.
-4. Do not duplicate workflow changes here; update the plugin-owned skill and scripts instead.
+- Internal docs, `AGENTS.md`, or `.agents/skills/**`: `bun run test:docs:kb`
+- User-facing `docs-site/**`: `bun run docs:test`, then `bun run docs:build`
+- Config/schema/defaults: `bun run test:config`
+  - If defaults or templates changed, also run `bun run generate:config-example` and `bun run verify:config-example`.
+- General TypeScript source: `bun run typecheck`, then `bun run test:fast`
+- Launcher or mpv plugin: `bun run test:launcher` or `bun run test:env`, based on the behavior changed
+- Runtime compatibility or dist-sensitive wiring: `bun run test:runtime:compat`
+- Stats dashboard: `bun run test:stats`
+- Build/release scripts: `bun run test:scripts`
+
+For substantial changes, use the full gate documented in `AGENTS.md` and `docs/workflow/verification.md`.
+
+## Runtime Escalation
+
+Real runtime checks are required when the claim depends on actual Electron, mpv, overlay, focus, window tracking, launch, or socket behavior. Run the relevant application flow when the environment supports it. Otherwise, report the missing runtime dependency and do not present cheaper checks as authoritative runtime validation.
+
+## Pre-Handoff Checks
+
+Before handoff, reconcile both questions:
+
+1. Do behavior, defaults, flags, shortcuts, ports, APIs, architecture, or workflow changes require documentation updates?
+2. Does the change require a current-outcome fragment under `changes/` according to `changes/README.md`?
+
+Complete required updates before handoff or report the blocker.
@@ -1,13 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
-REPO_ROOT=$(cd "$SCRIPT_DIR/../../../.." && pwd)
-TARGET="$REPO_ROOT/plugins/subminer-workflow/skills/subminer-change-verification/scripts/classify_subminer_diff.sh"
-
-if [[ ! -x "$TARGET" ]]; then
-  echo "Missing canonical script: $TARGET" >&2
-  exit 1
-fi
-
-exec "$TARGET" "$@"
@@ -1,13 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
-REPO_ROOT=$(cd "$SCRIPT_DIR/../../../.." && pwd)
-TARGET="$REPO_ROOT/plugins/subminer-workflow/skills/subminer-change-verification/scripts/verify_subminer_change.sh"
-
-if [[ ! -x "$TARGET" ]]; then
-  echo "Missing canonical script: $TARGET" >&2
-  exit 1
-fi
-
-exec "$TARGET" "$@"
@@ -1,18 +0,0 @@
---
-name: 'subminer-scrum-master'
-description: 'Compatibility shim. Canonical SubMiner scrum-master workflow now lives in the repo-local subminer-workflow plugin.'
---
-
-# Compatibility Shim
-
-Canonical source:
-
- `plugins/subminer-workflow/skills/subminer-scrum-master/SKILL.md`
-
-When this shim is invoked:
-
-1. Read the canonical plugin-owned skill.
-2. Follow the plugin-owned skill as the source of truth.
-3. Do not duplicate workflow changes here; update the plugin-owned skill instead.
-
-This shim exists so existing repo references and prompts keep resolving during the migration to the repo-local plugin workflow.
@@ -1,4 +1,4 @@
-# AGENTS.MD
+# AGENTS.md

 ## Internal Docs

@@ -13,7 +13,7 @@ Start here, then leave this file.

 `docs-site/` is user-facing. Do not treat it as the canonical internal source of truth.

-`CLAUDE.md` is a symlink to this file — there is one project instruction file, not two.
+`CLAUDE.md` is a symlink to this file; there is one project instruction file, not two.

 ## Quick Start

@@ -25,8 +25,9 @@ Start here, then leave this file.

 ## Build / Test

- Runtime/package manager: Bun (`packageManager: bun@1.3.5`)
- Default handoff gate:
+- Runtime/package manager: Bun; use the version pinned by `package.json`.
+- Follow [`docs/workflow/verification.md`](./docs/workflow/verification.md) and start with the cheapest sufficient lane.
+- Full handoff gate for substantial changes:
  `bun run typecheck`
  `bun run test:fast`
  `bun run test:env`
@@ -44,13 +45,15 @@ Start here, then leave this file.
 - Runtime-compat / dist-sensitive: `bun run test:runtime:compat`
 - Stats dashboard UI (`stats/`): `bun run test:stats`
 - Build/release scripts (`scripts/**`): `bun run test:scripts`
- Docs-only: `bun run docs:test`, then `bun run docs:build`
+- Internal docs, `AGENTS.md`, or repo skills: `bun run test:docs:kb`
+- User-facing `docs-site/`: `bun run docs:test`, then `bun run docs:build`
+- macOS mpv window helper: `bun test scripts/get-mpv-window-macos.test.ts`
 - Test lanes are directory-discovered via `scripts/test-lanes.ts`; never hand-list test files in `package.json`

 ## Docs Upkeep

 - Docs ship with the change, not after. If a change alters behavior, defaults, flags, shortcuts, ports, or APIs, update the matching docs in the same PR. Touching code without reconciling its docs is an incomplete change.
- Source of truth for config defaults is the generated `config.example.jsonc`. Never write a default value into prose you didn't read from it — and don't restate the same default across multiple docs; cite/link to one place so there's a single thing to update.
+- Source of truth for config defaults is the generated `config.example.jsonc`. Never write a default value into prose you didn't read from it, and don't restate the same default across multiple docs; cite/link to one place so there's a single thing to update.
 - Trigger map (touch left → update right):
  - `src/config/definitions/**` (schema/defaults/template) → `bun run generate:config-example`, then reconcile `docs-site/configuration.md` + any feature doc that cites that default
  - shortcuts/keybindings (`shortcuts.*`, `keybindings`, `stats.*Key`, `subtitleSidebar.toggleKey`, controller bindings) → `docs-site/shortcuts.md`
@@ -71,16 +74,10 @@ Start here, then leave this file.

 ## Release / PR Notes

- User-visible PRs need reconciled current-outcome fragment(s) in `changes/*.md` — format and rules in [`changes/README.md`](./changes/README.md) (`type` + `area` keys required; inspect existing same-PR fragments, then update/remove stale bullets or add only genuinely separate outcomes; apply the `skip-changelog` label to opt out)
+- User-visible PRs need reconciled current-outcome fragment(s) in `changes/*.md`. Format and rules live in [`changes/README.md`](./changes/README.md) (`type` + `area` keys required; inspect existing same-PR fragments, then update/remove stale bullets or add only genuinely separate outcomes; apply the `skip-changelog` label to opt out).
 - User-visible docs changes get a `type: docs` fragment
 - CI enforces `bun run changelog:lint` and `bun run changelog:pr-check`
 - PR review helpers:
-  - `gh pr view --json number,title,url --jq '"PR #\\(.number): \\(.title)\\n\\(.url)"'`
+  - `gh pr view --json number,title --jq '"PR #\\(.number): \\(.title)"'`
  - `gh api repos/:owner/:repo/pulls/<num>/comments --paginate`
-
-## Runtime Notes
-
- Use Codex background for long jobs; tmux only when persistence/interaction is required
- CI red: `gh run list/view`, rerun, fix, repeat until green
- TypeScript: keep files small; follow existing patterns
- Only Swift is the `scripts/get-mpv-window-macos.swift` helper (macOS mpv window detection); validate via `bun test scripts/get-mpv-window-macos.test.ts`
+- For CI debugging, inspect runs with `gh run list/view`; rerun or fix only within the requested scope.
@@ -1,5 +1,29 @@
 # Changelog

+## v0.19.3 (2026-08-13)
+
+### Added
+- Changelog Modal: Adds an in-app changelog you can open from the tray ("View Changelog") or the "What's New" button on the update notification, so the notification stays reachable while you read. It shows the newest published release notes (falling back to the bundled changelog if that fetch fails), folds older versions while keeping the current one expanded, and supports keyboard navigation (`J`/`K`/arrows, `Enter`, `R`, `Esc`).
+
+### Changed
+- Subtitle Tokenization Performance: Reworks subtitle dictionary lookups to cut per-line work roughly in half, cache repeated lookups across lines, and stop tokenization from competing with on-screen subtitle prefetching. Also fixes several accuracy issues along the way: dropped readings on trailing kana, character names being skipped after a dictionary sync, annotations not refreshing after mining a card, and halfwidth katakana character names losing their reading or being swallowed by other words.
+
+### Fixed
+- Character Dictionary Large Imports: Large character dictionaries (e.g. One Piece) no longer fail to install from a fixed timeout budget; the import now scales its time budget to dictionary size and reports detailed progress (page/character counts, image download progress, elapsed time) instead of one static message.
+- Stats Delete Responsiveness: Deleting sessions, episodes, or library entries no longer freezes the stats page or an active video player; deletes are now batched into a single transaction.
+- Styled Subtitle Cue Parsing: Heavily typeset subtitles (karaoke, signs) no longer flood the subtitle sidebar with garbage; vector drawing commands are no longer shown as text, and duplicate/animation-burst cues now collapse into one.
+- X11 mpv Renderer: Fixes an mpv crash on the first fullscreen toggle for X11/XWayland users with `gpu-next` shaders (e.g. ArtCNN), which was caused by X11 mode forcing the legacy OpenGL renderer.
+- X11 Overlay Display Scaling: Fixes the overlay appearing oversized and offset from mpv on X11/XWayland under fractional or mixed-monitor display scaling.
+
+<details>
+<summary>Internal changes</summary>
+
+### Internal
+- Subtitle text is now decoded from ASS exactly once at ingest, so the renderer, timing tracker, and tokenizer all share one decoded value instead of each re-deriving it.
+- Added per-stage debug timings (`scanMs`, `mecabMs`, `frequencyMs`, `annotateMs`) to the subtitle tokenization pipeline log.
+
+</details>
+
 ## v0.19.2 (2026-08-04)

 ### Changed
@@ -0,0 +1 @@
+AGENTS.md
@@ -41,7 +41,7 @@
    "fast-uri": "3.1.5",
    "form-data": "4.0.6",
    "ip-address": "10.2.0",
-    "js-yaml": "4.3.0",
+    "js-yaml": "4.3.1",
    "lodash": "4.18.0",
    "minimatch": "10.2.5",
    "picomatch": "4.0.4",
@@ -498,7 +498,7 @@

    "jiti": ["jiti@2.6.1", "", { "bin": { "jiti": "lib/jiti-cli.mjs" } }, "sha512-ekilCSN1jwRvIbgeg/57YFh8qQDNbwDb9xT/qu2DAHbFFZUicIl4ygVaAvzveMhMVr3LnpSKTNnwt8PoOfmKhQ=="],

-    "js-yaml": ["js-yaml@4.3.0", "", { "dependencies": { "argparse": "^2.0.1" }, "bin": { "js-yaml": "bin/js-yaml.js" } }, "sha512-1td788aAnnZ5qs7V2QIRl1owjtYpbKt749Y3xauqQgwIIGF/xXWz1wMTEBx5O3LK3lXLVuqXPdPxj2BoFHaW9Q=="],
+    "js-yaml": ["js-yaml@4.3.1", "", { "dependencies": { "argparse": "^2.0.1" }, "bin": { "js-yaml": "bin/js-yaml.js" } }, "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ=="],

    "json-buffer": ["json-buffer@3.0.1", "", {}, "sha512-4bV5BfR2mqfQTJm+V5tPPdf+ZpuhiIvTuAB5g8kcrXOZpTT/QwwVRWBywX1ozr6lEuPdbHxwaJlm9G6mI2sfSQ=="],

@@ -0,0 +1,4 @@
+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`.
@@ -0,0 +1,5 @@
+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.
@@ -75,8 +75,8 @@ src/
  renderer/                # Overlay renderer (modularized UI/runtime)
    handlers/              # Keyboard/mouse/gamepad interaction modules
    modals/                # Modal flows (Jimaku, Kiku, subsync, runtime options, session help,
-                           #   character dictionary, playlist browser, subtitle sidebar,
-                           #   YouTube track picker, controller config/debug/select)
+                           #   changelog, character dictionary, playlist browser, subtitle
+                           #   sidebar, YouTube track picker, controller config/debug/select)
    positioning/           # Subtitle position controller (drag-to-reposition)
  settings/                # Settings window UI (model, controls, markup)
  types/                   # Domain type modules (anki, config, integrations, ...)
@@ -1,5 +1,29 @@
 # Changelog

+## v0.19.3 (2026-08-13)
+
+**Added**
+- Changelog Modal: Adds an in-app changelog you can open from the tray ("View Changelog") or the "What's New" button on the update notification, so the notification stays reachable while you read. It shows the newest published release notes (falling back to the bundled changelog if that fetch fails), folds older versions while keeping the current one expanded, and supports keyboard navigation (`J`/`K`/arrows, `Enter`, `R`, `Esc`).
+
+**Changed**
+- Subtitle Tokenization Performance: Reworks subtitle dictionary lookups to cut per-line work roughly in half, cache repeated lookups across lines, and stop tokenization from competing with on-screen subtitle prefetching. Also fixes several accuracy issues along the way: dropped readings on trailing kana, character names being skipped after a dictionary sync, annotations not refreshing after mining a card, and halfwidth katakana character names losing their reading or being swallowed by other words.
+
+**Fixed**
+- Character Dictionary Large Imports: Large character dictionaries (e.g. One Piece) no longer fail to install from a fixed timeout budget; the import now scales its time budget to dictionary size and reports detailed progress (page/character counts, image download progress, elapsed time) instead of one static message.
+- Stats Delete Responsiveness: Deleting sessions, episodes, or library entries no longer freezes the stats page or an active video player; deletes are now batched into a single transaction.
+- Styled Subtitle Cue Parsing: Heavily typeset subtitles (karaoke, signs) no longer flood the subtitle sidebar with garbage; vector drawing commands are no longer shown as text, and duplicate/animation-burst cues now collapse into one.
+- X11 mpv Renderer: Fixes an mpv crash on the first fullscreen toggle for X11/XWayland users with `gpu-next` shaders (e.g. ArtCNN), which was caused by X11 mode forcing the legacy OpenGL renderer.
+- X11 Overlay Display Scaling: Fixes the overlay appearing oversized and offset from mpv on X11/XWayland under fractional or mixed-monitor display scaling.
+
+<details>
+<summary>Internal changes</summary>
+
+**Internal**
+- Subtitle text is now decoded from ASS exactly once at ingest, so the renderer, timing tracker, and tokenizer all share one decoded value instead of each re-deriving it.
+- Added per-stage debug timings (`scanMs`, `mecabMs`, `frequencyMs`, `annotateMs`) to the subtitle tokenization pipeline log.
+
+</details>
+
 ## v0.19.2 (2026-08-04)

 **Changed**
@@ -34,7 +34,7 @@ The same immersion data powers the stats dashboard.
 - In-app overlay: focus the visible overlay, then press the key from `stats.toggleKey` (default: `` ` `` / `Backquote`).
 - Launcher command: run `subminer stats` to start the local stats server on demand (it also opens the dashboard in your browser when `stats.autoOpenBrowser` is enabled; the default is `false`).
 - Background server: run `subminer stats -b` to start or reuse a dedicated background stats daemon without keeping the launcher attached, and `subminer stats -s` to stop that daemon.
- Maintenance commands: run `subminer stats cleanup` or `subminer stats cleanup -v` to backfill/repair vocabulary metadata (`headword`, `reading`, POS) and purge stale or excluded rows from `imm_words` on demand; `subminer stats cleanup -l` repairs lifetime summary tables. `subminer stats rebuild` and `subminer stats backfill` rebuild or backfill rollup data.
+- Maintenance commands: run `subminer stats cleanup` or `subminer stats cleanup -v` to backfill/repair vocabulary metadata (`headword`, `reading`, POS) and purge stale or excluded rows from `imm_words` on demand; `subminer stats cleanup -l` repairs lifetime summary tables; `subminer stats cleanup --duplicate-lines` collapses repeated lines left behind by typeset subtitles (see [Repeated Line Cleanup](#repeated-line-cleanup)). `subminer stats rebuild` and `subminer stats backfill` rebuild or backfill rollup data.
 - Browser page: open `http://127.0.0.1:6969` directly if the local stats server is already running.

 ### Dashboard Tabs
@@ -125,6 +125,34 @@ Secondary subtitle text (typically English translations) is stored alongside pri

 The Vocabulary tab toolbar includes an **Exclusions** button for hiding words from all vocabulary views. Excluded words are stored in the immersion database, with older browser localStorage exclusions imported on first load after upgrade. They can be managed (restored or cleared) from the exclusion modal. Exclusions affect stat cards, charts, the frequency rank table, and the word list.

+### Repeated Line Cleanup
+
+Karaoke openings and animated signs are authored as one subtitle event per animation frame, all carrying the same text. Playback reports every one of those frames, so a single OP lyric could be recorded hundreds of times and dominate "Top Repeated Words".
+
+Recording now collapses those runs as they happen, matching what the subtitle sidebar shows:
+
+- 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.
+
+For stats recorded before this, the Vocabulary tab toolbar has a **Duplicates** button:
+
+- Pick how far back to look (7 days, 30 days, 90 days, 1 year, or all time). A narrower window does less work and keeps older history untouched.
+- **Scan** reports the bursts found, the lines they added, and the word and kanji counts they inflated, without writing anything.
+- **Clean Up** applies exactly what the scan reported: each run collapses to its first line (extended to cover the run), and the removed lines' word and kanji occurrences are subtracted from the vocabulary aggregates.
+
+The same thing runs from the terminal:
+
+```bash
+subminer stats cleanup --duplicate-lines --dry-run --lookback-days 30
+subminer stats cleanup --duplicate-lines --lookback-days 30
+```
+
+`--duplicate-lines` (short: `-d`) picks the cleanup mode, so it cannot be combined with `--vocab` or `--lifetime`, and `--dry-run` and `--lookback-days <days>` only apply to it. Omitting `--lookback-days` scans all history; the value must be at least one day.
+
+The cleanup chains runs per line of text, so interleaved dual-line karaoke collapses each of its lines. It also removes the short residue the live rule stores before a run is long enough to recognize: a run one frame short of the usual minimum qualifies when every event is under the strict 0.1s bound.
+
+Runs never cross a session boundary, so rewatching an episode keeps both watches. Session telemetry (watch time, lines seen, tokens seen) and the rollups derived from it are left as recorded: they are cumulative samples taken during playback, and cannot be recomputed for sessions whose raw rows have since been pruned.
+
 ## Retention Defaults

 By default, SubMiner keeps all retention tables and raw data (`0` means keep all) while continuing daily/monthly rollup maintenance:
@@ -151,6 +151,7 @@ subminer stats -b                       # start background stats daemon
 | `subminer stats`                           | Start the stats server (opens the dashboard when `stats.autoOpenBrowser` is on)                   |
 | `subminer stats -b` / `-s`                 | Start/reuse or stop the background stats daemon                                                   |
 | `subminer stats cleanup`                   | Backfill vocabulary metadata and prune stale rows (`-v` vocab, `-l` lifetime summaries)           |
+| `subminer stats cleanup -d`                | Collapse repeated lines from typeset subs (`--dry-run`, `--lookback-days <n>`)                    |
 | `subminer stats rebuild` / `backfill`      | Rebuild or backfill rollup data                                                                   |
 | `subminer doctor`                          | Dependency + config + socket diagnostics (`--refresh-known-words` refreshes the known-word cache) |
 | `subminer settings`                        | Open the SubMiner settings window                                                                 |
@@ -171,7 +171,9 @@ Without FFmpeg, card creation still works but audio and image fields will be emp

 **Audio or screenshot generation hangs**

-Media generation has a 30-second timeout (60 seconds for animated AVIF). If your video file is on a slow network mount or the codec requires software decoding, generation may time out. Try:
+Audio extraction has a 2-minute timeout. SubMiner also limits FFmpeg probing when mpv provides the selected audio stream, which avoids scanning unrelated subtitle and font-attachment streams in large MKV files. Screenshots retain a 30-second timeout, and animated AVIF uses 60 seconds.
+
+If your video file is on a slow or unresponsive network mount, generation may still time out. Try:

 - Using a local copy of the video file.
 - Reducing `ankiConnect.media.imageQuality` or switching from `avif` to `static` image type.
@@ -405,8 +407,9 @@ On any Wayland session that is not Hyprland or Sway (KDE Plasma, GNOME, and othe

 SubMiner handles this automatically:

- It launches its own window under XWayland (it sets `--ozone-platform-hint=x11`).
- Every mpv it launches (via the `subminer` launcher, Jellyfin, or YouTube) is pinned to XWayland too - Wayland environment hints are stripped and an X11 GPU context (`--gpu-context=x11egl,x11`) is applied.
+- It launches its own window under XWayland (it sets `--ozone-platform=x11`).
+- Every mpv it launches (via the `subminer` launcher, Jellyfin, or YouTube) is pinned to XWayland too - Wayland environment hints are stripped and an X11 GPU context (`--gpu-context=x11vk,x11egl,x11`) is applied. Only the window context is overridden; your `vo`/`gpu-api` and user shaders are left alone.
+- Fractional and mixed-monitor display scaling is handled per screen when SubMiner maps XWayland mpv coordinates to the overlay.
 - While mpv is windowed, the overlay is a managed X11 window owned by the tracked mpv window (`WM_TRANSIENT_FOR`), so it stays above mpv while other foreground X11/Xwayland apps can still cover both windows.
 - While tracked mpv is fullscreen, SubMiner swaps the visible overlay to a focusable-false X11 override-redirect window. That path can stay above the active fullscreen mpv window without requiring a KDE/KWin-specific rule, and SubMiner hides/releases it when mpv is no longer the active X11/Xwayland window.
 - The visible overlay is shown inactive on Linux, so normal hover should not steal keyboard focus from mpv.
@@ -420,7 +423,7 @@ Requirements: `xdotool`, `xprop`, and `xwininfo` must be installed. SubMiner use
 This almost always means mpv came up as a **native Wayland** window that the XWayland overlay cannot cover. It happens when mpv is launched **manually** (your own command), because SubMiner can only force XWayland on the mpv processes it launches itself. Fix it one of these ways:

 - Launch playback through SubMiner (the `subminer` launcher or the tray), which forces XWayland for you, or
- Force XWayland in your own mpv invocation, e.g. `mpv --gpu-context=x11egl …`, or launch with `WAYLAND_DISPLAY= mpv …`, or set `gpu-context=x11egl` in your `mpv.conf`.
+- Force XWayland in your own mpv invocation, e.g. `mpv --gpu-context=x11vk,x11egl,x11 …`, or launch with `WAYLAND_DISPLAY= mpv …`, or set `gpu-context=x11vk` (Vulkan) / `gpu-context=x11egl` (OpenGL) in your `mpv.conf`.

 To confirm mpv is on XWayland, `xdotool search --class mpv` should return a window id (a native Wayland mpv returns nothing).

@@ -95,6 +95,8 @@ subminer texthooker               # Texthooker-only mode (-o also opens the brow
 subminer stats -b                 # Start/reuse the background stats daemon
 subminer stats -s                 # Stop the background stats daemon
 subminer stats cleanup            # Backfill vocabulary metadata, prune stale rows
+subminer stats cleanup -d --dry-run  # Preview cleanup of repeated typeset subtitle lines
+subminer stats cleanup -d --lookback-days 30  # Clean only lines recorded in the last 30 days
 subminer stats rebuild            # Rebuild rollup data
 subminer doctor --refresh-known-words  # Refresh the known-word cache
 subminer logs -e                  # Export a sanitized log ZIP and print its path
@@ -107,6 +109,8 @@ subminer app --stop               # Stop the background app
 subminer --version                # Print the launcher's version
 ```

+`stats cleanup` runs one mode per invocation: `-v`/`--vocab` (the default), `-l`/`--lifetime`, or `-d`/`--duplicate-lines`; explicitly selected modes cannot be combined. `--dry-run` and `--lookback-days <days>` apply to `--duplicate-lines` only and are rejected without it; `--lookback-days` must be at least one day, and leaving it off scans all history.
+
 Jellyfin, cross-machine sync, and character-dictionary commands have their own sections: [Jellyfin](/jellyfin-integration), [Sync Between Machines](/launcher-script#sync-between-machines), and [Character Dictionary](/character-dictionary).

 </details>
@@ -137,7 +141,7 @@ SubMiner.AppImage --start --log-level debug  # Verbose logging without dev mode
 SubMiner.AppImage --help                  # Show all options
 ```

-The remaining flags are internal or scripting-only surfaces: the `--jellyfin-*` family (login, library listing, item playback, cast announce), `--sync-cli` (the app's headless sync entrypoint that `subminer sync` proxies to), `--dictionary-candidates` / `--dictionary-select`, and `--playback-feedback <text>`. Run `SubMiner.AppImage --help` for the complete list. The previous `--open-animetosho` flag is still accepted as a deprecated alias for `--open-tsukihime`.
+The remaining flags are internal or scripting-only surfaces: the `--jellyfin-*` family (login, library listing, item playback, cast announce), `--sync-cli` (the app's headless sync entrypoint that `subminer sync` proxies to), the `--stats-cleanup-*` family that `subminer stats cleanup` forwards (`--stats-cleanup-vocab`, `--stats-cleanup-lifetime`, `--stats-cleanup-duplicate-lines`, and its `--stats-cleanup-dry-run` / `--stats-cleanup-lookback-days <days>` modifiers), `--dictionary-candidates` / `--dictionary-select`, and `--playback-feedback <text>`. Run `SubMiner.AppImage --help` for the complete list. The previous `--open-animetosho` flag is still accepted as a deprecated alias for `--open-tsukihime`.

 </details>

@@ -145,6 +149,8 @@ The tray menu includes `Export Logs`, which creates the same sanitized local-dat

 Once Jellyfin is configured, the tray menu includes `Jellyfin Discovery` for starting or stopping cast discovery in the current app session without changing config.

+The tray menu also includes `View Changelog`, which opens the in-app changelog modal. It fetches the changelog from the newest published release, so you see release notes for versions newer than the one you run; if the download fails it falls back to the changelog bundled with your install and says so. Versions in the current `0.x` line are expanded by default and older lines are folded, matching this site's [Changelog](/changelog). A badge marks the version you have installed, and newer versions are tagged `New`. The same modal opens from the `What's New` button on the update-available overlay notification.
+
 ### Logging and App Mode

 - `--log-level` controls logger verbosity.
@@ -368,6 +374,8 @@ Press `V` to cycle the primary SubMiner subtitle bar through hidden → visible

 `Ctrl/Cmd+/` opens the session help modal with the current overlay and mpv keybindings. The same help view is also available through the `y-h` chord in mpv.

+The changelog modal (tray > `View Changelog`) works the same way: it renders over mpv when a video is playing and in its own window otherwise. Use `J`/`K` or the arrow keys to move between versions, `Enter` to fold or unfold one, `R` to refetch, and `Esc` to close.
+
 Hovering over subtitle text pauses mpv by default; leaving resumes it. Yomitan popups also pause playback by default. Set `subtitleStyle.autoPauseVideoOnHover: false` or `subtitleStyle.autoPauseVideoOnYomitanPopup: false` to disable either behavior.

 ### Drag-and-Drop
@@ -64,18 +64,23 @@ External subtitle files only (SRT, VTT, ASS). Embedded subtitle tracks are out o
 A cue parser extracts both timing and text content from subtitle files for prefetching.

 **Parsed cue structure:**
+
 ```typescript
 interface SubtitleCue {
-  startTime: number;  // seconds
-  endTime: number;    // seconds
-  text: string;       // raw subtitle text
+  startTime: number; // seconds
+  endTime: number; // seconds
+  text: string; // plain text, decoded from the source format
 }
 ```

 **Supported formats:**
+
 - SRT/VTT: Regex-based parsing of timing lines + text content between timing blocks.
- ASS: Parse `[Events]` section, extract `Dialogue:` lines, split on the first 9 commas only (ASS v4+ has 10 fields; the last field is Text which can itself contain commas). Strip ASS override tags (`{\...}`) from the text before storing.
-  ASS text fields contain inline override tags like `{\b1}`, `{\an8}`, `{\fad(200,300)}`. The cue parser strips these during extraction so the tokenizer receives clean text.
+- 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 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`.

 #### Prefetch Service Lifecycle

@@ -153,6 +158,7 @@ tokens (already have frequencyRank values from parser-level applyFrequencyRanks)
 ### Dependency Analysis

 All annotations either depend on MeCab POS data or benefit from running after it:
+
 - **Known word marking:** Needs base tokens (surface/headword). No POS dependency, but no reason to run separately.
 - **Frequency filtering:** Uses `pos1Exclusions` and `pos2Exclusions` to clear frequency ranks on excluded tokens (particles, noise). Depends on MeCab POS data.
 - **JLPT marking:** Uses `shouldIgnoreJlptForMecabPos1` to filter. Depends on MeCab POS data.
@@ -169,18 +175,14 @@ function annotateTokens(tokens, deps, options): MergedToken[] {

  // Single pass: known word + frequency filtering + JLPT computed together
  const annotated = tokens.map((token) => {
-    const isKnown = nPlusOneEnabled
-      ? token.isKnown || computeIsKnown(token, deps)
-      : false;
+    const isKnown = nPlusOneEnabled ? token.isKnown || computeIsKnown(token, deps) : false;

    // Filter frequency rank using POS exclusions (rank values already set at parser level)
    const frequencyRank = frequencyEnabled
      ? filterFrequencyRank(token, pos1Exclusions, pos2Exclusions)
      : undefined;

-    const jlptLevel = jlptEnabled
-      ? computeJlptLevel(token, deps.getJlptLevel)
-      : undefined;
+    const jlptLevel = jlptEnabled ? computeJlptLevel(token, deps.getJlptLevel) : undefined;

    return { ...token, isKnown, frequencyRank, jlptLevel };
  });
@@ -221,6 +223,7 @@ Replace `document.createElement('span')` calls in the renderer with `templateSpa
 ### Current Behavior

 In `renderWithTokens` (`subtitle-render.ts`), each render cycle:
+
 1. Clears DOM with `innerHTML = ''`
 2. Creates a `DocumentFragment`
 3. Calls `document.createElement('span')` for each token (~10-15 per subtitle)
@@ -256,27 +259,30 @@ Full recycling (collecting old nodes, clearing attributes, reusing them) require

 ## Combined Impact Summary

-| Scenario | Before | After | Improvement |
-|----------|--------|-------|-------------|
-| Normal playback (prefetch-warmed) | ~200-320ms | ~30-50ms | ~80-85% |
-| Cache hit (repeated subtitle) | ~72ms | ~55-65ms | ~10-20% |
-| Cache miss (immediate seek) | ~200-320ms | ~150-260ms | ~20-25% |
+| Scenario                          | Before     | After      | Improvement |
+| --------------------------------- | ---------- | ---------- | ----------- |
+| Normal playback (prefetch-warmed) | ~200-320ms | ~30-50ms   | ~80-85%     |
+| Cache hit (repeated subtitle)     | ~72ms      | ~55-65ms   | ~10-20%     |
+| Cache miss (immediate seek)       | ~200-320ms | ~150-260ms | ~20-25%     |

 ---

 ## Files Summary

 ### New Files
+
 - `src/core/services/subtitle-prefetch.ts`
 - `src/core/services/subtitle-cue-parser.ts`

 ### Modified Files
+
 - `src/core/services/subtitle-processing-controller.ts` (expose `preCacheTokenization`)
 - `src/core/services/tokenizer/annotation-stage.ts` (batched single-pass)
 - `src/renderer/subtitle-render.ts` (template cloneNode)
 - `src/main.ts` (wire up prefetch service)

 ### Test Files
+
 - New tests for subtitle cue parser (SRT, VTT, ASS formats)
 - New tests for subtitle prefetch service (priority window, seek, pause/resume)
 - Updated tests for annotation stage (same behavior, new implementation)
@@ -25,6 +25,7 @@ Read when: you need to find the owner module for a behavior or test surface
 - Anki workflow: `src/anki-integration/`, `src/core/services/anki-jimaku*.ts`
 - Immersion tracking: `src/core/services/immersion-tracker/`
  Includes stats storage/query schema such as `imm_videos`, `imm_media_art`, and `imm_youtube_videos` for per-video and YouTube-specific library metadata.
+  `delete-maintenance-scheduler.ts` coalesces and serializes stats deletes; expensive deletion and summary rebuilds run in `delete-maintenance-worker-thread.ts` while the tracker queues playback writes. Each batch uses one transaction, lexical update, rollup refresh, and lifetime rebuild.
 - AniList tracking + character dictionary: `src/core/services/anilist/`, `src/main/runtime/composers/anilist-*`, `src/main/character-dictionary-runtime.ts`, `src/main/character-dictionary-runtime/`
 - Jellyfin integration: `src/core/services/jellyfin*.ts`, `src/main/runtime/composers/jellyfin-*`
 - Window trackers: `src/window-trackers/`
@@ -50,10 +50,22 @@ subtitles do not draw.
 7. Cache miss: call `refreshCurrentSubtitle(text)`. Normal processing emits a plain payload
   synchronously, then replaces it with the tokenized payload when ready.

-In `src/main.ts`, both `onSubtitleChange` and `refreshCurrentSubtitle` pause
-`subtitlePrefetchService`, notify it with `onSeek(lastObservedTimePos)`, and then call the matching
-`subtitleProcessingController` method. This gives the visible overlay priority over background
-prefetch work and re-centers prefetch around the live playback time.
+Both `onSubtitleChange` and `refreshCurrentSubtitle` pause `subtitlePrefetchService` and then call
+the matching `subtitleProcessingController` method, giving the visible overlay priority over
+background prefetch work. Prefetch is not re-centered here: restarting the run per line
+(`onSeek`) discarded the in-flight tokenization every time the subtitle changed, so only real
+seeks restart it (see `onTimePosUpdate` in `src/main.ts`).
+
+On an uncached autoplay prime the raw payload is emitted here and reported to the controller with
+`notePlainSubtitleEmitted`, so the controller skips its own plain emit for that line and the
+overlay receives one plain payload followed by the annotated one.
+
+The pause is released by the controller's `onProcessingSettled` callback, which fires once it has
+no work left. Emits do not release it: the first emit for an uncached line is the plain payload
+that precedes tokenization, and a run can finish without emitting at all (a suppressed duplicate,
+a failed tokenization). Both controller methods return whether processing is now pending, and the
+caller resumes immediately when it is not — a repeated subtitle schedules no work, so no settle is
+coming and prefetching would otherwise idle for the rest of the cue.

 ## Live Cue Delivery

@@ -3,7 +3,7 @@
 # Documentation Catalog

 Status: active  
-Last verified: 2026-05-23  
+Last verified: 2026-08-13
 Owner: Kyle Yasuda  
 Read when: finding internal docs or checking verification status

@@ -17,10 +17,10 @@ Read when: finding internal docs or checking verification status
 | KB rules | `docs/knowledge-base/README.md` | active | 2026-05-23 | maintenance policy |
 | Core beliefs | `docs/knowledge-base/core-beliefs.md` | active | 2026-03-13 | agent-first principles |
 | Quality scorecard | `docs/knowledge-base/quality.md` | active | 2026-03-13 | quality grades and gaps |
-| Workflow index | `docs/workflow/README.md` | active | 2026-05-23 | execution map |
+| 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 plugins | `docs/workflow/agent-plugins.md` | active | 2026-05-23 | repo-local agent workflow plugin ownership |
-| Verification guide | `docs/workflow/verification.md` | active | 2026-05-23 | maintained verification lanes |
+| Agent skills | `docs/workflow/agent-skills.md` | active | 2026-08-13 | 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 |

 ## Update Rules
@@ -1,184 +0,0 @@
-# Library Summary Replaces Per-Day Trends — Design
-
-**Status:** Draft
-**Date:** 2026-04-09
-**Scope:** `stats/` frontend, `src/core/services/immersion-tracker/query-trends.ts` backend
-
-## Problem
-
-The "Library — Per Day" section on the stats Trends tab (`stats/src/components/trends/TrendsTab.tsx:224-254`) renders six stacked-area charts — Videos, Watch Time, Cards, Words, Lookups, and Lookups/100w, each broken down per title per day.
-
-In practice these charts are not useful:
-
- Most titles only have activity on one or two days in a window, so they render as isolated bumps on a noisy baseline.
- Stacking 7+ titles with mostly-zero days makes individual lines hard to follow.
- The top "Activity" and "Period Trends" sections already answer "what am I doing per day" globally.
- The "Library — Cumulative" section directly below already answers "which titles am I progressing through" with less noise.
-
-The per-day section occupies significant vertical space without carrying its weight, and the user has confirmed it should be replaced.
-
-## Goal
-
-Replace the six per-day stacked charts with a single "Library — Summary" section that surfaces per-title aggregate statistics over the selected date range. The new view should make it trivially easy to answer: "For the selected window, which titles am I spending time on, how much mining output have they produced, and how efficient is my lookup rate on each?"
-
-## Non-goals
-
- Changing the "Library — Cumulative" section (stays as-is).
- Changing the "Activity", "Period Trends", or "Patterns" sections.
- Adding a new API endpoint — the existing dashboard endpoint is extended in place.
- Renaming internal `anime*` data-model identifiers (`animeId`, `imm_anime`, etc.). Those stay per the convention established in `c5e778d7`; only new fields/types/user-visible strings use generic "title"/"library" wording.
- Supporting a true all-time library view on the Trends tab. If that's ever wanted, it belongs on a different tab.
-
-## Solution Overview
-
-Delete the "Library — Per Day" section. In its place, add "Library — Summary", composed of:
-
-1. A horizontal-bar leaderboard chart of watch time per title (top 10, descending).
-2. A sortable table of every title with activity in the selected window, with columns: Title, Watch Time, Videos, Sessions, Cards, Words, Lookups, Lookups/100w, Date Range.
-
-Both controls are scoped to the top-of-page date range selector. The existing shared Anime Visibility filter continues to work — it now gates Summary + Cumulative instead of Per-Day + Cumulative.
-
-## Backend
-
-### New type
-
-Add to `stats/src/types/stats.ts` and the backend query module:
-
-```ts
-type LibrarySummaryRow = {
-  title: string;              // display title — anime series, YouTube video title, etc.
-  watchTimeMin: number;       // sum(total_active_min) across the window
-  videos: number;             // distinct video_id count
-  sessions: number;           // session count from imm_sessions
-  cards: number;              // sum(total_cards)
-  words: number;              // sum(total_tokens_seen)
-  lookups: number;            // sum(lookup_count) from imm_sessions
-  lookupsPerHundred: number | null; // lookups / words * 100, null when words == 0
-  firstWatched: number;       // min(rollup_day) as epoch day, within the window
-  lastWatched: number;        // max(rollup_day) as epoch day, within the window
-};
-```
-
-### Query changes in `src/core/services/immersion-tracker/query-trends.ts`
-
- Add `librarySummary: LibrarySummaryRow[]` to `TrendsDashboardQueryResult`.
- Populate it from a single aggregating query over `imm_daily_rollups` joined to `imm_videos` → `imm_anime`, filtered by `rollup_day` within the selected window. Session count and lookup count come from `imm_sessions` aggregated by `video_id` and then grouped by the parent library entry. Use a single query (or at most two joined/unioned) — no N+1.
- `imm_anime` is the generic library-grouping table; anime series, YouTube videos, and yt-dlp imports all land there. The internal table name stays `imm_anime`; only the new field uses generic naming.
- Return rows pre-sorted by `watchTimeMin` descending so the leaderboard is zero-cost and the table default sort matches.
- Emit `lookupsPerHundred: null` when `words == 0`.
-
-### Removed from API response
-
-Drop the entire `animePerDay` field from `TrendsDashboardQueryResult` (both backend in `src/core/services/immersion-tracker/query-trends.ts` and frontend in `stats/src/types/stats.ts`).
-
-Internally, the existing helpers (`buildPerAnimeFromDailyRollups`, `buildEpisodesPerAnimeFromDailyRollups`) are still used as intermediates to build `animeCumulative.*` via `buildCumulativePerAnime`. Keep those helpers — just scope their output to local variables inside `getTrendsDashboard` instead of exposing them on the response. The `buildPerAnimeFromSessions` call for lookups and the `buildLookupsPerHundredPerAnime` helper become unused and can be deleted.
-
-Before removing `animePerDay` from the frontend type, verify no other file under `stats/src/` references it. Based on current inspection, only `TrendsTab.tsx` and `stats/src/types/stats.ts` touch it.
-
-## Frontend
-
-### New component: `stats/src/components/trends/LibrarySummarySection.tsx`
-
-Owns the header, leaderboard chart, visibility-filtered data, and the table. Keeps `TrendsTab.tsx` from growing. Component props: `{ rows: LibrarySummaryRow[]; hiddenTitles: ReadonlySet<string>; windowStart: Date; windowEnd: Date }`.
-
-Internal state: `useState<{ column: ColumnId; direction: 'asc' | 'desc' }>` for sort, defaulting to `{ column: 'watchTimeMin', direction: 'desc' }`.
-
-### Layout
-
-Replaces `TrendsTab.tsx:224-254`:
-
-```
-[SectionHeader: "Library — Summary"]
-[AnimeVisibilityFilter — unchanged, shared with Cumulative below]
-[Card, col-span-full: Leaderboard — horizontal bar chart, ~260px tall]
-[Card, col-span-full: Sortable table, auto height up to ~480px with internal scroll]
-```
-
-Both cards use the existing chart/card wrapper styling.
-
-### Leaderboard chart
-
- Recharts horizontal bar chart (matches the rest of the page — existing charts use `recharts`, not ECharts).
- Top 10 titles by watch time. If fewer titles have activity, render what's there.
- Y-axis: title (category), truncated with ellipsis at container width; full title visible in the Recharts tooltip.
- X-axis: minutes (number).
- Use `layout="vertical"` with `YAxis dataKey="title" type="category"` and `XAxis type="number"`.
- Single series color: `#8aadf4` (matching the existing Watch Time color).
- Reuse `CHART_DEFAULTS`, `CHART_THEME`, `TOOLTIP_CONTENT_STYLE` from `stats/src/lib/chart-theme.ts` so theming matches the rest of the dashboard.
- Chart order is fixed at watch-time desc regardless of table sort — the leaderboard's meaning is fixed.
-
-### Table
-
- Plain HTML `<table>` with Tailwind classes. No new deps.
- Columns, in order:
-  1. **Title** — left-aligned, sticky, truncated with ellipsis, full title on hover.
-  2. **Watch Time** — formatted `Xh Ym` when ≥60 min, else `Xm`.
-  3. **Videos** — integer.
-  4. **Sessions** — integer.
-  5. **Cards** — integer.
-  6. **Words** — integer.
-  7. **Lookups** — integer.
-  8. **Lookups/100w** — one decimal place, `—` when null.
-  9. **Date Range** — `Mon D → Mon D` using the title's `firstWatched` / `lastWatched` within the window.
- Click a column header to sort; click again to reverse. Visual arrow on the active column.
- Numeric columns right-aligned.
- Null `lookupsPerHundred` sorts as the lowest value in both directions (consistent with "no data").
- Row hover highlight; no row click action (read-only view).
- Empty state: "No library activity in the selected window."
-
-### Visibility filter integration
-
-Hiding a title via `AnimeVisibilityFilter` removes it from both the leaderboard and the table. The filter's set of available titles is built from the union of titles that appear in `librarySummary` and the existing `animeCumulative.*` arrays (matches current behavior in `buildAnimeVisibilityOptions`).
-
-### `TrendsTab.tsx` changes
-
- Remove the `filteredEpisodesPerAnime`, `filteredWatchTimePerAnime`, `filteredCardsPerAnime`, `filteredWordsPerAnime`, `filteredLookupsPerAnime`, `filteredLookupsPerHundredPerAnime` locals.
- Remove the six `<StackedTrendChart>` calls in the "Library — Per Day" section.
- Remove the `<SectionHeader>Library — Per Day</SectionHeader>` and the `<AnimeVisibilityFilter>` from that position.
- Insert `<SectionHeader>Library — Summary</SectionHeader>` + `<AnimeVisibilityFilter>` + `<LibrarySummarySection>` in the same place.
- Update `buildAnimeVisibilityOptions` input to use `librarySummary` titles instead of the six dropped `animePerDay.*` arrays.
-
-## Data flow
-
-1. `useTrends(range, groupBy)` calls `/api/stats/trends/dashboard`.
-2. Response now includes `librarySummary` (sorted by watch time desc).
-3. `TrendsTab` holds the shared `hiddenAnime` set (unchanged).
-4. `LibrarySummarySection` receives `librarySummary` + `hiddenAnime`, filters out hidden rows, renders the leaderboard from the top-10 slice of the filtered list, renders the table from the filtered list with local sort state applied.
-5. Date-range selector changes trigger a new fetch; `groupBy` toggle does not affect the summary section (it's always window-total).
-
-## Edge cases
-
- **No activity in window:** Section renders header + empty-state card. Leaderboard card hidden. Visibility filter hidden.
- **One title only:** Leaderboard renders a single bar; table renders one row. No special-casing.
- **Title with zero words but non-zero lookups:** `lookupsPerHundred` is `null`, rendered as `—`. Sort treats null as lowest.
- **Title with zero cards/lookups/words but non-zero watch time:** Normal zero rendering, still shown.
- **Very long titles:** Ellipsis in chart y-axis labels and table title column; full title in `title` attribute / ECharts tooltip.
- **Mixed sources (anime + YouTube):** No special case — both land in `imm_anime` and are grouped uniformly.
-
-## Testing
-
-### Backend (`query-trends.ts`)
-
-New unit tests, following the existing pattern:
-
-1. Empty window returns `librarySummary: []`.
-2. Single title with a few rollups: all aggregates are correct; `firstWatched`/`lastWatched` match the bounding days within the window.
-3. Multiple titles: rows returned sorted by watch time desc.
-4. Mixed sources (anime-style + YouTube-style entries in `imm_anime`): both appear in the summary with their own aggregates.
-5. Title with `words == 0`: `lookupsPerHundred` is `null`.
-6. Date range excludes some rollups: excluded rollups are not counted; `firstWatched`/`lastWatched` reflect only within-window activity.
-7. `sessions` and `lookups` come from `imm_sessions`, not `imm_daily_rollups`, and are correctly attributed to the parent library entry.
-
-### Frontend
-
- Existing Trends tab smoke test should continue to pass after wiring.
- Optional: a targeted render test for `LibrarySummarySection` (empty state, single title, sort toggle, visibility filter interaction). Not required for merge if the smoke test exercises the happy path.
-
-## Release / docs
-
- One fragment in `changes/*.md` summarizing the replacement.
- No user-facing docs (`docs-site/`) changes unless the per-day section was documented there — verify during implementation.
-
-## Open items
-
-None.
@@ -1,347 +0,0 @@
-# Stats Dashboard Feedback Pass — Design
-
-Date: 2026-04-09
-Scope: Stats dashboard UX follow-ups from user feedback (items 1–7).
-Delivery: **Single PR**, broken into logically scoped commits.
-
-## Goals
-
-Address seven concrete pieces of feedback against the Statistics menu:
-
-1. Library — collapse episodes behind a per-series dropdown.
-2. Sessions — roll up multiple sessions of the same episode within a day.
-3. Trends — add a 365d range option.
-4. Library — delete an episode (video) from its detail view.
-5. Vocabulary — tighten spacing between word and reading in the Top 50 table.
-6. Episode detail — hide cards whose Anki notes have been deleted.
-7. Trend/watch charts — add gridlines, fix tick legibility, unify theming.
-
-Out of scope for this pass: English-token ingestion cleanup and Overview stat-card drill-downs (feedback items 8 and 9). Those require a larger design decision and a migration respectively.
-
-## Files touched (inventory)
-
-Dashboard (`stats/src/`):
- `components/library/LibraryTab.tsx` — collapsible groups (item 1).
- `components/library/MediaDetailView.tsx`, `components/library/MediaHeader.tsx` — delete-episode action (item 4).
- `components/sessions/SessionsTab.tsx`, `components/library/MediaSessionList.tsx` — episode rollup (item 2).
- `components/trends/DateRangeSelector.tsx`, `hooks/useTrends.ts`, `lib/api-client.ts`, `lib/api-client.test.ts` — 365d (item 3).
- `components/vocabulary/FrequencyRankTable.tsx` — word/reading column collapse (item 5).
- `components/anime/EpisodeDetail.tsx` — filter deleted Anki cards (item 6).
- `components/trends/TrendChart.tsx`, `components/trends/StackedTrendChart.tsx`, `components/overview/WatchTimeChart.tsx`, `lib/chart-theme.ts` — chart clarity (item 7).
- New file: `stats/src/lib/session-grouping.ts` + `session-grouping.test.ts`.
-
-Backend (`src/core/services/`):
- `immersion-tracker/query-trends.ts` — extend `TrendRange` and `TREND_DAY_LIMITS` (item 3).
- `immersion-tracker/__tests__/query.test.ts` — 365d coverage (item 3).
- `stats-server.ts` — passthrough if range validation lives here (check before editing).
- `__tests__/stats-server.test.ts` — 365d coverage (item 3).
-
-## Commit plan
-
-One PR, one feature per commit. Order picks low-risk mechanical changes first so failures in later commits don't block merging of earlier ones.
-
-1. `feat(stats): add 365d range to trends dashboard` (item 3)
-2. `fix(stats): tighten word/reading column in Top 50 table` (item 5)
-3. `fix(stats): hide cards deleted from Anki in episode detail` (item 6)
-4. `feat(stats): delete episode from library detail view` (item 4)
-5. `feat(stats): collapsible series groups in library` (item 1)
-6. `feat(stats): roll up same-episode sessions within a day` (item 2)
-7. `feat(stats): gridlines and unified theme for trend charts` (item 7)
-
-Each commit must pass `bun run typecheck`, `bun run test:fast`, and any change-specific checks listed below.
-
---
-
-## Item 1 — Library collapsible series groups
-
-### Current behavior
-
-`LibraryTab.tsx` groups media via `groupMediaLibraryItems` and always renders the full grid of `MediaCard`s beneath each group header.
-
-### Target behavior
-
-Each group header becomes clickable. Groups with `items.length > 1` default to **collapsed**; single-video groups stay expanded (collapsing them would be visual noise).
-
-### Implementation
-
- State: `const [collapsedGroups, setCollapsedGroups] = useState<Set<string>>(...)`. Initialize from `grouped` where `items.length > 1`.
- Toggle helper: `toggleGroup(key: string)` adds/removes from the set.
- Group header: wrap in a `<button>` with `aria-expanded` and a chevron icon (`▶`/`▼`). Keep the existing cover + title + subtitle layout inside the button.
- Children grid is conditionally rendered on `!collapsedGroups.has(group.key)`.
- Header summary (`N videos · duration · cards`) stays visible in both states so collapsed groups remain informative.
-
-### Tests
-
- New `LibraryTab.test.tsx` (if not already present — check first) covering:
-  - Multi-video group renders collapsed on first mount.
-  - Single-video group renders expanded on first mount.
-  - Clicking the header toggles visibility.
-  - Header summary is visible in both states.
-
---
-
-## Item 2 — Sessions episode rollup within a day
-
-### Current behavior
-
-`SessionsTab.tsx:10-24` groups sessions by day label only (`formatSessionDayLabel(startedAtMs)`). Multiple sessions of the same episode on the same day show as independent rows. `MediaSessionList.tsx` has the same problem inside the library detail view.
-
-### Target behavior
-
-Within each day, sessions with the same `videoId` collapse into one parent row showing combined totals. A chevron reveals the individual sessions. Single-session buckets render flat (no pointless nesting).
-
-### Implementation
-
- New helper in `stats/src/lib/session-grouping.ts`:
-  ```ts
-  export interface SessionBucket {
-    key: string;              // videoId as string, or `s-${sessionId}` for singletons
-    videoId: number | null;
-    sessions: SessionSummary[];
-    totalActiveMs: number;
-    totalCardsMined: number;
-    representativeSession: SessionSummary; // most recent, for header display
-  }
-  export function groupSessionsByVideo(sessions: SessionSummary[]): SessionBucket[];
-  ```
-  Sessions missing a `videoId` become singleton buckets.
-
- `SessionsTab.tsx`: after day grouping, pipe each `daySessions` through `groupSessionsByVideo`. Render each bucket:
-  - `sessions.length === 1`: existing `SessionRow` behavior, unchanged.
-  - `sessions.length >= 2`: render a **bucket row** that looks like `SessionRow` but shows combined totals and session count (e.g. `3 sessions · 1h 24m · 12 cards`). Chevron state stored in a second `Set<string>` on bucket key. Expanded buckets render the child `SessionRow`s indented (`pl-8`) beneath the header.
- `MediaSessionList.tsx`: within the media detail view, a single video's sessions are all the same `videoId` by definition — grouping here is by day only, and within a day multiple sessions render nested under a day header. Re-use the same visual pattern; factor the bucket row into a shared `SessionBucketRow` component.
-
-### Delete semantics
-
- Deleting a bucket header offers "Delete all N sessions in this group" (reuse `confirmDayGroupDelete` pattern with a bucket-specific message, or add `confirmBucketDelete`).
- Deleting an individual session from inside an expanded bucket keeps the existing single-delete flow.
-
-### Tests
-
- `session-grouping.test.ts`:
-  - Empty input → empty output.
-  - All unique videos → N singleton buckets.
-  - Two sessions same videoId → one bucket with correct totals and representative (most recent start time).
-  - Missing videoId → singleton bucket keyed by sessionId.
- `SessionsTab.test.tsx` (extend or add) verifying the rendered bucket rows expand/collapse and delete hooks fire with the right ID set.
-
---
-
-## Item 3 — 365d trends range
-
-### Backend
-
-`src/core/services/immersion-tracker/query-trends.ts`:
- `type TrendRange = '7d' | '30d' | '90d' | '365d' | 'all';`
- Add `'365d': 365` to `TREND_DAY_LIMITS`.
- `getTrendDayLimit` picks up the new key automatically because of the `Exclude<TrendRange, 'all'>` generic.
-
-`src/core/services/stats-server.ts`:
- Search for any hardcoded range validation (e.g. allow-list in the trends route handler) and extend it.
-
-### Frontend
-
- `hooks/useTrends.ts`: widen the `TimeRange` union.
- `components/trends/DateRangeSelector.tsx`: add `'365d'` to the options list. Display label stays as `365d`.
- `lib/api-client.ts` / `api-client.test.ts`: if the client validates ranges, add `365d`.
-
-### Tests
-
- `query.test.ts`: extend the existing range table to cover `365d` returning 365 days of data.
- `stats-server.test.ts`: ensure the route accepts `range=365d`.
- `api-client.test.ts`: ensure the client emits the new range.
-
-### Change-specific checks
-
- `bun run test:config` is not required here (no schema/defaults change).
- Run `bun run typecheck` + `bun run test:fast`.
-
---
-
-## Item 4 — Delete episode from library detail
-
-### Current behavior
-
-`MediaDetailView.tsx` provides session-level delete only. The backend `deleteVideo` exists (`query-maintenance.ts:509`), the API is exposed at `stats-server.ts:559`, and `api-client.deleteVideo` is already wired (`stats/src/lib/api-client.ts:146`). `EpisodeList.tsx:46` already uses it from the anime tab.
-
-### Target behavior
-
-A "Delete Episode" action in `MediaHeader` (top-right, small, `text-ctp-red`), gated by `confirmEpisodeDelete(title)`. On success, call `onBack()` and make sure the parent `LibraryTab` refetches.
-
-### Implementation
-
- Add an `onDeleteEpisode?: () => void` prop to `MediaHeader` and render the button only if provided.
- In `MediaDetailView`:
-  - New handler `handleDeleteEpisode` that calls `apiClient.deleteVideo(videoId)`, then `onBack()`.
-  - Reuse `confirmEpisodeDelete` from `stats/src/lib/delete-confirm.ts`.
- In `LibraryTab`:
-  - `useMediaLibrary` returns fresh data on mount. The simplest fix: pass a `refresh` function from the hook (extend the hook if it doesn't already expose one) and call it when the detail view signals back.
-  - Alternative: force a remount by incrementing a `libraryVersion` key on the library list. Prefer `refresh` for clarity.
-
-### Tests
-
- Extend the existing `MediaDetailView.test.tsx`: mock `apiClient.deleteVideo`, click the new button, confirm `onBack` fires after success.
- `useMediaLibrary.test.ts`: if we add a `refresh` method, cover it.
-
---
-
-## Item 5 — Vocabulary word/reading column collapse
-
-### Current behavior
-
-`FrequencyRankTable.tsx:110-144` uses a 5-column table: `Rank | Word | Reading | POS | Seen`. Word and Reading are auto-sized, producing a large gap.
-
-### Target behavior
-
-Merge Word + Reading into a single column titled "Word". Reading sits immediately after the headword in a muted, smaller style.
-
-### Implementation
-
- Drop the `<th>Reading</th>` header and cell.
- Word cell becomes:
-  ```tsx
-  <td className="py-1.5 pr-3">
-    <span className="text-ctp-text font-medium">{w.headword}</span>
-    {reading && (
-      <span className="text-ctp-subtext0 text-xs ml-1.5">
-        【{reading}】
-      </span>
-    )}
-  </td>
-  ```
-  where `reading = fullReading(w.headword, w.reading)` and differs from `headword`.
- Keep `fullReading` import from `reading-utils`.
-
-### Tests
-
- Extend `FrequencyRankTable.test.tsx` (if present — otherwise add a focused test) to assert:
-  - Headword renders.
-  - Reading renders when different from headword.
-  - Reading does not render when equal to headword.
-
---
-
-## Item 6 — Hide Anki-deleted cards in Cards Mined
-
-### Current behavior
-
-`EpisodeDetail.tsx:109-147` iterates `cardEvents`, fetches note info via `ankiNotesInfo(allNoteIds)`, and for each `noteId` renders a row even if no matching `info` came back — the user sees an empty word with an "Open in Anki" button that leads nowhere.
-
-### Target behavior
-
-After `ankiNotesInfo` resolves:
- Drop `noteId`s that are not in the resolved map.
- Drop `cardEvents` whose `noteIds` list was non-empty but is now empty after filtering.
- Card events with a positive `cardsDelta` but no `noteIds` (legacy rollup path) still render as `+N cards` — we have no way to cross-reference them, so leave them alone.
-
-### Implementation
-
- Compute `filteredCardEvents` as a `useMemo` depending on `data.cardEvents` and `noteInfos`.
- Iterate `filteredCardEvents` instead of `cardEvents` in the render.
- Surface a subtle note (optional, muted) "N cards hidden (deleted from Anki)" at the end of the list if any were filtered — helps the user understand why counts here diverge from session totals. Final decision on the note can be made at PR review; default: **show it**.
-
-### Tests
-
- Add a test in `EpisodeDetail.test.tsx` (add the file if not present) that stubs `ankiNotesInfo` to return only a subset of notes and verifies the missing ones are not rendered.
-
-### Other call sites
-
- Grep so far shows `ankiNotesInfo` is only used in `EpisodeDetail.tsx`. Re-verify before landing the commit; if another call site appears, apply the same filter.
-
---
-
-## Item 7 — Trend/watch chart clarity pass
-
-### Current behavior
-
-`TrendChart.tsx`, `StackedTrendChart.tsx`, and `WatchTimeChart.tsx` render Recharts components with:
- No `CartesianGrid` → no horizontal reference lines.
- 9px axis ticks → borderline unreadable.
- Height 120 → cramped.
- Tooltip uses raw labels (`04/04` etc.).
- No shared theme object; each chart redefines colors and tooltip styles inline.
-
-`stats/src/lib/chart-theme.ts` already exists and currently exports a single `CHART_THEME` constant with tick/tooltip colors and `barFill`. It will be extended, not replaced, to preserve existing consumers.
-
-### Target behavior
-
-All three charts share a theme, have horizontal gridlines, readable ticks, and sensible tooltips.
-
-### Implementation
-
-Extend `stats/src/lib/chart-theme.ts` with the additional shared defaults (keeping the existing `CHART_THEME` export intact so current consumers don't break):
-```ts
-export const CHART_THEME = {
-  tick: '#a5adcb',
-  tooltipBg: '#363a4f',
-  tooltipBorder: '#494d64',
-  tooltipText: '#cad3f5',
-  tooltipLabel: '#b8c0e0',
-  barFill: '#8aadf4',
-  grid: '#494d64',
-  axisLine: '#494d64',
-} as const;
-
-export const CHART_DEFAULTS = {
-  height: 160,
-  tickFontSize: 11,
-  margin: { top: 8, right: 8, bottom: 0, left: 0 },
-  grid: { strokeDasharray: '3 3', vertical: false },
-} as const;
-
-export const TOOLTIP_CONTENT_STYLE = {
-  background: CHART_THEME.tooltipBg,
-  border: `1px solid ${CHART_THEME.tooltipBorder}`,
-  borderRadius: 6,
-  color: CHART_THEME.tooltipText,
-  fontSize: 12,
-};
-```
-
-Apply to each chart:
- Import `CartesianGrid` from recharts.
- Insert `<CartesianGrid stroke={CHART_THEME.grid} {...CHART_DEFAULTS.grid} />` inside each chart container.
- `<XAxis tick={{ fontSize: CHART_DEFAULTS.tickFontSize, fill: CHART_THEME.tick }} />` and equivalent `YAxis`.
- `YAxis` gains `axisLine={{ stroke: CHART_THEME.axisLine }}`.
- `ResponsiveContainer` height changes from 120 → `CHART_DEFAULTS.height`.
- `Tooltip` `contentStyle` uses `TOOLTIP_CONTENT_STYLE`, and charts pass a `labelFormatter` when the label is a date key (e.g. show `Fri Apr 4`).
-
-### Unit formatters
-
- `TrendChart` already accepts a `formatter` prop — extend usage sites to pass unit-aware formatters where they aren't already (`formatDuration`, `formatNumber`, etc.).
-
-### Tests
-
- `chart-theme.test.ts` (if present — otherwise add a trivial snapshot to keep the shape stable).
- `TrendChart` snapshot/render tests: no regression, gridline element present.
-
---
-
-## Verification gate
-
-Before requesting code review, run:
-
-```
-bun run typecheck
-bun run test:fast
-bun run test:env
-bun run test:runtime:compat   # dist-sensitive check for the charts
-bun run build
-bun run test:smoke:dist
-```
-
-No docs-site changes are planned in this spec; if `docs-site/` ends up touched (e.g. screenshots), also run `bun run docs:test` and `bun run docs:build`.
-
-No config schema changes → `bun run test:config` and `bun run generate:config-example` are not required.
-
-## Risks and open questions
-
- **MediaDetailView refresh**: `useMediaLibrary` may not expose a `refresh` function. If it doesn't, the simplest path is adding one; the alternative (keying a remount) works but is harder to test. Decide during implementation.
- **Session bucket delete UX**: "Delete all N sessions in this group" is powerful. The copy must make it clear the underlying sessions are being removed, not just the grouping. Reuse `confirmBucketDelete` wording from existing confirm helpers if possible.
- **Anki-deleted-cards hidden notice**: Showing a subtle "N cards hidden" footer is a call that can be made at PR review.
- **Bucket delete helper**: `confirmBucketDelete` does not currently exist in `delete-confirm.ts`. Implementation either adds it or reuses `confirmDayGroupDelete` with bucket-specific wording — decide during the session-rollup commit.
-
-## Changelog entry
-
-User-visible PR → needs a fragment under `changes/*.md`. Suggested title:
-`Stats dashboard: collapsible series, session rollups, 365d trends, chart polish, episode delete.`
@@ -3,7 +3,7 @@
 # Workflow

 Status: active  
-Last verified: 2026-05-23  
+Last verified: 2026-08-13
 Owner: Kyle Yasuda  
 Read when: planning or executing nontrivial work in this repo

@@ -13,7 +13,7 @@ This section is the internal workflow map for contributors and agents.

 - [Planning](./planning.md) - when to write a lightweight plan vs a full execution plan
 - [Verification](./verification.md) - maintained test/build lanes and handoff gate
- [Agent Plugins](./agent-plugins.md) - repo-local plugin ownership for agent workflow skills
+- [Agent Skills](./agent-skills.md) - repo-local workflow skill ownership
 - [Release Guide](../RELEASING.md) - tagged release workflow

 ## Default Flow
@@ -1,27 +0,0 @@
-<!-- read_when: using or modifying repo-local agent plugins -->
-
-# Agent Plugins
-
-Status: active
-Last verified: 2026-05-23
-Owner: Kyle Yasuda
-Read when: packaging or migrating repo-local agent workflow skills into plugins
-
-## SubMiner Workflow Plugin
-
- Canonical plugin path: `plugins/subminer-workflow/`
- Marketplace catalog: `.agents/plugins/marketplace.json`
- Canonical skill sources:
-  - `plugins/subminer-workflow/skills/subminer-scrum-master/`
-  - `plugins/subminer-workflow/skills/subminer-change-verification/`
-
-## Migration Rule
-
- Plugin-owned skills are the source of truth.
- `.agents/skills/subminer-*` remain only as compatibility shims.
- Existing script entrypoints under `.agents/skills/subminer-change-verification/scripts/` stay as wrappers so historical commands do not break.
-
-## Verification
-
- For plugin/docs-only changes, start with `bun run test:docs:kb`.
- Use the plugin-owned verifier when the change crosses from docs into scripts or workflow logic.
@@ -0,0 +1,31 @@
+<!-- read_when: using or modifying repo-local agent skills -->
+
+# Agent Skills
+
+Status: active
+Last verified: 2026-08-13
+Owner: Kyle Yasuda
+Read when: using, adding, or changing a repo-local agent workflow skill
+
+## Canonical Skills
+
+- `.agents/skills/subminer-change-verification/`
+  - Selects the cheapest sufficient repo-native verification lane.
+  - Defers command ownership to `package.json` and `docs/workflow/verification.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.
+
+## Rules
+
+- Keep each skill focused on one repeatable repository task.
+- Prefer instructions over helper scripts unless deterministic tooling provides clear value beyond existing package commands.
+- Keep trigger descriptions narrow enough to avoid invoking skills for unrelated requests.
+- Update this page and the documentation catalog when skill ownership changes.
+
+## Verification
+
+For skill or internal workflow documentation changes, run:
+
+```bash
+bun run test:docs:kb
+```
@@ -3,14 +3,14 @@
 # Verification

 Status: active  
-Last verified: 2026-07-06  
+Last verified: 2026-08-13
 Owner: Kyle Yasuda  
 Read when: selecting the right verification lane for a change

 ## Lane Infrastructure

 - Lane membership is defined once in `scripts/test-lanes.ts` and discovered by
-  directory — new test files join their lane automatically; never hand-list test
+  directory, so new test files join their lane automatically; never hand-list test
  files in `package.json`.
 - `scripts/run-test-lane.mjs` runs each test file in its own `bun test` process
  (per-file isolation with a wall timeout) so a hanging test or leaked global in
@@ -43,8 +43,8 @@ bun run docs:build

 ## Cheap-First Lane Selection

- Docs-only boundary/content changes: `bun run docs:test`, `bun run docs:build`
- Internal KB / `AGENTS.md` changes: `bun run test:docs:kb`
+- User-facing `docs-site/` changes: `bun run docs:test`, `bun run docs:build`
+- Internal KB, `AGENTS.md`, or `.agents/skills/**` changes: `bun run test:docs:kb`
 - Config/schema/defaults: `bun run test:config`, then `bun run generate:config-example` if template/defaults changed
 - Launcher/plugin: `bun run test:launcher` or `bun run test:env`
 - Runtime-compat / compiled behavior: `bun run test:runtime:compat`
@@ -157,6 +157,15 @@ export async function runStatsCommand(
    if (args.statsCleanupLifetime) {
      forwarded.push('--stats-cleanup-lifetime');
    }
+    if (args.statsCleanupDuplicateLines) {
+      forwarded.push('--stats-cleanup-duplicate-lines');
+    }
+    if (args.statsCleanupDryRun) {
+      forwarded.push('--stats-cleanup-dry-run');
+    }
+    if (args.statsCleanupLookbackDays) {
+      forwarded.push('--stats-cleanup-lookback-days', String(args.statsCleanupLookbackDays));
+    }
    if (shouldForwardLogLevel(args.logLevel)) {
      forwarded.push('--log-level', args.logLevel);
    }
@@ -134,6 +134,9 @@ test('applyInvocationsToArgs maps config and jellyfin invocation state', () => {
    statsCleanup: false,
    statsCleanupVocab: false,
    statsCleanupLifetime: false,
+    statsCleanupDuplicateLines: false,
+    statsCleanupDryRun: false,
+    statsCleanupLookbackDays: null,
    statsLogLevel: null,
    syncTriggered: false,
    syncCliTokens: [],
@@ -185,6 +188,9 @@ test('applyInvocationsToArgs maps settings invocation to settings window', () =>
    statsCleanup: false,
    statsCleanupVocab: false,
    statsCleanupLifetime: false,
+    statsCleanupDuplicateLines: false,
+    statsCleanupDryRun: false,
+    statsCleanupLookbackDays: null,
    statsLogLevel: null,
    syncTriggered: false,
    syncCliTokens: [],
@@ -229,6 +235,9 @@ test('applyInvocationsToArgs fails when config invocation has no action', () =>
      statsCleanup: false,
      statsCleanupVocab: false,
      statsCleanupLifetime: false,
+      statsCleanupDuplicateLines: false,
+      statsCleanupDryRun: false,
+      statsCleanupLookbackDays: null,
      statsLogLevel: null,
      syncTriggered: false,
      syncCliTokens: [],
@@ -271,6 +280,9 @@ test('applyInvocationsToArgs maps texthooker browser-open request', () => {
    statsCleanup: false,
    statsCleanupVocab: false,
    statsCleanupLifetime: false,
+    statsCleanupDuplicateLines: false,
+    statsCleanupDryRun: false,
+    statsCleanupLookbackDays: null,
    statsLogLevel: null,
    syncTriggered: false,
    syncCliTokens: [],
@@ -162,6 +162,8 @@ export function createDefaultArgs(
    statsCleanup: false,
    statsCleanupVocab: false,
    statsCleanupLifetime: false,
+    statsCleanupDuplicateLines: false,
+    statsCleanupDryRun: false,
    doctor: false,
    doctorRefreshKnownWords: false,
    logsExport: false,
@@ -258,6 +260,11 @@ export function applyInvocationsToArgs(parsed: Args, invocations: CliInvocations
  if (invocations.statsCleanup) parsed.statsCleanup = true;
  if (invocations.statsCleanupVocab) parsed.statsCleanupVocab = true;
  if (invocations.statsCleanupLifetime) parsed.statsCleanupLifetime = true;
+  if (invocations.statsCleanupDuplicateLines) parsed.statsCleanupDuplicateLines = true;
+  if (invocations.statsCleanupDryRun) parsed.statsCleanupDryRun = true;
+  if (invocations.statsCleanupLookbackDays !== null) {
+    parsed.statsCleanupLookbackDays = invocations.statsCleanupLookbackDays;
+  }
  if (invocations.dictionaryTarget) {
    parsed.dictionaryTarget = parseDictionaryTarget(invocations.dictionaryTarget);
  } else if (
@@ -37,6 +37,9 @@ export interface CliInvocations {
  statsCleanup: boolean;
  statsCleanupVocab: boolean;
  statsCleanupLifetime: boolean;
+  statsCleanupDuplicateLines: boolean;
+  statsCleanupDryRun: boolean;
+  statsCleanupLookbackDays: number | null;
  statsLogLevel: string | null;
  syncTriggered: boolean;
  syncCliTokens: string[];
@@ -53,6 +56,16 @@ export interface CliInvocations {
  texthookerOpenBrowser: boolean;
 }

+/** `--lookback-days` narrows the duplicate-line cleanup; fractions are floored. */
+function parseStatsLookbackDays(value: unknown): number | null {
+  if (typeof value !== 'string' && typeof value !== 'number') return null;
+  const days = Number(value);
+  if (!Number.isFinite(days) || days < 1) {
+    throw new Error('Stats --lookback-days must be at least one day.');
+  }
+  return Math.floor(days);
+}
+
 function applyRootOptions(program: Command): void {
  program
    .option(
@@ -169,6 +182,9 @@ export function parseCliPrograms(
  let statsCleanup = false;
  let statsCleanupVocab = false;
  let statsCleanupLifetime = false;
+  let statsCleanupDuplicateLines = false;
+  let statsCleanupDryRun = false;
+  let statsCleanupLookbackDays: number | null = null;
  let statsLogLevel: string | null = null;
  let syncTriggered = false;
  let syncCliTokens: string[] = [];
@@ -269,6 +285,9 @@ export function parseCliPrograms(
    .option('-s, --stop', 'Stop the background stats server')
    .option('-v, --vocab', 'Clean vocabulary rows in the stats database')
    .option('-l, --lifetime', 'Rebuild lifetime summary rows from retained data')
+    .option('-d, --duplicate-lines', 'Collapse repeated subtitle lines from typeset animations')
+    .option('--dry-run', 'Report what a cleanup would remove without changing anything')
+    .option('--lookback-days <days>', 'Only clean lines recorded in the last N days')
    .option('--log-level <level>', 'Log level')
    .action((action: string | undefined, options: Record<string, unknown>) => {
      statsTriggered = true;
@@ -289,13 +308,35 @@ export function parseCliPrograms(
      if (normalizedAction && (statsBackground || statsStop)) {
        throw new Error('Stats background and stop flags cannot be combined with stats actions.');
      }
-      if (normalizedAction !== 'cleanup' && (options.vocab === true || options.lifetime === true)) {
-        throw new Error('Stats --vocab and --lifetime flags require the cleanup action.');
+      if (
+        normalizedAction !== 'cleanup' &&
+        (options.vocab === true || options.lifetime === true || options.duplicateLines === true)
+      ) {
+        throw new Error(
+          'Stats --vocab, --lifetime and --duplicate-lines flags require the cleanup action.',
+        );
+      }
+      if (
+        options.duplicateLines !== true &&
+        (options.dryRun === true || options.lookbackDays !== undefined)
+      ) {
+        throw new Error('Stats --dry-run and --lookback-days require --duplicate-lines.');
      }
      if (normalizedAction === 'cleanup') {
        statsCleanup = true;
        statsCleanupLifetime = options.lifetime === true;
-        statsCleanupVocab = statsCleanupLifetime ? false : options.vocab !== false;
+        statsCleanupDuplicateLines = options.duplicateLines === true;
+        const explicitModeCount = [options.vocab, options.lifetime, options.duplicateLines].filter(
+          (value) => value === true,
+        ).length;
+        if (explicitModeCount > 1) {
+          throw new Error('Stats cleanup runs one mode at a time.');
+        }
+        // Vocabulary cleanup stays the default so `stats cleanup` keeps its old meaning.
+        statsCleanupVocab =
+          statsCleanupLifetime || statsCleanupDuplicateLines ? false : options.vocab !== false;
+        statsCleanupDryRun = options.dryRun === true;
+        statsCleanupLookbackDays = parseStatsLookbackDays(options.lookbackDays);
      } else if (normalizedAction === 'rebuild' || normalizedAction === 'backfill') {
        statsCleanup = true;
        statsCleanupLifetime = true;
@@ -483,6 +524,9 @@ export function parseCliPrograms(
      statsCleanup,
      statsCleanupVocab,
      statsCleanupLifetime,
+      statsCleanupDuplicateLines,
+      statsCleanupDryRun,
+      statsCleanupLookbackDays,
      statsLogLevel,
      syncTriggered,
      syncCliTokens,
@@ -222,7 +222,7 @@ test('buildMpvEnv preserves native Wayland env for supported Hyprland and Sway a
  });
 });

-test('buildMpvBackendArgs forces an explicit X11 renderer stack when backend resolves to x11', () => {
+test('buildMpvBackendArgs pins the X11 window context when backend resolves to x11', () => {
  withPlatform('linux', () => {
    assert.deepEqual(
      buildMpvBackendArgs(makeArgs({ backend: 'x11' }), {
@@ -230,12 +230,12 @@ test('buildMpvBackendArgs forces an explicit X11 renderer stack when backend res
        WAYLAND_DISPLAY: 'wayland-0',
        XDG_SESSION_TYPE: 'wayland',
      }),
-      ['--vo=gpu', '--gpu-api=opengl', '--gpu-context=x11egl,x11'],
+      ['--gpu-context=x11vk,x11egl,x11'],
    );
  });
 });

-test('buildMpvBackendArgs forces the same X11 renderer stack for unsupported Wayland auto fallback', () => {
+test('buildMpvBackendArgs pins the same X11 window context for unsupported Wayland auto fallback', () => {
  withPlatform('linux', () => {
    assert.deepEqual(
      buildMpvBackendArgs(makeArgs({ backend: 'auto' }), {
@@ -245,7 +245,7 @@ test('buildMpvBackendArgs forces the same X11 renderer stack for unsupported Way
        XDG_CURRENT_DESKTOP: 'KDE',
        XDG_SESSION_DESKTOP: 'plasma',
      }),
-      ['--vo=gpu', '--gpu-api=opengl', '--gpu-context=x11egl,x11'],
+      ['--gpu-context=x11vk,x11egl,x11'],
    );
  });
 });
@@ -292,9 +292,7 @@ test('buildConfiguredMpvDefaultArgs appends maximized launch mode to configured
        '--secondary-sub-visibility=no',
        '--alang=ja,jp,jpn,japanese,en,eng,english,enus,en-us',
        '--slang=ja,jp,jpn,japanese,en,eng,english,enus,en-us',
-        '--vo=gpu',
-        '--gpu-api=opengl',
-        '--gpu-context=x11egl,x11',
+        '--gpu-context=x11vk,x11egl,x11',
        '--window-maximized=yes',
      ],
    );
@@ -232,6 +232,75 @@ test('parseArgs maps lifetime stats cleanup flag', () => {
  assert.equal(parsed.statsCleanupLifetime, true);
 });

+test('parseArgs maps duplicate-line stats cleanup flags', () => {
+  const parsed = parseArgs(
+    ['stats', 'cleanup', '--duplicate-lines', '--dry-run', '--lookback-days', '30'],
+    'subminer',
+    {},
+  );
+
+  assert.equal(parsed.statsCleanup, true);
+  assert.equal(parsed.statsCleanupVocab, false);
+  assert.equal(parsed.statsCleanupDuplicateLines, true);
+  assert.equal(parsed.statsCleanupDryRun, true);
+  assert.equal(parsed.statsCleanupLookbackDays, 30);
+
+  const fractional = parseArgs(
+    ['stats', 'cleanup', '--duplicate-lines', '--lookback-days', '1.5'],
+    'subminer',
+    {},
+  );
+  assert.equal(fractional.statsCleanupLookbackDays, 1);
+});
+
+test('parseArgs rejects duplicate-line flags without the duplicate-lines mode', () => {
+  const error = withProcessExitIntercept(() => {
+    parseArgs(['stats', 'cleanup', '--dry-run'], 'subminer', {});
+  });
+
+  assert.equal(error.code, 1);
+  assert.match(error.stderr, /--dry-run and --lookback-days require --duplicate-lines/);
+});
+
+test('parseArgs rejects an empty lookback value outside duplicate-line cleanup', () => {
+  const error = withProcessExitIntercept(() => {
+    parseArgs(['stats', '--lookback-days', ''], 'subminer', {});
+  });
+
+  assert.equal(error.code, 1);
+  assert.match(error.stderr, /--dry-run and --lookback-days require --duplicate-lines/);
+});
+
+test('parseArgs rejects combining explicit cleanup modes', () => {
+  for (const modes of [
+    ['--lifetime', '--duplicate-lines'],
+    ['--vocab', '--duplicate-lines'],
+    ['--vocab', '--lifetime'],
+  ]) {
+    const error = withProcessExitIntercept(() => {
+      parseArgs(['stats', 'cleanup', ...modes], 'subminer', {});
+    });
+
+    assert.equal(error.code, 1);
+    assert.match(error.stderr, /Stats cleanup runs one mode at a time/);
+  }
+});
+
+test('parseArgs rejects unusable lookback windows', () => {
+  for (const value of ['0', '0.5', '-5', 'soon']) {
+    const error = withProcessExitIntercept(() => {
+      parseArgs(
+        ['stats', 'cleanup', '--duplicate-lines', '--lookback-days', value],
+        'subminer',
+        {},
+      );
+    });
+
+    assert.equal(error.code, 1);
+    assert.match(error.stderr, /--lookback-days must be at least one day/);
+  }
+});
+
 test('parseArgs rejects cleanup-only stats flags without cleanup action', () => {
  const error = withProcessExitIntercept(() => {
    parseArgs(['stats', '--vocab'], 'subminer', {});
@@ -239,7 +308,10 @@ test('parseArgs rejects cleanup-only stats flags without cleanup action', () =>

  assert.equal(error.code, 1);
  assert.match(error.message, /exit:1/);
-  assert.match(error.stderr, /Stats --vocab and --lifetime flags require the cleanup action/);
+  assert.match(
+    error.stderr,
+    /Stats --vocab, --lifetime and --duplicate-lines flags require the cleanup action/,
+  );
 });

 test('parseArgs maps stats rebuild action to cleanup lifetime mode', () => {
@@ -142,6 +142,9 @@ export interface Args {
  statsCleanup?: boolean;
  statsCleanupVocab?: boolean;
  statsCleanupLifetime?: boolean;
+  statsCleanupDuplicateLines?: boolean;
+  statsCleanupDryRun?: boolean;
+  statsCleanupLookbackDays?: number;
  dictionaryTarget?: string;
  doctor: boolean;
  doctorRefreshKnownWords: boolean;
@@ -2,7 +2,7 @@
  "name": "subminer",
  "productName": "SubMiner",
  "desktopName": "SubMiner.desktop",
-  "version": "0.19.2",
+  "version": "0.19.3",
  "description": "All-in-one sentence mining overlay with AnkiConnect and dictionary integration",
  "packageManager": "bun@1.3.5",
  "main": "dist/main-entry.js",
@@ -89,7 +89,7 @@
    "fast-uri": "3.1.5",
    "form-data": "4.0.6",
    "ip-address": "10.2.0",
-    "js-yaml": "4.3.0",
+    "js-yaml": "4.3.1",
    "lodash": "4.18.0",
    "minimatch": "10.2.5",
    "picomatch": "4.0.4",
@@ -260,6 +260,10 @@
      {
        "from": "dist/launcher/subminer",
        "to": "launcher/subminer"
+      },
+      {
+        "from": "CHANGELOG.md",
+        "to": "CHANGELOG.md"
      }
    ]
  },
@@ -1,30 +0,0 @@
-{
-  "name": "subminer-workflow",
-  "version": "0.1.0",
-  "description": "Repo-local SubMiner agent workflow plugin for orchestration and change verification.",
-  "author": {
-    "name": "Kyle Yasuda",
-    "email": "suda@sudacode.com",
-    "url": "https://github.com/sudacode"
-  },
-  "homepage": "https://github.com/sudacode/SubMiner/tree/main/plugins/subminer-workflow",
-  "repository": "https://github.com/sudacode/SubMiner",
-  "license": "GPL-3.0-or-later",
-  "keywords": ["subminer", "workflow", "verification", "skills"],
-  "skills": "./skills/",
-  "interface": {
-    "displayName": "SubMiner Workflow",
-    "shortDescription": "SubMiner orchestration and verification.",
-    "longDescription": "Canonical repo-local plugin for SubMiner agent workflow packaging. Owns the scrum-master and change-verification skills plus helper scripts used to plan, verify, and validate changes reproducibly inside this repo.",
-    "developerName": "Kyle Yasuda",
-    "category": "Productivity",
-    "capabilities": ["Interactive", "Write"],
-    "websiteURL": "https://github.com/sudacode/SubMiner",
-    "defaultPrompt": [
-      "Use SubMiner workflow to plan and ship a feature.",
-      "Verify a SubMiner change with the plugin-owned verifier.",
-      "Plan and ship this SubMiner task."
-    ],
-    "brandColor": "#2F6B4F"
-  }
-}
@@ -1,40 +0,0 @@
-<!-- read_when: migrating or using the repo-local SubMiner workflow plugin -->
-
-# SubMiner Workflow Plugin
-
-Status: active
-Last verified: 2026-03-26
-Owner: Kyle Yasuda
-Read when: using or updating the repo-local plugin that owns SubMiner agent workflow skills
-
-This plugin is the canonical source of truth for the SubMiner agent workflow packaging.
-
-## Contents
-
- `skills/subminer-scrum-master/`
-  - intake, planning, dispatch, and handoff workflow
- `skills/subminer-change-verification/`
-  - cheap-first verification workflow plus helper scripts
-
-## Compatibility
-
- `.agents/skills/subminer-scrum-master/` is a compatibility shim that redirects to the plugin-owned skill.
- `.agents/skills/subminer-change-verification/` is a compatibility shim.
- `.agents/skills/subminer-change-verification/scripts/*.sh` remain as wrapper entrypoints so existing docs and shell history keep working.
-
-## Verification
-
-For plugin/doc/shim changes, prefer:
-
-```bash
-bun run test:docs:kb
-bash plugins/subminer-workflow/skills/subminer-change-verification/scripts/verify_subminer_change.sh --lane docs --lane core \
-  plugins/subminer-workflow \
-  .agents/skills/subminer-scrum-master/SKILL.md \
-  .agents/skills/subminer-change-verification/SKILL.md \
-  .agents/skills/subminer-change-verification/scripts/classify_subminer_diff.sh \
-  .agents/skills/subminer-change-verification/scripts/verify_subminer_change.sh \
-  .agents/plugins/marketplace.json \
-  docs/workflow/README.md \
-  docs/workflow/agent-plugins.md
-```
@@ -1,143 +0,0 @@
---
-name: 'subminer-change-verification'
-description: 'Use when working in the SubMiner repo and you need to verify code changes actually work. Covers targeted regression checks during debugging and pre-handoff verification, with cheap-first lane selection for config, docs, launcher/plugin, runtime-compat, and optional real-runtime escalation.'
---
-
-# SubMiner Change Verification
-
-Canonical source: this plugin path.
-
-Use this skill for SubMiner code changes. Default to cheap, repo-native verification first. Escalate only when the changed behavior actually depends on Electron, mpv, overlay/window tracking, or other GUI-sensitive runtime behavior.
-
-## Scripts
-
- `scripts/classify_subminer_diff.sh`
-  - Emits suggested lanes and flags from explicit paths or current git changes.
- `scripts/verify_subminer_change.sh`
-  - Runs selected lanes, captures artifacts, and writes a compact summary.
-
-If you need an explicit installed path, use the directory that contains this `SKILL.md`. The helper scripts live under:
-
-```bash
-export SUBMINER_VERIFY_SKILL="<path-to-plugin-skill>"
-```
-
-## Default workflow
-
-1. Inspect the changed files or user-requested area.
-2. Run the classifier unless you already know the right lane.
-3. Run the verifier with the cheapest sufficient lane set.
-4. If the classifier emits `flag:real-runtime-candidate`, do not jump straight to runtime verification. First run the non-runtime lanes.
-5. Escalate to explicit `--lane real-runtime --allow-real-runtime` only when cheaper lanes cannot validate the behavior claim.
-6. Return:
-   - verification summary
-   - exact commands run
-   - artifact paths
-   - skipped lanes and blockers
-
-## Quick start
-
-Plugin-source quick start:
-
-```bash
-bash plugins/subminer-workflow/skills/subminer-change-verification/scripts/classify_subminer_diff.sh
-```
-
-Installed-skill quick start:
-
-```bash
-bash "$SUBMINER_VERIFY_SKILL/scripts/classify_subminer_diff.sh"
-```
-
-Compatibility entrypoint:
-
-```bash
-bash .agents/skills/subminer-change-verification/scripts/classify_subminer_diff.sh
-```
-
-Classify explicit files:
-
-```bash
-bash plugins/subminer-workflow/skills/subminer-change-verification/scripts/classify_subminer_diff.sh \
-  launcher/main.ts \
-  plugin/subminer/lifecycle.lua \
-  src/main/runtime/mpv-client-runtime-service.ts
-```
-
-Run automatic lane selection:
-
-```bash
-bash plugins/subminer-workflow/skills/subminer-change-verification/scripts/verify_subminer_change.sh
-```
-
-Installed-skill form:
-
-```bash
-bash "$SUBMINER_VERIFY_SKILL/scripts/verify_subminer_change.sh"
-```
-
-Compatibility entrypoint:
-
-```bash
-bash .agents/skills/subminer-change-verification/scripts/verify_subminer_change.sh
-```
-
-Run targeted lanes:
-
-```bash
-bash plugins/subminer-workflow/skills/subminer-change-verification/scripts/verify_subminer_change.sh \
-  --lane launcher-plugin \
-  --lane runtime-compat
-```
-
-Dry-run to inspect planned commands and artifact layout:
-
-```bash
-bash plugins/subminer-workflow/skills/subminer-change-verification/scripts/verify_subminer_change.sh \
-  --dry-run \
-  launcher/main.ts \
-  src/main.ts
-```
-
-## Lane guidance
-
- `docs`
-  - For `docs-site/`, `docs/`, and doc-only edits.
- `config`
-  - For `src/config/` and config-template-sensitive edits.
- `stats`
-  - For `stats/` dashboard UI edits.
- `core`
-  - For general source changes where `typecheck` + `test:fast` is the best cheap signal.
- `launcher-plugin`
-  - For `launcher/`, `plugin/subminer/`, plugin gating scripts, and wrapper/mpv routing work.
- `runtime-compat`
-  - For `src/main*`, runtime/composer wiring, mpv/overlay services, window trackers, and dist-sensitive behavior.
- `real-runtime`
-  - Only after deliberate escalation.
-
-## Real Runtime Escalation
-
-Escalate only when the change claim depends on actual runtime behavior, for example:
-
- overlay appears, hides, or tracks a real mpv window
- mpv launch flags or pause-until-ready behavior
- plugin/socket/auto-start handshake under a real player
- macOS/window-tracker/focus-sensitive behavior
-
-If the environment cannot support authoritative runtime verification, report the blocker explicitly. Do not silently downgrade a runtime-required claim to a pass.
-
-## Artifact contract
-
-The verifier writes under `.tmp/skill-verification/<timestamp>/`:
-
- `summary.json`
- `summary.txt`
- `classification.txt`
- `env.txt`
- `lanes.txt`
- `steps.tsv`
- `steps/*.stdout.log`
- `steps/*.stderr.log`
-
-On failure, quote the exact failing command and point at the artifact directory.
@@ -1,171 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-usage() {
-  cat <<'EOF'
-Usage: classify_subminer_diff.sh [path ...]
-
-Emit suggested verification lanes for explicit paths or current local git changes.
-
-Output format:
-  lane:<name>
-  flag:<name>
-  reason:<text>
-EOF
-}
-
-has_item() {
-  local needle=$1
-  shift || true
-  local item
-  for item in "$@"; do
-    if [[ "$item" == "$needle" ]]; then
-      return 0
-    fi
-  done
-  return 1
-}
-
-add_lane() {
-  local lane=$1
-  if ! has_item "$lane" "${LANES[@]:-}"; then
-    LANES+=("$lane")
-  fi
-}
-
-add_flag() {
-  local flag=$1
-  if ! has_item "$flag" "${FLAGS[@]:-}"; then
-    FLAGS+=("$flag")
-  fi
-}
-
-add_reason() {
-  REASONS+=("$1")
-}
-
-collect_git_paths() {
-  local top_level
-  if ! top_level=$(git rev-parse --show-toplevel 2>/dev/null); then
-    return 0
-  fi
-
-  (
-    cd "$top_level"
-    if git rev-parse --verify HEAD >/dev/null 2>&1; then
-      git diff --name-only --relative HEAD --
-      git diff --name-only --relative --cached --
-    else
-      git diff --name-only --relative --
-      git diff --name-only --relative --cached --
-    fi
-    git ls-files --others --exclude-standard
-  ) | awk 'NF' | sort -u
-}
-
-if [[ "${1:-}" == "--help" || "${1:-}" == "-h" ]]; then
-  usage
-  exit 0
-fi
-
-declare -a PATHS=()
-declare -a LANES=()
-declare -a FLAGS=()
-declare -a REASONS=()
-
-if [[ $# -gt 0 ]]; then
-  while [[ $# -gt 0 ]]; do
-    PATHS+=("$1")
-    shift
-  done
-else
-  while IFS= read -r line; do
-    [[ -n "$line" ]] && PATHS+=("$line")
-  done < <(collect_git_paths)
-fi
-
-if [[ ${#PATHS[@]} -eq 0 ]]; then
-  add_lane "core"
-  add_reason "no changed paths detected -> default to core"
-fi
-
-for path in "${PATHS[@]}"; do
-  specialized=0
-
-  case "$path" in
-    docs-site/*|docs/*|changes/*|README.md)
-      add_lane "docs"
-      add_reason "$path -> docs"
-      specialized=1
-      ;;
-  esac
-
-  case "$path" in
-    src/config/*|src/generate-config-example.ts|src/verify-config-example.ts|docs-site/public/config.example.jsonc|config.example.jsonc)
-      add_lane "config"
-      add_reason "$path -> config"
-      specialized=1
-      ;;
-  esac
-
-  case "$path" in
-    stats/*)
-      add_lane "stats"
-      add_reason "$path -> stats"
-      specialized=1
-      ;;
-  esac
-
-  case "$path" in
-    launcher/*|plugin/subminer/*|plugin/subminer.conf|scripts/test-plugin-*|scripts/get-mpv-window-*|scripts/configure-plugin-binary-path.mjs)
-      add_lane "launcher-plugin"
-      add_reason "$path -> launcher-plugin"
-      add_flag "real-runtime-candidate"
-      add_reason "$path -> real-runtime-candidate"
-      specialized=1
-      ;;
-  esac
-
-  case "$path" in
-    src/main.ts|src/main-entry.ts|src/preload.ts|src/main/*|src/core/services/mpv*|src/core/services/overlay*|src/renderer/*|src/window-trackers/*|scripts/prepare-build-assets.mjs)
-      add_lane "runtime-compat"
-      add_reason "$path -> runtime-compat"
-      add_flag "real-runtime-candidate"
-      add_reason "$path -> real-runtime-candidate"
-      specialized=1
-      ;;
-  esac
-
-  if [[ "$specialized" == "0" ]]; then
-    case "$path" in
-      src/*|package.json|tsconfig*.json|scripts/*|Makefile)
-        add_lane "core"
-        add_reason "$path -> core"
-        ;;
-    esac
-  fi
-
-  case "$path" in
-    package.json|src/main.ts|src/main-entry.ts|src/preload.ts)
-      add_flag "broad-impact"
-      add_reason "$path -> broad-impact"
-      ;;
-  esac
-done
-
-if [[ ${#LANES[@]} -eq 0 ]]; then
-  add_lane "core"
-  add_reason "no lane-specific matches -> default to core"
-fi
-
-for lane in "${LANES[@]}"; do
-  printf 'lane:%s\n' "$lane"
-done
-
-for flag in "${FLAGS[@]}"; do
-  printf 'flag:%s\n' "$flag"
-done
-
-for reason in "${REASONS[@]}"; do
-  printf 'reason:%s\n' "$reason"
-done
@@ -1,537 +0,0 @@
-#!/usr/bin/env bash
-set -euo pipefail
-
-usage() {
-  cat <<'EOF'
-Usage: verify_subminer_change.sh [options] [path ...]
-
-Options:
-  --lane <name>             Force a verification lane. Repeatable.
-  --artifact-dir <dir>      Use an explicit artifact directory.
-  --allow-real-runtime      Allow explicit real-runtime execution.
-  --allow-real-gui          Deprecated alias for --allow-real-runtime.
-  --dry-run                 Record planned steps without executing commands.
-  --help                    Show this help text.
-
-If no lanes are supplied, the script classifies the provided paths. If no paths are
-provided, it classifies the current local git changes.
-
-Authoritative real-runtime verification should be requested with explicit path
-arguments instead of relying on inferred local git changes.
-EOF
-}
-
-timestamp() {
-  date +%Y%m%d-%H%M%S
-}
-
-timestamp_iso() {
-  date -u +%Y-%m-%dT%H:%M:%SZ
-}
-
-generate_session_id() {
-  local tmp_dir
-  tmp_dir=$(mktemp -d "${TMPDIR:-/tmp}/subminer-verify-$(timestamp)-XXXXXX")
-  basename "$tmp_dir"
-  rmdir "$tmp_dir"
-}
-
-has_item() {
-  local needle=$1
-  shift || true
-  local item
-  for item in "$@"; do
-    if [[ "$item" == "$needle" ]]; then
-      return 0
-    fi
-  done
-  return 1
-}
-
-normalize_lane_name() {
-  case "$1" in
-    real-gui)
-      printf '%s' "real-runtime"
-      ;;
-    *)
-      printf '%s' "$1"
-      ;;
-  esac
-}
-
-add_lane() {
-  local lane
-  lane=$(normalize_lane_name "$1")
-  if ! has_item "$lane" "${SELECTED_LANES[@]:-}"; then
-    SELECTED_LANES+=("$lane")
-  fi
-}
-
-add_blocker() {
-  BLOCKERS+=("$1")
-  BLOCKED=1
-}
-
-validate_artifact_dir() {
-  local candidate=$1
-  if [[ ! "$candidate" =~ ^[A-Za-z0-9._/@:+-]+$ ]]; then
-    echo "Invalid characters in --artifact-dir path" >&2
-    exit 2
-  fi
-}
-
-append_step_record() {
-  printf '%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\n' \
-    "$1" "$2" "$3" "$4" "$5" "$6" "$7" "$8" >>"$STEPS_TSV"
-}
-
-record_env() {
-  {
-    printf 'repo_root=%s\n' "$REPO_ROOT"
-    printf 'session_id=%s\n' "$SESSION_ID"
-    printf 'artifact_dir=%s\n' "$ARTIFACT_DIR"
-    printf 'path_selection_mode=%s\n' "$PATH_SELECTION_MODE"
-    printf 'dry_run=%s\n' "$DRY_RUN"
-    printf 'allow_real_runtime=%s\n' "$ALLOW_REAL_RUNTIME"
-    printf 'session_home=%s\n' "$SESSION_HOME"
-    printf 'session_xdg_config_home=%s\n' "$SESSION_XDG_CONFIG_HOME"
-    printf 'session_mpv_dir=%s\n' "$SESSION_MPV_DIR"
-    printf 'session_logs_dir=%s\n' "$SESSION_LOGS_DIR"
-    printf 'session_mpv_log=%s\n' "$SESSION_MPV_LOG"
-    printf 'pwd=%s\n' "$(pwd)"
-    git rev-parse --short HEAD 2>/dev/null | sed 's/^/git_head=/' || true
-    git status --short 2>/dev/null || true
-    if [[ ${#PATH_ARGS[@]} -gt 0 ]]; then
-      printf 'requested_paths=\n'
-      printf '  %s\n' "${PATH_ARGS[@]}"
-    fi
-  } >"$ARTIFACT_DIR/env.txt"
-}
-
-run_step() {
-  local lane=$1
-  local name=$2
-  local command=$3
-  local note=${4:-}
-  local lane_slug=${lane//[^a-zA-Z0-9_-]/-}
-  local slug=${name//[^a-zA-Z0-9_-]/-}
-  local step_slug="${lane_slug}--${slug}"
-  local stdout_rel="steps/${step_slug}.stdout.log"
-  local stderr_rel="steps/${step_slug}.stderr.log"
-  local stdout_path="$ARTIFACT_DIR/$stdout_rel"
-  local stderr_path="$ARTIFACT_DIR/$stderr_rel"
-  local status exit_code
-
-  COMMANDS_RUN+=("$command")
-  printf '%s\n' "$command" >"$ARTIFACT_DIR/steps/${step_slug}.command.txt"
-
-  if [[ "$DRY_RUN" == "1" ]]; then
-    printf '[dry-run] %s\n' "$command" >"$stdout_path"
-    : >"$stderr_path"
-    status="dry-run"
-    exit_code=0
-  else
-    if HOME="$SESSION_HOME" \
-      XDG_CONFIG_HOME="$SESSION_XDG_CONFIG_HOME" \
-      SUBMINER_SESSION_LOGS_DIR="$SESSION_LOGS_DIR" \
-      SUBMINER_SESSION_MPV_LOG="$SESSION_MPV_LOG" \
-      bash -c "cd \"$REPO_ROOT\" && $command" >"$stdout_path" 2>"$stderr_path"; then
-      status="passed"
-      exit_code=0
-      EXECUTED_REAL_STEPS=1
-    else
-      exit_code=$?
-      status="failed"
-      FAILED=1
-    fi
-  fi
-
-  append_step_record "$lane" "$name" "$status" "$exit_code" "$command" "$stdout_rel" "$stderr_rel" "$note"
-  printf '%s\t%s\t%s\n' "$lane" "$name" "$status"
-
-  if [[ "$status" == "failed" ]]; then
-    FAILURE_STEP="$name"
-    FAILURE_COMMAND="$command"
-    FAILURE_STDOUT="$stdout_rel"
-    FAILURE_STDERR="$stderr_rel"
-    return "$exit_code"
-  fi
-}
-
-record_nonpassing_step() {
-  local lane=$1
-  local name=$2
-  local status=$3
-  local note=$4
-  local lane_slug=${lane//[^a-zA-Z0-9_-]/-}
-  local slug=${name//[^a-zA-Z0-9_-]/-}
-  local step_slug="${lane_slug}--${slug}"
-  local stdout_rel="steps/${step_slug}.stdout.log"
-  local stderr_rel="steps/${step_slug}.stderr.log"
-  printf '%s\n' "$note" >"$ARTIFACT_DIR/$stdout_rel"
-  : >"$ARTIFACT_DIR/$stderr_rel"
-  append_step_record "$lane" "$name" "$status" "0" "" "$stdout_rel" "$stderr_rel" "$note"
-  printf '%s\t%s\t%s\n' "$lane" "$name" "$status"
-}
-
-record_skipped_step() {
-  record_nonpassing_step "$1" "$2" "skipped" "$3"
-}
-
-record_blocked_step() {
-  add_blocker "$3"
-  record_nonpassing_step "$1" "$2" "blocked" "$3"
-}
-
-record_failed_step() {
-  FAILED=1
-  FAILURE_STEP=$2
-  FAILURE_COMMAND=${FAILURE_COMMAND:-"(validation)"}
-  local lane_slug=${1//[^a-zA-Z0-9_-]/-}
-  local step_slug=${2//[^a-zA-Z0-9_-]/-}
-  FAILURE_STDOUT="steps/${lane_slug}--${step_slug}.stdout.log"
-  FAILURE_STDERR="steps/${lane_slug}--${step_slug}.stderr.log"
-  add_blocker "$3"
-  record_nonpassing_step "$1" "$2" "failed" "$3"
-}
-
-find_real_runtime_helper() {
-  local candidate
-  for candidate in \
-    "$SCRIPT_DIR/run_real_runtime_smoke.sh" \
-    "$SCRIPT_DIR/run_real_mpv_smoke.sh"; do
-    if [[ -x "$candidate" ]]; then
-      printf '%s' "$candidate"
-      return 0
-    fi
-  done
-  return 1
-}
-
-acquire_real_runtime_lease() {
-  local lease_root="$REPO_ROOT/.tmp/skill-verification/locks"
-  local lease_dir="$lease_root/exclusive-real-runtime"
-  mkdir -p "$lease_root"
-  if mkdir "$lease_dir" 2>/dev/null; then
-    REAL_RUNTIME_LEASE_DIR="$lease_dir"
-    printf '%s\n' "$SESSION_ID" >"$lease_dir/session_id"
-    return 0
-  fi
-
-  local owner=""
-  if [[ -f "$lease_dir/session_id" ]]; then
-    owner=$(cat "$lease_dir/session_id")
-  fi
-  REAL_RUNTIME_LEASE_ERROR="real-runtime lease already held${owner:+ by $owner}"
-  return 1
-}
-
-release_real_runtime_lease() {
-  if [[ -n "$REAL_RUNTIME_LEASE_DIR" && -d "$REAL_RUNTIME_LEASE_DIR" ]]; then
-    if [[ -f "$REAL_RUNTIME_LEASE_DIR/session_id" ]]; then
-      local owner
-      owner=$(cat "$REAL_RUNTIME_LEASE_DIR/session_id")
-      if [[ "$owner" != "$SESSION_ID" ]]; then
-        return 0
-      fi
-    fi
-    rm -rf "$REAL_RUNTIME_LEASE_DIR"
-  fi
-}
-
-compute_final_status() {
-  if [[ "$FAILED" == "1" ]]; then
-    FINAL_STATUS="failed"
-  elif [[ "$BLOCKED" == "1" ]]; then
-    FINAL_STATUS="blocked"
-  elif [[ "$EXECUTED_REAL_STEPS" == "1" ]]; then
-    FINAL_STATUS="passed"
-  else
-    FINAL_STATUS="skipped"
-  fi
-}
-
-write_summary_files() {
-  local lane_lines
-  lane_lines=$(printf '%s\n' "${SELECTED_LANES[@]}")
-  printf '%s\n' "$lane_lines" >"$ARTIFACT_DIR/lanes.txt"
-  # bash 3.2 raises "unbound variable" under set -u when expanding an empty
-  # array, so guard on length (matching the idiom used elsewhere here).
-  if [[ ${#BLOCKERS[@]} -gt 0 ]]; then
-    printf '%s\n' "${BLOCKERS[@]}" >"$ARTIFACT_DIR/blockers.txt"
-  else
-    : >"$ARTIFACT_DIR/blockers.txt"
-  fi
-  if [[ ${#PATH_ARGS[@]} -gt 0 ]]; then
-    printf '%s\n' "${PATH_ARGS[@]}" >"$ARTIFACT_DIR/requested-paths.txt"
-  else
-    : >"$ARTIFACT_DIR/requested-paths.txt"
-  fi
-
-  ARTIFACT_DIR_ENV="$ARTIFACT_DIR" \
-  SESSION_ID_ENV="$SESSION_ID" \
-  FINAL_STATUS_ENV="$FINAL_STATUS" \
-  PATH_SELECTION_MODE_ENV="$PATH_SELECTION_MODE" \
-  ALLOW_REAL_RUNTIME_ENV="$ALLOW_REAL_RUNTIME" \
-  SESSION_HOME_ENV="$SESSION_HOME" \
-  SESSION_XDG_CONFIG_HOME_ENV="$SESSION_XDG_CONFIG_HOME" \
-  SESSION_MPV_DIR_ENV="$SESSION_MPV_DIR" \
-  SESSION_LOGS_DIR_ENV="$SESSION_LOGS_DIR" \
-  SESSION_MPV_LOG_ENV="$SESSION_MPV_LOG" \
-  STARTED_AT_ENV="$STARTED_AT" \
-  FINISHED_AT_ENV="$FINISHED_AT" \
-  FAILED_ENV="$FAILED" \
-  FAILURE_COMMAND_ENV="${FAILURE_COMMAND:-}" \
-  FAILURE_STDOUT_ENV="${FAILURE_STDOUT:-}" \
-  FAILURE_STDERR_ENV="${FAILURE_STDERR:-}" \
-  bun -e '
-    const fs = require("fs");
-    const path = require("path");
-
-    const lines = fs
-      .readFileSync(path.join(process.env.ARTIFACT_DIR_ENV, "steps.tsv"), "utf8")
-      .trim()
-      .split("\n")
-      .filter(Boolean)
-      .slice(1)
-      .map((line) => {
-        const [lane, name, status, exitCode, command, stdout, stderr, note] = line.split("\t");
-        return { lane, name, status, exitCode: Number(exitCode), command, stdout, stderr, note };
-      });
-
-    const payload = {
-      sessionId: process.env.SESSION_ID_ENV,
-      startedAt: process.env.STARTED_AT_ENV,
-      finishedAt: process.env.FINISHED_AT_ENV,
-      status: process.env.FINAL_STATUS_ENV,
-      pathSelectionMode: process.env.PATH_SELECTION_MODE_ENV,
-      allowRealRuntime: process.env.ALLOW_REAL_RUNTIME_ENV === "1",
-      sessionHome: process.env.SESSION_HOME_ENV,
-      sessionXdgConfigHome: process.env.SESSION_XDG_CONFIG_HOME_ENV,
-      sessionMpvDir: process.env.SESSION_MPV_DIR_ENV,
-      sessionLogsDir: process.env.SESSION_LOGS_DIR_ENV,
-      sessionMpvLog: process.env.SESSION_MPV_LOG_ENV,
-      failed: process.env.FAILED_ENV === "1",
-      failure: process.env.FAILURE_COMMAND_ENV
-        ? {
-            command: process.env.FAILURE_COMMAND_ENV,
-            stdout: process.env.FAILURE_STDOUT_ENV,
-            stderr: process.env.FAILURE_STDERR_ENV,
-          }
-        : null,
-      blockers: fs
-        .readFileSync(path.join(process.env.ARTIFACT_DIR_ENV, "blockers.txt"), "utf8")
-        .split("\n")
-        .filter(Boolean),
-      lanes: fs
-        .readFileSync(path.join(process.env.ARTIFACT_DIR_ENV, "lanes.txt"), "utf8")
-        .split("\n")
-        .filter(Boolean),
-      requestedPaths: fs
-        .readFileSync(path.join(process.env.ARTIFACT_DIR_ENV, "requested-paths.txt"), "utf8")
-        .split("\n")
-        .filter(Boolean),
-      steps: lines,
-    };
-
-    fs.writeFileSync(
-      path.join(process.env.ARTIFACT_DIR_ENV, "summary.json"),
-      JSON.stringify(payload, null, 2) + "\n",
-    );
-
-    const summaryLines = [
-      `status: ${payload.status}`,
-      `session: ${payload.sessionId}`,
-      `artifacts: ${process.env.ARTIFACT_DIR_ENV}`,
-      `lanes: ${payload.lanes.join(", ") || "(none)"}`,
-    ];
-
-    if (payload.requestedPaths.length > 0) {
-      summaryLines.push("requested paths:");
-      for (const entry of payload.requestedPaths) {
-        summaryLines.push(`- ${entry}`);
-      }
-    }
-
-    if (payload.failure) {
-      summaryLines.push(`failure command: ${payload.failure.command}`);
-      summaryLines.push(`failure stdout: ${payload.failure.stdout}`);
-      summaryLines.push(`failure stderr: ${payload.failure.stderr}`);
-    }
-
-    if (payload.blockers.length > 0) {
-      summaryLines.push("blockers:");
-      for (const blocker of payload.blockers) {
-        summaryLines.push(`- ${blocker}`);
-      }
-    }
-
-    summaryLines.push("steps:");
-    for (const step of payload.steps) {
-      summaryLines.push(`- ${step.lane}/${step.name}: ${step.status}`);
-    }
-
-    fs.writeFileSync(
-      path.join(process.env.ARTIFACT_DIR_ENV, "summary.txt"),
-      summaryLines.join("\n") + "\n",
-    );
-  '
-}
-
-SCRIPT_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
-SKILL_DIR=$(cd "$SCRIPT_DIR/.." && pwd)
-REPO_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
-
-declare -a PATH_ARGS=()
-declare -a SELECTED_LANES=()
-declare -a COMMANDS_RUN=()
-declare -a BLOCKERS=()
-
-ALLOW_REAL_RUNTIME=0
-DRY_RUN=0
-FAILED=0
-BLOCKED=0
-EXECUTED_REAL_STEPS=0
-FAILURE_STEP=""
-FAILURE_COMMAND=""
-FAILURE_STDOUT=""
-FAILURE_STDERR=""
-REAL_RUNTIME_LEASE_DIR=""
-REAL_RUNTIME_LEASE_ERROR=""
-PATH_SELECTION_MODE="auto"
-
-trap 'release_real_runtime_lease' EXIT
-
-while [[ $# -gt 0 ]]; do
-  case "$1" in
-    --lane)
-      shift
-      [[ $# -gt 0 ]] || {
-        echo "Missing value for --lane" >&2
-        exit 2
-      }
-      add_lane "$1"
-      PATH_SELECTION_MODE="explicit-lanes"
-      ;;
-    --artifact-dir)
-      shift
-      [[ $# -gt 0 ]] || {
-        echo "Missing value for --artifact-dir" >&2
-        exit 2
-      }
-      ARTIFACT_DIR=$1
-      ;;
-    --allow-real-runtime|--allow-real-gui)
-      ALLOW_REAL_RUNTIME=1
-      ;;
-    --dry-run)
-      DRY_RUN=1
-      ;;
-    --help|-h)
-      usage
-      exit 0
-      ;;
-    *)
-      PATH_ARGS+=("$1")
-      ;;
-  esac
-  shift || true
-done
-
-if [[ -z "${ARTIFACT_DIR:-}" ]]; then
-  SESSION_ID=$(generate_session_id)
-  ARTIFACT_DIR="$REPO_ROOT/.tmp/skill-verification/$SESSION_ID"
-else
-  validate_artifact_dir "$ARTIFACT_DIR"
-  SESSION_ID=$(basename "$ARTIFACT_DIR")
-fi
-
-mkdir -p "$ARTIFACT_DIR/steps"
-STEPS_TSV="$ARTIFACT_DIR/steps.tsv"
-printf 'lane\tstep\tstatus\texit_code\tcommand\tstdout\tstderr\tnote\n' >"$STEPS_TSV"
-
-STARTED_AT=$(timestamp_iso)
-SESSION_HOME="$REPO_ROOT/.tmp/skill-verification/runtime/$SESSION_ID/home"
-SESSION_XDG_CONFIG_HOME="$REPO_ROOT/.tmp/skill-verification/runtime/$SESSION_ID/xdg-config"
-SESSION_MPV_DIR="$SESSION_XDG_CONFIG_HOME/mpv"
-SESSION_LOGS_DIR="$REPO_ROOT/.tmp/skill-verification/runtime/$SESSION_ID/logs"
-SESSION_MPV_LOG="$SESSION_LOGS_DIR/mpv.log"
-mkdir -p "$SESSION_HOME" "$SESSION_MPV_DIR" "$SESSION_LOGS_DIR"
-
-CLASSIFIER_OUTPUT="$ARTIFACT_DIR/classification.txt"
-if [[ ${#SELECTED_LANES[@]} -eq 0 ]]; then
-  if [[ ${#PATH_ARGS[@]} -gt 0 ]]; then
-    PATH_SELECTION_MODE="explicit-paths"
-  fi
-  if "$SCRIPT_DIR/classify_subminer_diff.sh" "${PATH_ARGS[@]}" >"$CLASSIFIER_OUTPUT"; then
-    while IFS= read -r line; do
-      case "$line" in
-        lane:*)
-          add_lane "${line#lane:}"
-          ;;
-      esac
-    done <"$CLASSIFIER_OUTPUT"
-  else
-    record_failed_step "meta" "classify" "classification failed"
-  fi
-else
-  : >"$CLASSIFIER_OUTPUT"
-fi
-
-record_env
-
-if [[ ${#SELECTED_LANES[@]} -eq 0 ]]; then
-  add_lane "core"
-fi
-
-for lane in "${SELECTED_LANES[@]}"; do
-  case "$lane" in
-    docs)
-      run_step "$lane" "docs-kb" "bun run test:docs:kb" || break
-      ;;
-    config)
-      run_step "$lane" "config" "bun run test:config" || break
-      ;;
-    stats)
-      run_step "$lane" "stats" "bun run test:stats" || break
-      ;;
-    core)
-      run_step "$lane" "typecheck" "bun run typecheck" || break
-      run_step "$lane" "fast-tests" "bun run test:fast" || break
-      ;;
-    launcher-plugin)
-      run_step "$lane" "launcher" "bun run test:launcher" || break
-      run_step "$lane" "plugin-src" "bun run test:plugin:src" || break
-      ;;
-    runtime-compat)
-      run_step "$lane" "runtime-compat" "bun run test:runtime:compat" || break
-      ;;
-    real-runtime)
-      if [[ "$ALLOW_REAL_RUNTIME" != "1" ]]; then
-        record_blocked_step "$lane" "real-runtime" "real-runtime requested without --allow-real-runtime"
-        continue
-      fi
-      if ! acquire_real_runtime_lease; then
-        record_blocked_step "$lane" "real-runtime-lease" "$REAL_RUNTIME_LEASE_ERROR"
-        continue
-      fi
-      helper=$(find_real_runtime_helper || true)
-      if [[ -z "${helper:-}" ]]; then
-        record_blocked_step "$lane" "real-runtime-helper" "no real-runtime helper script available in $SCRIPT_DIR"
-        continue
-      fi
-      run_step "$lane" "real-runtime" "\"$helper\" \"$SESSION_ID\" \"$ARTIFACT_DIR\"" || break
-      ;;
-    *)
-      record_blocked_step "$lane" "unknown-lane" "unknown lane: $lane"
-      ;;
-  esac
-done
-
-release_real_runtime_lease
-FINISHED_AT=$(timestamp_iso)
-compute_final_status
-write_summary_files
-
-printf 'summary:%s\n' "$ARTIFACT_DIR/summary.txt"
-cat "$ARTIFACT_DIR/summary.txt"
@@ -1,118 +0,0 @@
---
-name: 'subminer-scrum-master'
-description: 'Use in the SubMiner repo when a request should be turned into planned work and driven through execution. Records a plan, dispatches one or more subagents when useful, and requires verification before handoff.'
---
-
-# SubMiner Scrum Master
-
-Canonical source: this plugin path.
-
-Own workflow, not code by default.
-
-Use this skill when the user gives a feature request, bug report, issue, refactor, or implementation ask and the agent should manage intake, planning, worker dispatch, and verification through completion.
-
-## Core Rules
-
-1. Keep the process light for questions, obvious mechanical edits, and tiny isolated changes.
-2. Record a plan before dispatching coding work.
-3. Split multi-part work into clear phases and ownership areas.
-4. Dispatch conservatively. Parallelize only disjoint write scopes.
-5. Require verification before handoff, typically via `subminer-change-verification`.
-6. Report dispatched workers, verification, blockers, and remaining risks.
-
-## Intake Workflow
-
-1. Parse the request.
-   Classify it as question, mechanical edit, bugfix, feature, refactor, investigation, or follow-up.
-2. Write a short working plan in-thread when the work is nontrivial.
-3. Choose execution mode:
-   - no subagents for trivial work
-   - one worker for focused work
-   - parallel workers only for disjoint scopes
-4. Run verification before handoff.
-
-## Dispatch Rules
-
-The scrum master orchestrates. Workers implement.
-
- Do not become the default implementer unless delegation is unnecessary.
- Do not parallelize overlapping files or tightly coupled runtime work.
- Give every worker explicit ownership of files/modules.
- Tell every worker other agents may be active and they must not revert unrelated edits.
- Require each worker to report:
-  - changed files
-  - tests run
-  - blockers
-
-Use worker agents for implementation and explorer agents only for bounded codebase questions.
-
-## Verification
-
-Every nontrivial code task gets verification.
-
-Preferred flow:
-
-1. use `subminer-change-verification`
-2. start with the cheapest sufficient lane
-3. escalate only when needed
-4. if worker verification is sufficient, accept it or run one final consolidating pass
-
-Never hand off nontrivial work without stating what was verified and what was skipped.
-
-## Pre-Handoff Policy Checks
-
-Before handoff, always ask and answer both questions explicitly:
-
-1. Docs update required?
-2. Changelog fragment required?
-
-Rules:
-
- Do not assume silence implies "no."
- If the answer is yes, complete the update or report the blocker.
- Include final yes/no answers in the handoff summary even when both answers are "no."
-
-## Failure / Scope Handling
-
- If a worker hits ambiguity, pause and ask the user.
- If verification fails, either:
-  - send the worker back with exact failure context, or
-  - fix it directly if it is tiny and clearly in scope
- If new scope appears, pause and re-plan before silently expanding work.
-
-## Representative Flows
-
-### Trivial work
-
- keep a short plan
- implement directly or with one worker if helpful
- run targeted verification
- report outcome concisely
-
-### Focused implementation
-
- record plan
- dispatch one worker
- integrate
- verify
- report outcome
-
-### Multi-part execution
-
- define distinct deliverables/phases
- record sequencing in the plan
- dispatch workers only where scopes are disjoint
- integrate
- run consolidated verification
- report outcome
-
-## Output Expectations
-
-At the end, report:
-
- which workers were dispatched and what they owned
- what verification ran
- explicit answers to:
-  - docs update required?
-  - changelog fragment required?
- blockers, skips, and risks
@@ -1,30 +1,38 @@
 ## Highlights
-### Changed
+### Added
+- **In-App Changelog**
+  - View release notes without leaving the app, from the tray menu ("View Changelog") or the "What's New" button on update notifications.
+  - Shows notes for the latest published release even when it's newer than your installed build, and falls back to the notes bundled with your install if the download fails.
+  - Older versions fold automatically, your installed version is badged, and newer ones are tagged "New"; navigate with `J`/`K` or the arrow keys, `Enter` to expand/collapse, `R` to refresh, and `Esc` to close.

- **Subsync Reference & Target Picker**
-  - You can now choose both sides of a sync run: which subtitle is the timing reference and which one gets retimed.
-  - The video file itself can be used as the reference for local files (audio-based sync), though a subtitle track stays the default.
-  - Works for both alass and ffsubsync, and retiming the secondary subtitle track no longer overwrites your primary one.
+### Changed
+- **Faster Subtitle Tokenization**
+  - Subtitle lines are parsed and looked up roughly twice as efficiently, with results cached across lines so repeated words and grammar no longer re-query the dictionary.
+  - Enabling a character dictionary no longer slows subtitle scanning as much, since name lookups now only check positions where a known name can actually start.
+  - Fixed related accuracy issues along the way: readings that could go missing on certain word endings, subtitle text that stayed unannotated after mining a card, character names that could drop out of disambiguation rules, and halfwidth-katakana character names that weren't recognized or read correctly.

 ### Fixed
-
- **Startup Logging**
-  - Background startup now respects your configured log level even when no `--log-level` flag is passed.
- **Streaming Subtitle Tokenization**
-  - Jellyfin playback now seeds tokenization straight from the downloaded subtitle file, so episodes no longer fall back to slow, line-by-line tokenizing while waiting on playback events.
-  - Subtitle cues are no longer dropped when switching to a subtitle track embedded in the stream.
-  - Prefetching now runs through the whole episode instead of stopping once the cache filled, and the cache clears between episodes so slowdowns don't carry over to later titles.
-  - The tokenization cache was expanded from 256 to 2,500 lines, leaving more room for repeated lines (like openings and endings) to stay cached across episodes.
- **Subtitle Line Display**
-  - Subtitle lines now appear immediately at their cue time even if tokenization hasn't finished, upgrading in place with annotations once ready.
-  - A failed tokenization attempt is no longer cached as plain text, so the line gets another chance at full annotations later.
+- **Large Character Dictionary Generation**
+  - Big character dictionaries (long-running series like One Piece) no longer fail to install with a timeout error; the import time budget now scales with dictionary size instead of using a fixed 7-second limit.
+  - The "Generating character dictionary" notification now shows real progress (character/page counts, image download progress with an ETA, name-processing progress) and an elapsed-time clock, so a long-running import no longer looks frozen.
+- **Stats Deletion Responsiveness**
+  - Deleting sessions, episodes, or library entries on the stats page no longer freezes the page or an active video player; deletes are now batched into a single transaction.
+- **Subtitle Sidebar Clutter from Styled Subtitles**
+  - Heavily typeset subtitles (karaoke openings/endings, stylized signs) no longer flood the subtitle sidebar with garbled vector-drawing text or duplicate "shadow" copies of the same line.
+  - Subtitle text is now decoded consistently in one place, matching what mpv actually renders on screen, so it can no longer diverge or get cached inconsistently.
+- **X11/XWayland Playback and Overlay Fixes**
+  - Fixed a crash on the first fullscreen toggle when using an mpv `gpu-next` shader (e.g. ArtCNN) in X11/XWayland mode; SubMiner no longer forces mpv onto its older OpenGL renderer.
+  - Fixed the overlay appearing oversized and offset from the video under fractional or mixed-monitor display scaling in X11/XWayland mode.

 ## What's Changed

- feat(subsync): add reference and target subtitle track picker by @ksyasuda in #181
- fix(logging): surface subtitle processing debug/warn logs by @ksyasuda in #182
- fix(streaming): keep subtitle tokenization prefetch warm for full episodes by @ksyasuda in #183
- fix(overlay): show plain subtitle line immediately on tokenization cache miss by @ksyasuda in #184
+- perf(tokenizer): single-pass Yomitan scan with cross-line caching and prefetch fixes by @ksyasuda in #185
+- fix(subtitles): collapse duplicate ASS events and decode text once by @ksyasuda in #186
+- feat(overlay): add in-app changelog modal by @ksyasuda in #187
+- fix(playback): stop forcing legacy OpenGL renderer on X11 mpv backend by @ksyasuda in #188
+- fix(dictionary): stop large character dictionaries from timing out by @ksyasuda in #189
+- fix(overlay): handle X11 display scaling across monitors by @ksyasuda in #193
+- fix(stats): batch deletes off the main thread by @ksyasuda in #194

 ## Installation

@@ -19,6 +19,7 @@ const requiredDocs = [
  'docs/knowledge-base/catalog.md',
  'docs/knowledge-base/quality.md',
  'docs/workflow/README.md',
+  'docs/workflow/agent-skills.md',
  'docs/workflow/planning.md',
  'docs/workflow/verification.md',
 ] as const;
@@ -1,202 +0,0 @@
-import assert from 'node:assert/strict';
-import fs from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
-import { spawnSync } from 'node:child_process';
-import test from 'node:test';
-
-const repoRoot = process.cwd();
-const classifyScript = path.join(
-  repoRoot,
-  '.agents/skills/subminer-change-verification/scripts/classify_subminer_diff.sh',
-);
-const verifyScript = path.join(
-  repoRoot,
-  '.agents/skills/subminer-change-verification/scripts/verify_subminer_change.sh',
-);
-
-function withTempDir<T>(fn: (dir: string) => T): T {
-  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-change-verification-test-'));
-  try {
-    return fn(dir);
-  } finally {
-    fs.rmSync(dir, { recursive: true, force: true });
-  }
-}
-
-function runBash(args: string[]) {
-  return spawnSync('bash', args, {
-    cwd: repoRoot,
-    env: process.env,
-    encoding: 'utf8',
-  });
-}
-
-function parseArtifactDir(stdout: string): string {
-  const match = stdout.match(/^artifacts: (.+)$/m);
-  assert.ok(match, `expected artifact_dir in stdout, got:\n${stdout}`);
-  return match[1] ?? '';
-}
-
-function readSummaryJson(artifactDir: string) {
-  return JSON.parse(fs.readFileSync(path.join(artifactDir, 'summary.json'), 'utf8')) as {
-    sessionId: string;
-    status: string;
-    lanes: string[];
-    blockers?: string[];
-    artifactDir: string;
-    pathSelectionMode?: string;
-    steps: Array<{
-      lane: string;
-      name: string;
-      stdout: string;
-      stderr: string;
-      note: string;
-    }>;
-  };
-}
-
-test('classifier marks launcher and plugin paths as real-runtime candidates', () => {
-  const result = runBash([classifyScript, 'launcher/mpv.ts', 'plugin/subminer/process.lua']);
-
-  assert.equal(result.status, 0, result.stderr || result.stdout);
-  assert.match(result.stdout, /^lane:launcher-plugin$/m);
-  assert.match(result.stdout, /^flag:real-runtime-candidate$/m);
-  assert.doesNotMatch(result.stdout, /real-gui-candidate/);
-});
-
-test('verifier blocks requested real-runtime lane when runtime execution is not allowed', () => {
-  withTempDir((root) => {
-    const artifactDir = path.join(root, 'artifacts');
-    const result = runBash([
-      verifyScript,
-      '--dry-run',
-      '--artifact-dir',
-      artifactDir,
-      '--lane',
-      'real-runtime',
-      'launcher/mpv.ts',
-    ]);
-
-    assert.equal(result.status, 0, result.stdout);
-
-    const summary = readSummaryJson(artifactDir);
-    assert.equal(summary.status, 'blocked');
-    assert.deepEqual(summary.lanes, ['real-runtime']);
-    assert.ok(summary.sessionId.length > 0);
-    assert.ok(summary.blockers?.some((entry) => entry.includes('--allow-real-runtime')));
-    assert.equal(fs.existsSync(path.join(artifactDir, 'summary.json')), true);
-  });
-});
-
-test('verifier fails closed for unknown lanes', () => {
-  withTempDir((root) => {
-    const artifactDir = path.join(root, 'artifacts');
-    const result = runBash([
-      verifyScript,
-      '--dry-run',
-      '--artifact-dir',
-      artifactDir,
-      '--lane',
-      'not-a-lane',
-      'src/main.ts',
-    ]);
-
-    assert.equal(result.status, 0, result.stdout);
-
-    const summary = readSummaryJson(artifactDir);
-    assert.equal(summary.status, 'blocked');
-    assert.deepEqual(summary.lanes, ['not-a-lane']);
-    assert.ok(summary.blockers?.some((entry) => entry.includes('unknown lane')));
-  });
-});
-
-test('verifier keeps non-passing step artifacts distinct across lanes', () => {
-  withTempDir((root) => {
-    const artifactDir = path.join(root, 'artifacts');
-    const result = runBash([
-      verifyScript,
-      '--dry-run',
-      '--artifact-dir',
-      artifactDir,
-      '--lane',
-      'docs',
-      '--lane',
-      'not-a-lane',
-      'src/main.ts',
-    ]);
-
-    assert.equal(result.status, 0, result.stdout);
-
-    const summary = readSummaryJson(artifactDir);
-    const docsStep = summary.steps.find((step) => step.lane === 'docs' && step.name === 'docs-kb');
-    const unknownStep = summary.steps.find(
-      (step) => step.lane === 'not-a-lane' && step.name === 'unknown-lane',
-    );
-
-    assert.ok(docsStep);
-    assert.ok(unknownStep);
-    assert.notEqual(docsStep?.stdout, unknownStep?.stdout);
-    assert.equal(fs.existsSync(path.join(artifactDir, docsStep!.stdout)), true);
-    assert.equal(fs.existsSync(path.join(artifactDir, unknownStep!.stdout)), true);
-  });
-});
-
-test('verifier records the real-runtime lease blocker once', () => {
-  withTempDir((root) => {
-    const artifactDir = path.join(root, 'artifacts');
-    const leaseDir = path.join(
-      repoRoot,
-      '.tmp',
-      'skill-verification',
-      'locks',
-      'exclusive-real-runtime',
-    );
-    fs.mkdirSync(leaseDir, { recursive: true });
-    fs.writeFileSync(path.join(leaseDir, 'session_id'), 'other-session');
-
-    try {
-      const result = runBash([
-        verifyScript,
-        '--dry-run',
-        '--artifact-dir',
-        artifactDir,
-        '--allow-real-runtime',
-        '--lane',
-        'real-runtime',
-        'launcher/mpv.ts',
-      ]);
-
-      assert.equal(result.status, 0, result.stdout);
-
-      const summary = readSummaryJson(artifactDir);
-      assert.deepEqual(summary.blockers, ['real-runtime lease already held by other-session']);
-    } finally {
-      fs.rmSync(leaseDir, { recursive: true, force: true });
-    }
-  });
-});
-
-test('verifier allocates unique session ids and artifact roots by default', () => {
-  const first = runBash([verifyScript, '--dry-run', '--lane', 'core', 'src/main.ts']);
-  const second = runBash([verifyScript, '--dry-run', '--lane', 'core', 'src/main.ts']);
-
-  assert.equal(first.status, 0, first.stderr || first.stdout);
-  assert.equal(second.status, 0, second.stderr || second.stdout);
-
-  const firstArtifactDir = parseArtifactDir(first.stdout);
-  const secondArtifactDir = parseArtifactDir(second.stdout);
-
-  try {
-    const firstSummary = readSummaryJson(firstArtifactDir);
-    const secondSummary = readSummaryJson(secondArtifactDir);
-
-    assert.notEqual(firstSummary.sessionId, secondSummary.sessionId);
-    assert.notEqual(firstArtifactDir, secondArtifactDir);
-    assert.equal(firstSummary.pathSelectionMode, 'explicit-lanes');
-    assert.equal(secondSummary.pathSelectionMode, 'explicit-lanes');
-  } finally {
-    fs.rmSync(firstArtifactDir, { recursive: true, force: true });
-    fs.rmSync(secondArtifactDir, { recursive: true, force: true });
-  }
-});
@@ -399,6 +399,30 @@ test('hasExplicitCommand and shouldStartApp preserve command intent', () => {
  assert.equal(statsLifetimeRebuild.statsCleanupLifetime, true);
  assert.equal(statsLifetimeRebuild.statsCleanupVocab, false);

+  assert.throws(
+    () =>
+      parseArgs([
+        '--stats',
+        '--stats-cleanup',
+        '--stats-cleanup-duplicate-lines',
+        '--stats-cleanup-lookback-days',
+        '0.5',
+      ]),
+    /at least one day/,
+  );
+  assert.equal(
+    parseArgs([
+      '--stats',
+      '--stats-cleanup',
+      '--stats-cleanup-duplicate-lines',
+      '--stats-cleanup-lookback-days',
+      '1.5',
+    ]).statsCleanupLookbackDays,
+    1,
+  );
+  assert.equal(parseArgs(['--stats-cleanup-lookback-days=30']).statsCleanupLookbackDays, 30);
+  assert.throws(() => parseArgs(['--stats-cleanup-lookback-days=30=oops']), /at least one day/);
+
  const jellyfinLibraries = parseArgs(['--jellyfin-libraries']);
  assert.equal(jellyfinLibraries.jellyfinLibraries, true);
  assert.equal(hasExplicitCommand(jellyfinLibraries), true);
@@ -64,6 +64,9 @@ export interface CliArgs {
  statsCleanup?: boolean;
  statsCleanupVocab?: boolean;
  statsCleanupLifetime?: boolean;
+  statsCleanupDuplicateLines?: boolean;
+  statsCleanupDryRun?: boolean;
+  statsCleanupLookbackDays?: number;
  statsResponsePath?: string;
  jellyfin: boolean;
  jellyfinLogin: boolean;
@@ -109,6 +112,14 @@ export interface CliArgs {

 export type CliCommandSource = 'initial' | 'second-instance';

+function parseStatsCleanupLookbackDays(value: string | undefined): number {
+  const days = Number(value);
+  if (!Number.isFinite(days) || days < 1) {
+    throw new Error('Stats --lookback-days must be at least one day.');
+  }
+  return Math.floor(days);
+}
+
 export function parseArgs(argv: string[]): CliArgs {
  const args: CliArgs = {
    background: false,
@@ -167,6 +178,8 @@ export function parseArgs(argv: string[]): CliArgs {
    statsCleanup: false,
    statsCleanupVocab: false,
    statsCleanupLifetime: false,
+    statsCleanupDuplicateLines: false,
+    statsCleanupDryRun: false,
    jellyfin: false,
    jellyfinLogin: false,
    jellyfinLogout: false,
@@ -368,7 +381,15 @@ export function parseArgs(argv: string[]): CliArgs {
    } else if (arg === '--stats-cleanup') args.statsCleanup = true;
    else if (arg === '--stats-cleanup-vocab') args.statsCleanupVocab = true;
    else if (arg === '--stats-cleanup-lifetime') args.statsCleanupLifetime = true;
-    else if (arg.startsWith('--stats-response-path=')) {
+    else if (arg === '--stats-cleanup-duplicate-lines') args.statsCleanupDuplicateLines = true;
+    else if (arg === '--stats-cleanup-dry-run') args.statsCleanupDryRun = true;
+    else if (arg.startsWith('--stats-cleanup-lookback-days=')) {
+      args.statsCleanupLookbackDays = parseStatsCleanupLookbackDays(
+        arg.slice('--stats-cleanup-lookback-days='.length),
+      );
+    } else if (arg === '--stats-cleanup-lookback-days') {
+      args.statsCleanupLookbackDays = parseStatsCleanupLookbackDays(readValue(argv[i + 1]));
+    } else if (arg.startsWith('--stats-response-path=')) {
      const value = arg.split('=', 2)[1];
      if (value) args.statsResponsePath = value;
    } else if (arg === '--stats-response-path') {
@@ -1032,6 +1032,180 @@ describe('stats server API routes', () => {
    ]);
  });

+  it('POST /api/stats/maintenance/duplicate-lines forwards the window and dry-run flag', async () => {
+    let seenOptions: unknown = null;
+    const summary = {
+      dryRun: true,
+      lookbackDays: 30,
+      scannedLines: 900,
+      burstGroups: 2,
+      removedLines: 180,
+      removedWordOccurrences: 540,
+      removedKanjiOccurrences: 120,
+      samples: [],
+    };
+    const app = createStatsApp(
+      createMockTracker({
+        cleanupDuplicateSubtitleLines: async (options: unknown) => {
+          seenOptions = options;
+          return summary;
+        },
+      }),
+    );
+
+    const res = await app.request('/api/stats/maintenance/duplicate-lines', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ dryRun: true, lookbackDays: 30 }),
+    });
+
+    assert.equal(res.status, 200);
+    assert.deepEqual(await res.json(), summary);
+    assert.deepEqual(seenOptions, { dryRun: true, lookbackDays: 30 });
+  });
+
+  it('POST /api/stats/maintenance/duplicate-lines rejects cross-origin simple requests', async () => {
+    let cleanupCalls = 0;
+    const app = createStatsApp(
+      createMockTracker({
+        cleanupDuplicateSubtitleLines: async () => {
+          cleanupCalls += 1;
+          throw new Error('cleanup must not run');
+        },
+      }),
+    );
+
+    const res = await app.request('/api/stats/maintenance/duplicate-lines', {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'text/plain',
+        Origin: 'https://attacker.example',
+      },
+      body: JSON.stringify({ dryRun: false, lookbackDays: null }),
+    });
+
+    assert.equal(res.status, 415);
+    assert.equal(cleanupCalls, 0);
+  });
+
+  it('POST /api/stats/maintenance/duplicate-lines rejects a window shorter than a day', async () => {
+    let cleanupCalls = 0;
+    const app = createStatsApp(
+      createMockTracker({
+        cleanupDuplicateSubtitleLines: async () => {
+          cleanupCalls += 1;
+          return {
+            dryRun: true,
+            lookbackDays: null,
+            scannedLines: 0,
+            burstGroups: 0,
+            removedLines: 0,
+            removedWordOccurrences: 0,
+            removedKanjiOccurrences: 0,
+            samples: [],
+          };
+        },
+      }),
+    );
+
+    const res = await app.request('/api/stats/maintenance/duplicate-lines', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ dryRun: true, lookbackDays: 0.5 }),
+    });
+
+    assert.equal(res.status, 400);
+    assert.equal(cleanupCalls, 0);
+  });
+
+  it('POST /api/stats/maintenance/duplicate-lines floors a fractional multi-day window', async () => {
+    let seenOptions: unknown = null;
+    const app = createStatsApp(
+      createMockTracker({
+        cleanupDuplicateSubtitleLines: async (options: unknown) => {
+          seenOptions = options;
+          return {
+            dryRun: true,
+            lookbackDays: 1,
+            scannedLines: 0,
+            burstGroups: 0,
+            removedLines: 0,
+            removedWordOccurrences: 0,
+            removedKanjiOccurrences: 0,
+            samples: [],
+          };
+        },
+      }),
+    );
+
+    const res = await app.request('/api/stats/maintenance/duplicate-lines', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ dryRun: true, lookbackDays: 1.5 }),
+    });
+
+    assert.equal(res.status, 200);
+    assert.deepEqual(seenOptions, { dryRun: true, lookbackDays: 1 });
+  });
+
+  it('POST /api/stats/maintenance/duplicate-lines accepts an explicit empty object for all history', async () => {
+    let seenOptions: unknown = null;
+    const app = createStatsApp(
+      createMockTracker({
+        cleanupDuplicateSubtitleLines: async (options: unknown) => {
+          seenOptions = options;
+          return {
+            dryRun: false,
+            lookbackDays: null,
+            scannedLines: 0,
+            burstGroups: 0,
+            removedLines: 0,
+            removedWordOccurrences: 0,
+            removedKanjiOccurrences: 0,
+            samples: [],
+          };
+        },
+      }),
+    );
+
+    const res = await app.request('/api/stats/maintenance/duplicate-lines', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: '{}',
+    });
+
+    assert.equal(res.status, 200);
+    assert.deepEqual(seenOptions, { dryRun: false, lookbackDays: null });
+  });
+
+  for (const malformed of [
+    { name: 'a missing body', body: undefined },
+    { name: 'malformed JSON', body: '{' },
+    { name: 'JSON null', body: 'null' },
+    { name: 'a JSON array', body: '[]' },
+  ]) {
+    it(`POST /api/stats/maintenance/duplicate-lines rejects ${malformed.name}`, async () => {
+      let cleanupCalls = 0;
+      const app = createStatsApp(
+        createMockTracker({
+          cleanupDuplicateSubtitleLines: async () => {
+            cleanupCalls += 1;
+            throw new Error('cleanup must not run');
+          },
+        }),
+      );
+
+      const res = await app.request('/api/stats/maintenance/duplicate-lines', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: malformed.body,
+      });
+
+      assert.equal(res.status, 400);
+      assert.equal(cleanupCalls, 0);
+    });
+  }
+
  it('PUT /api/stats/excluded-words rejects malformed rows', async () => {
    const app = createStatsApp(createMockTracker());

@@ -0,0 +1,195 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import {
+  assOverrideSignature,
+  assToPlainText,
+  collectAssOverrideCommands,
+  extractAssOverrideBlocks,
+  hasAssTemporalOverride,
+  isAnimatedAssEffectKind,
+  isAssTemporalCommand,
+  normalizePlainSubtitleText,
+  parseAssEffectField,
+} from './ass-text';
+
+test('assToPlainText drops vector drawing runs', () => {
+  assert.equal(
+    assToPlainText(
+      '{\\an5\\pos(730,1042)\\p1\\blur1}m 20 0 b 10 0 0 10 0 20 b 0 31 10 40 20 40 {\\p0}',
+    ),
+    '',
+  );
+});
+
+test('assToPlainText keeps text around drawing runs on the same event', () => {
+  assert.equal(
+    assToPlainText('{\\p1}m 0 0 l 10 10{\\p0}本文{\\p1}m 5 5 l 6 6{\\p0}続き'),
+    '本文続き',
+  );
+});
+
+test('assToPlainText leaves \\pos alone when no drawing mode is active', () => {
+  assert.equal(assToPlainText('{\\pos(960,1068)\\bord3}位置指定'), '位置指定');
+});
+
+test('assToPlainText does not read \\pos as a drawing tag', () => {
+  assert.equal(assToPlainText('{\\p1\\pos(1,2)}m 0 0 l 5 5'), '');
+});
+
+test('assToPlainText resolves line-break and space escapes', () => {
+  assert.equal(assToPlainText('一行目\\N二行目'), '一行目\n二行目');
+  assert.equal(assToPlainText('一行目\\n二行目'), '一行目\n二行目');
+  assert.equal(assToPlainText('一行目\\N二行目', ' '), '一行目 二行目');
+  assert.equal(assToPlainText('間\\h隔'), '間 隔');
+});
+
+test('assToPlainText matches mpv on brace and backslash sequences', () => {
+  // mpv has no `\{` / `\}` / `\\` escapes: the backslashes are literal text and the
+  // braces still open and close an override block.
+  assert.equal(assToPlainText('\\{注\\}'), '\\');
+  assert.equal(assToPlainText('\\\\N'), '\\\n');
+});
+
+test('assToPlainText renders an unclosed override block verbatim', () => {
+  // mpv shows the stray brace; guessing where the block ended can eat a whole line.
+  assert.equal(assToPlainText('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
+});
+
+test('assToPlainText is idempotent', () => {
+  const samples = [
+    '{\\an5\\p1}m 0 0 l 5 5{\\p0}本文',
+    '\\{注\\}',
+    '\\\\N',
+    '本文{\\pos(1,2)',
+    '一行目\\N二行目\\h終わり',
+  ];
+
+  for (const sample of samples) {
+    const once = assToPlainText(sample);
+    assert.equal(assToPlainText(once), once, sample);
+  }
+});
+
+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('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)');
+  assert.equal(normalizePlainSubtitleText('一行目\\N二行目'), '一行目\n二行目');
+  assert.equal(
+    normalizePlainSubtitleText('一行目\\N二行目', { collapseLineBreaks: true }),
+    '一行目 二行目',
+  );
+  assert.equal(normalizePlainSubtitleText('  余白  ', { trim: false }), '  余白  ');
+});
+
+test('normalizePlainSubtitleText is idempotent', () => {
+  for (const sample of ['一行目\\N二行目', '間\\h隔', '本文{\\pos(1,2)', '  余白  ']) {
+    const once = normalizePlainSubtitleText(sample);
+    assert.equal(normalizePlainSubtitleText(once), once, sample);
+  }
+});
+
+test('extractAssOverrideBlocks returns block contents', () => {
+  assert.deepEqual(extractAssOverrideBlocks('{\\an8}上{\\fad(200,200)}下'), [
+    '\\an8',
+    '\\fad(200,200)',
+  ]);
+  assert.deepEqual(extractAssOverrideBlocks('括弧なし'), []);
+});
+
+test('collectAssOverrideCommands captures names and arguments from blocks only', () => {
+  const commands = collectAssOverrideCommands('{\\pos(1,2)\\1c&HFFFFFF&\\kf30}歌詞');
+
+  assert.deepEqual(commands, [
+    { name: 'pos', args: '1,2', animated: false },
+    { name: '1c', args: '&HFFFFFF&', animated: false },
+    { name: 'kf', args: '30', animated: false },
+  ]);
+
+  // A `\pos(...)` sitting in visible text is not typesetting markup.
+  assert.deepEqual(collectAssOverrideCommands('\\pos(730,1042) と書いてある'), []);
+});
+
+test('collectAssOverrideCommands marks tags animated by a wrapping \\t', () => {
+  const commands = collectAssOverrideCommands('{\\clip(0,0,10,10)\\t(0,500,\\frz30)}文字');
+
+  assert.deepEqual(
+    commands.map((command) => [command.name, command.animated]),
+    [
+      ['clip', false],
+      ['t', false],
+      ['frz', true],
+    ],
+  );
+  assert.equal(hasAssTemporalOverride(commands), true);
+});
+
+test('collectAssOverrideCommands stops descending into deeply nested \\t tags', () => {
+  // Nested far past the recursion cap. Uncapped, this recurses once per level, and a
+  // pathological line (real files reach one or two levels) overflows the stack.
+  const nesting = 32;
+  const block = `{${'\\t(0,500,'.repeat(nesting)}\\frz30${')'.repeat(nesting)}}文字`;
+
+  const commands = collectAssOverrideCommands(block);
+
+  // The outer `\t` plus one per allowed recursion level, and nothing from below the cap.
+  assert.equal(commands.length, 9);
+  assert.deepEqual(new Set(commands.map((command) => command.name)), new Set(['t']));
+  assert.equal(hasAssTemporalOverride(commands), true);
+});
+
+test('hasAssTemporalOverride ignores static placement and shape tags', () => {
+  assert.equal(
+    hasAssTemporalOverride(collectAssOverrideCommands('{\\pos(1,2)\\clip(m 1 1)\\blur2}文字')),
+    false,
+  );
+  assert.equal(hasAssTemporalOverride(collectAssOverrideCommands('{\\move(1,2,3,4)}文字')), true);
+});
+
+test('isAssTemporalCommand covers only intrinsically animated tags', () => {
+  for (const command of ['t', 'move', 'k', 'kf', 'ko', 'K']) {
+    assert.equal(isAssTemporalCommand(command), true, command);
+  }
+  for (const command of ['clip', 'iclip', 'frz', 'fscx', 'blur', 'be', 'pos', 'fad']) {
+    assert.equal(isAssTemporalCommand(command), false, command);
+  }
+});
+
+test('assOverrideSignature distinguishes events by their override values', () => {
+  const first = assOverrideSignature(collectAssOverrideCommands('{\\clip(m 1 1)}歌詞'));
+  const second = assOverrideSignature(collectAssOverrideCommands('{\\clip(m 2 2)}歌詞'));
+  const repeat = assOverrideSignature(collectAssOverrideCommands('{\\clip(m 1 1)}別の行'));
+
+  assert.notEqual(first, second);
+  assert.equal(first, repeat);
+});
+
+test('parseAssEffectField classifies the event-level Effect column', () => {
+  assert.equal(parseAssEffectField(''), 'none');
+  assert.equal(parseAssEffectField('   '), 'none');
+  assert.equal(parseAssEffectField('Banner;20;1;0'), 'banner');
+  assert.equal(parseAssEffectField('Scroll up;0;0;30;10'), 'scroll');
+  assert.equal(parseAssEffectField('Scroll down;0;0;30;10'), 'scroll');
+  assert.equal(parseAssEffectField('Karaoke'), 'karaoke');
+  assert.equal(parseAssEffectField('fx-template'), 'other');
+});
+
+test('parseAssEffectField matches stock effect names exactly', () => {
+  // Custom effect names that merely start with a stock name are not stock effects.
+  assert.equal(parseAssEffectField('scrolling-credit'), 'other');
+  assert.equal(parseAssEffectField('bannerfx;1'), 'other');
+  assert.equal(parseAssEffectField('karaoke-template'), 'other');
+  assert.equal(parseAssEffectField('Scroll'), 'other');
+});
+
+test('isAnimatedAssEffectKind covers the stock animated effects only', () => {
+  assert.equal(isAnimatedAssEffectKind('karaoke'), true);
+  assert.equal(isAnimatedAssEffectKind('banner'), true);
+  assert.equal(isAnimatedAssEffectKind('scroll'), true);
+  // Typesetting groups put static template names in this column too.
+  assert.equal(isAnimatedAssEffectKind('other'), false);
+  assert.equal(isAnimatedAssEffectKind('none'), false);
+});
@@ -0,0 +1,280 @@
+/*
+ * ASS/SSA text handling, split into two deliberately distinct contracts:
+ *
+ *   assToPlainText()             raw ASS event text -> plain text. Ingestion only.
+ *   normalizePlainSubtitleText() already-decoded text -> display/lookup form.
+ *
+ * Subtitle text is decoded from ASS exactly once, at the point it enters the app: the
+ * file cue parser does it for sidecar/embedded scripts, and mpv does it for live text
+ * (`sub-text` is already run through mpv's own `ass_to_plaintext`). Everything
+ * downstream -- renderer, timing tracker, tokenizer, tokenization cache keys -- gets
+ * plain text and only normalizes whitespace, so no layer decodes the same string twice.
+ *
+ * assToPlainText mirrors mpv's `ass_to_plaintext` rather than inventing its own rules,
+ * so a cue parsed from a file reads the same as the same line arriving live:
+ *   - `{...}` override blocks are markup
+ *   - `\pN ... \p0` runs are vector paths, not dialogue
+ *   - `\N`, `\n` and `\h` are the only escapes; `\{`, `\}` and `\\` are NOT escapes,
+ *     so `\{注\}` decodes to a lone backslash exactly as mpv renders it
+ *   - an unclosed `{` is rendered verbatim instead of swallowing the rest of the line
+ * Because the decoder never emits an escape or a closed brace, running it twice is a
+ * no-op -- but downstream code should still use normalizePlainSubtitleText.
+ */
+
+/** What `\N` and `\n` become. */
+export type AssLineBreak = '\n' | ' ';
+
+// `\p<n>` with n > 0 switches libass into vector-drawing mode: everything until the
+// next `\p0` is a path (`m 20 0 b 10 0 ...`), not dialogue. The negative lookahead keeps
+// `\pos(...)` from being read as a drawing tag.
+const ASS_DRAWING_SCALE_PATTERN = /\\p(?![a-zA-Z])(\d*)/g;
+
+function readDrawingScale(block: string): number | null {
+  ASS_DRAWING_SCALE_PATTERN.lastIndex = 0;
+  let scale: number | null = null;
+  let match: RegExpExecArray | null;
+  // Drawing mode is whatever the last `\p` tag in this block set it to.
+  while ((match = ASS_DRAWING_SCALE_PATTERN.exec(block)) !== null) {
+    scale = match[1] ? Number(match[1]) : 0;
+  }
+  return scale;
+}
+
+/** Resolve `\N`, `\n` and `\h`. The only text-level escapes libass recognises. */
+function resolveWhitespaceEscapes(text: string, lineBreak: AssLineBreak): string {
+  return text.replace(/\\([Nnh])/g, (_match, escaped: string) =>
+    escaped === 'h' ? ' ' : lineBreak,
+  );
+}
+
+/** Strip `{...}` override blocks and the drawing runs they enable. */
+function stripAssMarkup(raw: string): string {
+  let out = '';
+  let cursor = 0;
+  let drawing = false;
+
+  while (cursor < raw.length) {
+    if (raw[cursor] !== '{') {
+      if (!drawing) {
+        out += raw[cursor];
+      }
+      cursor += 1;
+      continue;
+    }
+
+    const close = raw.indexOf('}', cursor + 1);
+    if (close === -1) {
+      // mpv shows an unclosed `{` and everything after it. Guessing where the block was
+      // meant to end can eat a whole line of dialogue.
+      if (!drawing) {
+        out += raw.slice(cursor);
+      }
+      break;
+    }
+
+    const scale = readDrawingScale(raw.slice(cursor, close + 1));
+    if (scale !== null) {
+      drawing = scale > 0;
+    }
+    cursor = close + 1;
+  }
+
+  return out;
+}
+
+/**
+ * Decode a raw ASS/SSA event text field. Call this once, where the text enters the app;
+ * downstream layers take the result as plain text.
+ */
+export function assToPlainText(text: string, lineBreak: AssLineBreak = '\n'): string {
+  if (!text) return '';
+  return resolveWhitespaceEscapes(stripAssMarkup(text.replace(/\r\n/g, '\n')), lineBreak);
+}
+
+export interface NormalizePlainSubtitleTextOptions {
+  /** Fold every line break into a single space. */
+  collapseLineBreaks?: boolean;
+  trim?: boolean;
+}
+
+/**
+ * Whitespace normalization for text that has already been decoded -- by mpv for live
+ * subtitles, by the cue parser for files. Override blocks and drawing runs are none of
+ * this function's business; a `{` that reaches here is literal text mpv chose to show.
+ *
+ * `\N`/`\n`/`\h` are still folded, because subtitle sources outside the ASS path (asbplayer
+ * and other websocket clients) forward them raw and the display layer has to cope.
+ */
+export function normalizePlainSubtitleText(
+  text: string,
+  options: NormalizePlainSubtitleTextOptions = {},
+): string {
+  if (!text) return '';
+  const { collapseLineBreaks = false, trim = true } = options;
+
+  let normalized = resolveWhitespaceEscapes(
+    text.replace(/\r\n/g, '\n'),
+    collapseLineBreaks ? ' ' : '\n',
+  );
+  if (collapseLineBreaks) {
+    normalized = normalized.replace(/\n/g, ' ').replace(/\s+/g, ' ');
+  }
+
+  return trim ? normalized.trim() : normalized;
+}
+
+/** The contents of each `{...}` block, without the braces. */
+export function extractAssOverrideBlocks(text: string): string[] {
+  const blocks: string[] = [];
+  let cursor = 0;
+
+  while (cursor < text.length) {
+    const open = text.indexOf('{', cursor);
+    if (open === -1) {
+      break;
+    }
+    const close = text.indexOf('}', open + 1);
+    if (close === -1) {
+      break;
+    }
+    blocks.push(text.slice(open + 1, close));
+    cursor = close + 1;
+  }
+
+  return blocks;
+}
+
+export interface AssOverrideCommand {
+  /** Tag name without the backslash, e.g. `pos`, `kf`, `1c`. */
+  name: string;
+  /** Everything the tag was given, e.g. `960,1068` for `\pos(960,1068)`. */
+  args: string;
+  /** Nested inside a `\t(...)` argument, so its value is animated over the event. */
+  animated: boolean;
+}
+
+const ASS_OVERRIDE_NAME_PATTERN = /[1-4]?[a-zA-Z]+/y;
+
+function readCommandArgs(block: string, start: number): { args: string; next: number } {
+  if (block[start] === '(') {
+    let depth = 0;
+    for (let i = start; i < block.length; i += 1) {
+      if (block[i] === '(') depth += 1;
+      else if (block[i] === ')') {
+        depth -= 1;
+        if (depth === 0) {
+          return { args: block.slice(start + 1, i), next: i + 1 };
+        }
+      }
+    }
+    return { args: block.slice(start + 1), next: block.length };
+  }
+
+  const nextTag = block.indexOf('\\', start);
+  const end = nextTag === -1 ? block.length : nextTag;
+  return { args: block.slice(start, end), next: end };
+}
+
+// `\t(...)` can wrap another `\t(...)`, and nothing in the format stops an author (or a
+// malformed file) from nesting them thousands deep. Real typesetting never goes past one
+// or two levels, so stop recursing well before the call stack is at risk.
+const MAX_ANIMATION_NESTING_DEPTH = 8;
+
+function parseOverrideBlock(
+  block: string,
+  animated: boolean,
+  into: AssOverrideCommand[],
+  depth = 0,
+): void {
+  let cursor = 0;
+
+  while (cursor < block.length) {
+    if (block[cursor] !== '\\') {
+      cursor += 1;
+      continue;
+    }
+
+    ASS_OVERRIDE_NAME_PATTERN.lastIndex = cursor + 1;
+    const nameMatch = ASS_OVERRIDE_NAME_PATTERN.exec(block);
+    if (!nameMatch) {
+      cursor += 1;
+      continue;
+    }
+
+    const name = nameMatch[0];
+    const { args, next } = readCommandArgs(block, cursor + 1 + name.length);
+    into.push({ name, args: args.trim(), animated });
+    // `\t(0,500,\frz30)` animates whatever it wraps, so record the inner tags too.
+    if (name === 't' && args.includes('\\') && depth < MAX_ANIMATION_NESTING_DEPTH) {
+      parseOverrideBlock(args, true, into, depth + 1);
+    }
+    cursor = next;
+  }
+}
+
+/**
+ * Override commands with their arguments, in source order. Only `{...}` blocks are
+ * inspected, so a `\pos(...)` sitting in visible text is never mistaken for markup.
+ */
+export function collectAssOverrideCommands(text: string): AssOverrideCommand[] {
+  const commands: AssOverrideCommand[] = [];
+  for (const block of extractAssOverrideBlocks(text)) {
+    parseOverrideBlock(block, false, commands);
+  }
+  return commands;
+}
+
+// Tags that are animated by definition: `\t` interpolates, `\move` travels, and the
+// karaoke tags advance a highlight across the event's own duration. Everything else --
+// `\pos`, `\clip`, `\frz`, `\blur`, `\fad` -- is a static value for the event, so its
+// presence says nothing about whether neighbouring events form one animation.
+const ASS_TEMPORAL_COMMANDS = new Set(['t', 'move', 'k', 'kf', 'ko', 'K']);
+
+export function isAssTemporalCommand(name: string): boolean {
+  return ASS_TEMPORAL_COMMANDS.has(name);
+}
+
+/** True when the event animates on its own, or animates a static tag through `\t(...)`. */
+export function hasAssTemporalOverride(commands: readonly AssOverrideCommand[]): boolean {
+  return commands.some((command) => command.animated || isAssTemporalCommand(command.name));
+}
+
+/**
+ * Canonical form of an event's override values, for comparing consecutive events. Two
+ * events with the same signature were typeset identically, so neither is a frame of an
+ * animation the other belongs to.
+ */
+export function assOverrideSignature(commands: readonly AssOverrideCommand[]): string {
+  return commands.map((command) => `${command.name}(${command.args})`).join('|');
+}
+
+export type AssEffectKind = 'none' | 'banner' | 'scroll' | 'karaoke' | 'other';
+
+// The stock effects, matched exactly. Typesetting groups put their own template names in
+// this column -- `scrolling-credit` is a static sign, not libass's `Scroll up` -- so a
+// prefix match would hand out animation evidence to arbitrary custom effects.
+const STOCK_ASS_EFFECTS = new Map<string, AssEffectKind>([
+  ['banner', 'banner'],
+  ['scroll up', 'scroll'],
+  ['scroll down', 'scroll'],
+  ['karaoke', 'karaoke'],
+]);
+
+/**
+ * The event-level `Effect` column. The stock values (`Banner;...`, `Scroll up;...`,
+ * `Scroll down;...`, `Karaoke`) all animate; anything else is a custom name and lands in
+ * `other`.
+ */
+export function parseAssEffectField(raw: string): AssEffectKind {
+  const value = raw.trim().toLowerCase();
+  if (!value) return 'none';
+
+  const name = value.split(';', 1)[0]!.trim();
+  return STOCK_ASS_EFFECTS.get(name) ?? 'other';
+}
+
+const ANIMATED_ASS_EFFECT_KINDS = new Set<AssEffectKind>(['banner', 'scroll', 'karaoke']);
+
+export function isAnimatedAssEffectKind(kind: AssEffectKind): boolean {
+  return ANIMATED_ASS_EFFECT_KINDS.has(kind);
+}
@@ -1414,6 +1414,353 @@ test('deleteSession ignores the currently active session and keeps new writes fl
  }
 });

+test('deleteSession yields the main event loop while delete maintenance is pending', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const deleteGate: { release?: () => void } = {};
+  let deleteRunnerCalled = false;
+  let bufferedWritesAtDeleteStart = -1;
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    const createdTracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async () => {
+          deleteRunnerCalled = true;
+          bufferedWritesAtDeleteStart = (tracker as unknown as { queue: unknown[] }).queue.length;
+          await new Promise<void>((resolve) => {
+            deleteGate.release = resolve;
+          });
+        },
+      },
+    );
+    tracker = createdTracker;
+    createdTracker.handleMediaChange('/tmp/delete-yield-first.mkv', 'Delete Yield First');
+    createdTracker.handleMediaChange('/tmp/delete-yield-active.mkv', 'Delete Yield Active');
+
+    const privateApi = createdTracker as unknown as {
+      db: DatabaseSync;
+      queue: unknown[];
+      flushNow: () => void;
+    };
+    const sessionId = (
+      privateApi.db
+        .prepare(
+          `SELECT session_id AS sessionId
+           FROM imm_sessions
+           WHERE ended_at_ms IS NOT NULL
+           ORDER BY session_id
+           LIMIT 1`,
+        )
+        .get() as { sessionId: number } | null
+    )?.sessionId;
+    assert.ok(sessionId);
+
+    const deletePromise = createdTracker.deleteSession(sessionId);
+    let timerAdvanced = false;
+    setTimeout(() => {
+      timerAdvanced = true;
+    }, 0);
+
+    await waitForCondition(() => deleteRunnerCalled);
+    assert.equal(deleteRunnerCalled, true, 'delete should be dispatched to the maintenance runner');
+    assert.equal(
+      bufferedWritesAtDeleteStart,
+      0,
+      'writes buffered before delete should flush first',
+    );
+    await waitForCondition(() => timerAdvanced);
+
+    createdTracker.recordSubtitleLine('queued during delete', 0, 1);
+    privateApi.flushNow();
+    assert.ok(privateApi.queue.length > 0, 'tracking writes should wait for delete maintenance');
+
+    assert.ok(deleteGate.release);
+    deleteGate.release();
+    await deletePromise;
+  } finally {
+    deleteGate.release?.();
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('delete maintenance flushes the entire write queue before locking writes', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const deleteGate: { release?: () => void } = {};
+  let queuedWritesAtDeleteStart = -1;
+  let writeLockedAtDeleteStart = false;
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async () => {
+          const privateApi = tracker as unknown as {
+            queue: unknown[];
+            writeLock: { locked: boolean };
+          };
+          queuedWritesAtDeleteStart = privateApi.queue.length;
+          writeLockedAtDeleteStart = privateApi.writeLock.locked;
+          await new Promise<void>((resolve) => {
+            deleteGate.release = resolve;
+          });
+        },
+      },
+    );
+
+    const privateApi = tracker as unknown as {
+      batchSize: number;
+      flushNow: () => void;
+      queue: unknown[];
+    };
+    privateApi.batchSize = 1;
+    privateApi.queue.push({}, {}, {});
+    privateApi.flushNow = () => {
+      privateApi.queue.shift();
+    };
+
+    const deletePromise = tracker.deleteSession(101);
+    await waitForCondition(() => deleteGate.release !== undefined);
+
+    assert.equal(queuedWritesAtDeleteStart, 0);
+    assert.equal(writeLockedAtDeleteStart, true);
+
+    deleteGate.release?.();
+    await deletePromise;
+  } finally {
+    deleteGate.release?.();
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('delete maintenance tasks stay serialized under concurrent requests', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const releases: Array<() => void> = [];
+  let activeTasks = 0;
+  let maxActiveTasks = 0;
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async () => {
+          activeTasks += 1;
+          maxActiveTasks = Math.max(maxActiveTasks, activeTasks);
+          await new Promise<void>((resolve) => {
+            releases.push(resolve);
+          });
+          activeTasks -= 1;
+        },
+      },
+    );
+
+    const firstDelete = tracker.deleteSession(101);
+    await waitForCondition(() => releases.length === 1);
+    assert.equal(maxActiveTasks, 1);
+
+    const secondDelete = tracker.deleteSession(102);
+
+    releases[0]?.();
+    await waitForCondition(() => releases.length === 2);
+    assert.equal(maxActiveTasks, 1);
+
+    releases[1]?.();
+    await Promise.all([firstDelete, secondDelete]);
+  } finally {
+    for (const release of releases) release();
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('concurrent delete requests share one maintenance worker batch', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const tasks: unknown[] = [];
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async (_path, task) => {
+          tasks.push(task);
+        },
+      },
+    );
+
+    const firstDelete = tracker.deleteSession(201);
+    const secondDelete = tracker.deleteSessions([202, 203]);
+    const thirdDelete = tracker.deleteVideo(204);
+    await Promise.all([firstDelete, secondDelete, thirdDelete]);
+
+    assert.equal(tasks.length, 1, 'concurrent deletes should use one maintenance pass');
+    assert.deepEqual(tasks[0], {
+      kind: 'batch',
+      tasks: [
+        { kind: 'session', sessionId: 201 },
+        { kind: 'sessions', sessionIds: [202, 203] },
+        { kind: 'video', videoId: 204 },
+      ],
+    });
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('destroy rejects delete requests waiting behind active maintenance', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  let releaseFirstTask: () => void = () => {};
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    let markFirstTaskStarted: () => void = () => {};
+    const firstTaskStarted = new Promise<void>((resolve) => {
+      markFirstTaskStarted = resolve;
+    });
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async () => {
+          markFirstTaskStarted();
+          await new Promise<void>((resolve) => {
+            releaseFirstTask = resolve;
+          });
+        },
+      },
+    );
+
+    const firstDelete = tracker.deleteSession(301);
+    await firstTaskStarted;
+    const queuedDelete = tracker.deleteSession(302);
+    tracker.destroy();
+
+    const queuedOutcome = await Promise.race([
+      queuedDelete.then(
+        () => 'resolved',
+        (error: unknown) =>
+          error instanceof Error && /shutting down/.test(error.message)
+            ? 'rejected'
+            : 'wrong-error',
+      ),
+      new Promise<'pending'>((resolve) => setTimeout(() => resolve('pending'), 25)),
+    ]);
+    assert.equal(queuedOutcome, 'rejected');
+    releaseFirstTask();
+    await firstDelete;
+  } finally {
+    releaseFirstTask();
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('delete requested after destroy rejects without running maintenance', async () => {
+  const dbPath = makeDbPath();
+  let maintenanceCalls = 0;
+  const Ctor = await loadTrackerCtor();
+  const tracker = new Ctor(
+    { dbPath },
+    {
+      runDeleteMaintenanceTask: async () => {
+        maintenanceCalls += 1;
+      },
+    },
+  );
+
+  tracker.destroy();
+
+  await assert.rejects(tracker.deleteSession(303), /shutting down/);
+  assert.equal(maintenanceCalls, 0);
+  cleanupDbPath(dbPath);
+});
+
+test('deleteSessions skips maintenance when no sessions are deletable', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const tasks: unknown[] = [];
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async (_path, task) => {
+          tasks.push(task);
+        },
+      },
+    );
+
+    await tracker.deleteSessions([]);
+
+    assert.deepEqual(tasks, []);
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('queued video delete is skipped when that video becomes active before dispatch', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+  const tasks: Array<{ kind: string }> = [];
+  let releaseFirstTask: () => void = () => {};
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    const createdTracker = new Ctor(
+      { dbPath },
+      {
+        runDeleteMaintenanceTask: async (_path, task) => {
+          tasks.push(task);
+          if (tasks.length === 1) {
+            await new Promise<void>((resolve) => {
+              releaseFirstTask = resolve;
+            });
+          }
+        },
+      },
+    );
+    tracker = createdTracker;
+    createdTracker.handleMediaChange('/tmp/delete-race-target.mkv', 'Delete Race Target');
+    createdTracker.handleMediaChange('/tmp/delete-race-other.mkv', 'Delete Race Other');
+
+    const privateApi = createdTracker as unknown as { db: DatabaseSync };
+    const targetVideoId = (
+      privateApi.db
+        .prepare(`SELECT video_id AS videoId FROM imm_videos WHERE video_key LIKE '%target.mkv'`)
+        .get() as { videoId: number } | null
+    )?.videoId;
+    assert.ok(targetVideoId);
+
+    const firstDelete = createdTracker.deleteSession(999_001);
+    await waitForCondition(() => tasks.length === 1);
+
+    const queuedVideoDelete = createdTracker.deleteVideo(targetVideoId);
+    createdTracker.handleMediaChange('/tmp/delete-race-target.mkv', 'Delete Race Target');
+    releaseFirstTask();
+    await Promise.all([firstDelete, queuedVideoDelete]);
+
+    assert.deepEqual(
+      tasks.map((task) => task.kind),
+      ['session'],
+    );
+  } finally {
+    releaseFirstTask();
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
 test('deleteVideo ignores the currently active video and keeps new writes flushable', async () => {
  const dbPath = makeDbPath();
  let tracker: ImmersionTrackerService | null = null;
@@ -83,14 +83,20 @@ import {
 } from './immersion-tracker/query-library';
 import {
  cleanupVocabularyStats,
-  deleteAnime as deleteAnimeQuery,
-  deleteSession as deleteSessionQuery,
-  deleteSessions as deleteSessionsQuery,
-  deleteVideo as deleteVideoQuery,
  getVideoDurationMs,
  markVideoWatched,
  upsertCoverArt,
 } from './immersion-tracker/query-maintenance';
+import {
+  DeleteMaintenanceWorkerRuntime,
+  type RunDeleteMaintenanceTask,
+} from './immersion-tracker/delete-maintenance-worker-runtime';
+import { DeleteMaintenanceScheduler } from './immersion-tracker/delete-maintenance-scheduler';
+import {
+  cleanupDuplicateSubtitleLines,
+  type DuplicateSubtitleLineCleanupOptions,
+  type DuplicateSubtitleLineCleanupSummary,
+} from './immersion-tracker/duplicate-line-cleanup';
 import { repairJellyfinStreamVideoLinks } from './immersion-tracker/jellyfin-link-repair';
 import {
  repairLegacySeasonlessAnimeRows,
@@ -182,6 +188,7 @@ const YOUTUBE_SCREENSHOT_MAX_SECONDS = 120;
 const YOUTUBE_OEMBED_ENDPOINT = 'https://www.youtube.com/oembed';
 const YOUTUBE_ID_PATTERN = /^[A-Za-z0-9_-]{6,}$/;
 const YOUTUBE_METADATA_REFRESH_MS = 24 * 60 * 60 * 1000;
+const DELETE_MAINTENANCE_BATCH_WINDOW_MS = 10;

 function isValidYouTubeVideoId(value: string | null): boolean {
  return Boolean(value && YOUTUBE_ID_PATTERN.test(value));
@@ -385,6 +392,8 @@ export class ImmersionTrackerService {
  private readonly vacuumIntervalMs: number;
  private readonly dbPath: string;
  private readonly writeLock = { locked: false };
+  private readonly destroyDeleteMaintenanceRunner: () => void;
+  private readonly deleteMaintenanceScheduler: DeleteMaintenanceScheduler;
  private flushTimer: ReturnType<typeof setTimeout> | null = null;
  private maintenanceTimer: ReturnType<typeof setInterval> | null = null;
  private flushScheduled = false;
@@ -406,9 +415,38 @@ export class ImmersionTrackerService {
    | ((row: LegacyVocabularyPosRow) => Promise<LegacyVocabularyPosResolution | null>)
    | undefined;

-  constructor(options: ImmersionTrackerOptions) {
+  constructor(
+    options: ImmersionTrackerOptions,
+    dependencies: {
+      runDeleteMaintenanceTask?: RunDeleteMaintenanceTask;
+      destroyDeleteMaintenanceRunner?: () => void;
+    } = {},
+  ) {
    this.dbPath = options.dbPath;
    this.resolveLegacyVocabularyPos = options.resolveLegacyVocabularyPos;
+    let runDeleteMaintenanceTask: RunDeleteMaintenanceTask;
+    if (dependencies.runDeleteMaintenanceTask) {
+      runDeleteMaintenanceTask = dependencies.runDeleteMaintenanceTask;
+      this.destroyDeleteMaintenanceRunner =
+        dependencies.destroyDeleteMaintenanceRunner ?? (() => {});
+    } else {
+      const deleteMaintenanceRuntime = new DeleteMaintenanceWorkerRuntime();
+      runDeleteMaintenanceTask = (dbPath, task) => deleteMaintenanceRuntime.run(dbPath, task);
+      this.destroyDeleteMaintenanceRunner = () => deleteMaintenanceRuntime.destroy();
+    }
+    this.deleteMaintenanceScheduler = new DeleteMaintenanceScheduler({
+      batchWindowMs: DELETE_MAINTENANCE_BATCH_WINDOW_MS,
+      runTask: (task) => runDeleteMaintenanceTask(this.dbPath, task),
+      onBusy: () => {
+        this.flushTelemetry(true);
+        while (this.queue.length > 0) this.flushNow();
+        this.writeLock.locked = true;
+      },
+      onIdle: () => {
+        this.writeLock.locked = false;
+        if (!this.isDestroyed && this.queue.length > 0) this.scheduleFlush(0);
+      },
+    });
    const parentDir = path.dirname(this.dbPath);
    if (!fs.existsSync(parentDir)) {
      fs.mkdirSync(parentDir, { recursive: true });
@@ -512,6 +550,8 @@ export class ImmersionTrackerService {
    }
    this.finalizeActiveSession();
    this.isDestroyed = true;
+    this.deleteMaintenanceScheduler.destroy();
+    this.destroyDeleteMaintenanceRunner();
    this.db.close();
  }

@@ -595,6 +635,18 @@ export class ImmersionTrackerService {
    });
  }

+  /**
+   * Collapse animation bursts that earlier versions recorded frame by frame. The whole
+   * queue is drained first so a burst still waiting to be written is scanned as stored
+   * rows rather than surviving the cleanup and landing a moment after it.
+   */
+  async cleanupDuplicateSubtitleLines(
+    options: DuplicateSubtitleLineCleanupOptions = {},
+  ): Promise<DuplicateSubtitleLineCleanupSummary> {
+    this.drainQueue();
+    return cleanupDuplicateSubtitleLines(this.db, options);
+  }
+
  async rebuildLifetimeSummaries(): Promise<LifetimeRebuildSummary> {
    this.flushTelemetry(true);
    this.flushNow();
@@ -709,51 +761,66 @@ export class ImmersionTrackerService {
      this.logger.warn(`Ignoring delete request for active immersion session ${sessionId}`);
      return;
    }
-    deleteSessionQuery(this.db, sessionId);
+    await this.enqueueDeleteMaintenanceTask(() => ({ kind: 'session', sessionId }));
  }

  async deleteSessions(sessionIds: number[]): Promise<void> {
-    const activeSessionId = this.sessionState?.sessionId;
-    const deletableSessionIds =
-      activeSessionId === undefined
-        ? sessionIds
-        : sessionIds.filter((sessionId) => sessionId !== activeSessionId);
-    if (deletableSessionIds.length !== sessionIds.length) {
-      this.logger.warn(
-        `Ignoring bulk delete request for active immersion session ${activeSessionId}`,
-      );
-    }
-    deleteSessionsQuery(this.db, deletableSessionIds);
+    await this.enqueueDeleteMaintenanceTask(() => {
+      const activeSessionId = this.sessionState?.sessionId;
+      const deletableSessionIds =
+        activeSessionId === undefined
+          ? sessionIds
+          : sessionIds.filter((sessionId) => sessionId !== activeSessionId);
+      if (deletableSessionIds.length !== sessionIds.length) {
+        this.logger.warn(
+          `Ignoring bulk delete request for active immersion session ${activeSessionId}`,
+        );
+      }
+      if (deletableSessionIds.length === 0) return null;
+      return { kind: 'sessions', sessionIds: deletableSessionIds };
+    });
  }

  async deleteVideo(videoId: number): Promise<void> {
-    if (this.sessionState?.videoId === videoId) {
-      this.logger.warn(`Ignoring delete request for active immersion video ${videoId}`);
-      return;
-    }
-    deleteVideoQuery(this.db, videoId);
+    await this.enqueueDeleteMaintenanceTask(() => {
+      if (this.sessionState?.videoId === videoId) {
+        this.logger.warn(`Ignoring delete request for active immersion video ${videoId}`);
+        return null;
+      }
+      return { kind: 'video', videoId };
+    });
  }

  async deleteAnime(animeId: number): Promise<void> {
-    // The active video's anime link is assigned asynchronously after the title
-    // is parsed, so a guard reading imm_videos too early sees a null and lets
-    // the delete through — then the late update recreates the anime row.
-    const pendingVideoId = this.sessionState?.videoId;
-    if (pendingVideoId !== undefined) {
-      await this.pendingAnimeMetadataUpdates.get(pendingVideoId);
-    }
-
-    const activeVideoId = this.sessionState?.videoId;
-    if (activeVideoId !== undefined) {
-      const activeAnime = this.db
-        .prepare('SELECT anime_id FROM imm_videos WHERE video_id = ?')
-        .get(activeVideoId) as { anime_id: number | null } | null;
-      if (activeAnime?.anime_id === animeId) {
-        this.logger.warn(`Ignoring delete request for active immersion anime ${animeId}`);
-        return;
+    await this.enqueueDeleteMaintenanceTask(async () => {
+      // Resolve this at dispatch time because another queued delete can leave
+      // enough time for playback to switch to an episode of this anime.
+      const pendingVideoId = this.sessionState?.videoId;
+      if (pendingVideoId !== undefined) {
+        await this.pendingAnimeMetadataUpdates.get(pendingVideoId);
      }
+
+      const activeVideoId = this.sessionState?.videoId;
+      if (activeVideoId !== undefined) {
+        const activeAnime = this.db
+          .prepare('SELECT anime_id FROM imm_videos WHERE video_id = ?')
+          .get(activeVideoId) as { anime_id: number | null } | null;
+        if (activeAnime?.anime_id === animeId) {
+          this.logger.warn(`Ignoring delete request for active immersion anime ${animeId}`);
+          return null;
+        }
+      }
+      return { kind: 'anime', animeId };
+    });
+  }
+
+  private enqueueDeleteMaintenanceTask(
+    resolveTask: Parameters<DeleteMaintenanceScheduler['enqueue']>[0],
+  ): Promise<void> {
+    if (this.isDestroyed) {
+      return Promise.reject(new Error('Immersion tracker is shutting down'));
    }
-    deleteAnimeQuery(this.db, animeId);
+    return this.deleteMaintenanceScheduler.enqueue(resolveTask);
  }

  async reassignAnimeAnilist(
@@ -1799,6 +1866,24 @@ export class ImmersionTrackerService {
    }
  }

+  /**
+   * Write out everything queued, not just the next batch.
+   *
+   * `flushNow` writes at most `batchSize` entries and does nothing at all while the write
+   * lock is held, so a maintenance pass that runs straight after it can still be reading
+   * a database that is missing rows. Each pass has to shrink the queue to continue: a
+   * failed flush puts its batch back, and looping on that would never finish.
+   */
+  private drainQueue(): void {
+    while (this.queue.length > 0) {
+      const pendingBefore = this.queue.length;
+      this.flushNow();
+      if (this.queue.length >= pendingBefore) {
+        return;
+      }
+    }
+  }
+
  private flushSingle(write: QueuedWrite): void {
    executeQueuedWrite(write, this.preparedStatements);
  }
@@ -1811,7 +1896,7 @@ export class ImmersionTrackerService {
  }

  private runMaintenance(): void {
-    if (this.isDestroyed) return;
+    if (this.isDestroyed || this.writeLock.locked) return;
    try {
      this.flushTelemetry(true);
      this.flushNow();
@@ -0,0 +1,467 @@
+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';
+import { ensureSchema } from '../storage.js';
+import { cleanupDuplicateSubtitleLines } from '../duplicate-line-cleanup.js';
+
+const DAY_MS = 86_400_000;
+const BASE_MS = 1_700_000_000_000;
+const WORD_ID = 1;
+
+interface SeedLine {
+  session: number;
+  text: string;
+  startMs: number;
+  endMs: number;
+  /** Recording wall-clock, i.e. what the lookback window filters on. */
+  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(`
+    INSERT INTO imm_anime(anime_id, normalized_title_key, canonical_title, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'show', 'Show', ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_videos(video_id, video_key, anime_id, canonical_title, source_type, watched, duration_ms, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'v1', 1, 'Ep 1', 1, 1, 1440000, ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_sessions(session_id, session_uuid, video_id, started_at_ms, ended_at_ms, status, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 's1', 1, '${BASE_MS}', '${BASE_MS + 1000}', 2, ${BASE_MS}, ${BASE_MS}),
+             (2, 's2', 1, '${BASE_MS + DAY_MS}', '${BASE_MS + DAY_MS + 1000}', 2, ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_words(id, headword, word, reading, part_of_speech, pos1, first_seen, last_seen, frequency)
+      VALUES (${WORD_ID}, '飛び上がる', '飛び上がる', '', 'verb', '動詞', ${Math.floor(BASE_MS / 1000)}, ${Math.floor(BASE_MS / 1000)}, 0);
+  `);
+
+  const insertLine = db.prepare(
+    `INSERT INTO imm_subtitle_lines(
+       line_id, session_id, video_id, anime_id, line_index,
+       segment_start_ms, segment_end_ms, text, CREATED_DATE, LAST_UPDATE_DATE)
+     VALUES (?, ?, 1, 1, ?, ?, ?, ?, ?, ?)`,
+  );
+  const insertOccurrence = db.prepare(
+    `INSERT INTO imm_word_line_occurrences(line_id, word_id, occurrence_count, seen_ms)
+     VALUES (?, ?, 1, ?)`,
+  );
+
+  lines.forEach((line, index) => {
+    const lineId = index + 1;
+    const lineIndex = index + 1;
+    const createdMs = line.createdMs ?? BASE_MS;
+    insertLine.run(
+      lineId,
+      line.session,
+      lineIndex,
+      line.startMs,
+      line.endMs,
+      line.text,
+      createdMs,
+      createdMs,
+    );
+    insertOccurrence.run(lineId, WORD_ID, createdMs);
+  });
+
+  db.exec(`
+    UPDATE imm_words SET frequency = (
+      SELECT COALESCE(SUM(o.occurrence_count), 0)
+      FROM imm_word_line_occurrences o WHERE o.word_id = imm_words.id
+    )
+  `);
+}
+
+function createDb(lines: SeedLine[]): { db: DatabaseSync; dbPath: string } {
+  const dbPath = makeDbPath();
+  const db = new Database(dbPath);
+  ensureSchema(db);
+  seed(db, lines);
+  return { db, dbPath };
+}
+
+/** A typeset line mpv reported once per animation frame. */
+function karaokeFrames(
+  session: number,
+  text: string,
+  startMs: number,
+  frames: number,
+  frameMs: number,
+): SeedLine[] {
+  return Array.from({ length: frames }, (_, index) => ({
+    session,
+    text,
+    startMs: startMs + index * frameMs,
+    endMs: startMs + (index + 1) * frameMs,
+  }));
+}
+
+function countLines(db: DatabaseSync): number {
+  return (db.prepare('SELECT COUNT(*) AS total FROM imm_subtitle_lines').get() as { total: number })
+    .total;
+}
+
+function wordFrequency(db: DatabaseSync): number {
+  const row = db.prepare('SELECT frequency FROM imm_words WHERE id = ?').get(WORD_ID) as {
+    frequency: number;
+  } | null;
+  return row?.frequency ?? 0;
+}
+
+test('a karaoke burst collapses to one line and gives back its word counts', () => {
+  const { db, dbPath } = 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 },
+  ]);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 1);
+    assert.equal(summary.removedLines, 39);
+    assert.equal(summary.removedWordOccurrences, 39);
+    assert.equal(countLines(db), 2);
+    assert.equal(wordFrequency(db), 2);
+
+    // The surviving line covers the whole run, the way the parsed cue would.
+    const kept = db
+      .prepare(
+        'SELECT segment_start_ms AS startMs, segment_end_ms AS endMs FROM imm_subtitle_lines WHERE line_id = 1',
+      )
+      .get() as { startMs: number; endMs: number };
+    assert.equal(kept.startMs, 10_000);
+    assert.equal(kept.endMs, 10_000 + 40 * 40);
+
+    assert.equal(summary.samples.length, 1);
+    assert.equal(summary.samples[0]!.text, '飛び上がる');
+    assert.equal(summary.samples[0]!.frames, 40);
+    assert.equal(summary.samples[0]!.videoTitle, 'Ep 1');
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('ordinary repeated dialogue survives', () => {
+  // Six contiguous `飛び上がる`, each held for a normal beat rather than a frame.
+  const lines = Array.from({ length: 6 }, (_, index) => ({
+    session: 1,
+    text: '飛び上がる',
+    startMs: 5_000 + index * 800,
+    endMs: 5_000 + (index + 1) * 800,
+  }));
+  const { db, dbPath } = createDb(lines);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 0);
+    assert.equal(summary.removedLines, 0);
+    assert.equal(countLines(db), 6);
+    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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 1);
+    assert.equal(summary.removedLines, 5);
+    assert.equal(countLines(db), 1);
+    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([
+    ...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 },
+  ]);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 1);
+    assert.equal(summary.removedLines, 8);
+    assert.equal(countLines(db), 1);
+    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([
+    ...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 },
+  ]);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 0);
+    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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 0);
+    assert.equal(countLines(db), 6);
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 1);
+    assert.equal(summary.removedLines, 3);
+    assert.equal(countLines(db), 1);
+    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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 0);
+    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));
+
+  try {
+    const preview = cleanupDuplicateSubtitleLines(db, { dryRun: true });
+    assert.equal(preview.burstGroups, 1);
+
+    const summary = cleanupDuplicateSubtitleLines(db, { minRunLength: 6 });
+    assert.equal(summary.burstGroups, 0);
+    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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db, { maxFrameSeconds: 0.2 });
+
+    assert.equal(summary.burstGroups, 0);
+    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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db, { maxFrameSeconds: Infinity });
+
+    assert.equal(summary.burstGroups, 0);
+    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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db, { sampleLimit: 0 });
+
+    assert.equal(summary.removedLines, 39);
+    assert.deepEqual(summary.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));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 0);
+    assert.equal(countLines(db), 3);
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('interleaved dual-line karaoke collapses each line to one row', () => {
+  // Kanji and romaji lines frame-flipped together, the way fansub OPs are typeset. The
+  // rows arrive interleaved in time order; each text must still chain into its own run.
+  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);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 2);
+    assert.equal(summary.removedLines, 38);
+    assert.equal(countLines(db), 2);
+    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([
+    ...karaokeFrames(1, '飛び上がる', 10_000, 6, 40),
+    ...karaokeFrames(2, '飛び上がる', 10_000, 6, 40),
+  ]);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 2);
+    assert.equal(summary.removedLines, 10);
+    // One surviving line per session, not one across both.
+    assert.equal(countLines(db), 2);
+    assert.equal(wordFrequency(db), 2);
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('a gap between runs splits them', () => {
+  const { db, dbPath } = createDb([
+    ...karaokeFrames(1, '飛び上がる', 10_000, 6, 40),
+    ...karaokeFrames(1, '飛び上がる', 60_000, 6, 40),
+  ]);
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 2);
+    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));
+
+  try {
+    const preview = cleanupDuplicateSubtitleLines(db, { dryRun: true });
+
+    assert.equal(preview.dryRun, true);
+    assert.equal(preview.removedLines, 39);
+    assert.equal(countLines(db), 40);
+    assert.equal(wordFrequency(db), 40);
+
+    const applied = cleanupDuplicateSubtitleLines(db);
+    assert.equal(applied.removedLines, preview.removedLines);
+    assert.equal(applied.removedWordOccurrences, preview.removedWordOccurrences);
+    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([
+    ...karaokeFrames(1, '飛び上がる', 10_000, 6, 40).map((line) => ({
+      ...line,
+      createdMs: oldMs,
+    })),
+    ...karaokeFrames(2, '飛び上がる', 10_000, 6, 40).map((line) => ({
+      ...line,
+      createdMs: recentMs,
+    })),
+  ]);
+
+  globalThis.__subminerTestNowMs = BASE_MS;
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db, { lookbackDays: 30 });
+
+    assert.equal(summary.lookbackDays, 30);
+    assert.equal(summary.scannedLines, 6);
+    assert.equal(summary.burstGroups, 1);
+    assert.equal(summary.removedLines, 5);
+    // Six untouched old frames plus the one surviving recent line.
+    assert.equal(countLines(db), 7);
+    assert.equal(wordFrequency(db), 7);
+  } finally {
+    globalThis.__subminerTestNowMs = undefined;
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
@@ -50,6 +50,7 @@ import {
  updateAnimeAnilistInfo,
  upsertCoverArt,
 } from '../query-maintenance.js';
+import { deleteMaintenanceBatch } from '../query-delete-maintenance.js';
 import { getLocalEpochDay } from '../query-shared.js';
 import { EVENT_CARD_MINED, EVENT_SUBTITLE_LINE, SOURCE_TYPE_LOCAL } from '../types.js';

@@ -985,3 +986,197 @@ test('split maintenance helpers delete multiple sessions and whole videos with d
    cleanupDbPath(dbPath);
  }
 });
+
+test('delete maintenance batch preserves retained data across overlapping session, video, and anime targets', () => {
+  const { db, dbPath, stmts } = createDb();
+
+  try {
+    const retainedAnimeId = getOrCreateAnimeRecord(db, {
+      parsedTitle: 'Retained Anime',
+      canonicalTitle: 'Retained Anime',
+      anilistId: null,
+      titleRomaji: null,
+      titleEnglish: null,
+      titleNative: null,
+      metadataJson: null,
+    });
+    const deletedAnimeId = getOrCreateAnimeRecord(db, {
+      parsedTitle: 'Deleted Anime',
+      canonicalTitle: 'Deleted Anime',
+      anilistId: null,
+      titleRomaji: null,
+      titleEnglish: null,
+      titleNative: null,
+      metadataJson: null,
+    });
+    const retainedVideoId = getOrCreateVideoRecord(db, 'local:/tmp/batch-retain.mkv', {
+      canonicalTitle: 'Batch Retain',
+      sourcePath: '/tmp/batch-retain.mkv',
+      sourceUrl: null,
+      sourceType: SOURCE_TYPE_LOCAL,
+    });
+    const deletedVideoId = getOrCreateVideoRecord(db, 'local:/tmp/batch-video.mkv', {
+      canonicalTitle: 'Batch Video',
+      sourcePath: '/tmp/batch-video.mkv',
+      sourceUrl: null,
+      sourceType: SOURCE_TYPE_LOCAL,
+    });
+    const animeVideoId = getOrCreateVideoRecord(db, 'local:/tmp/batch-anime.mkv', {
+      canonicalTitle: 'Batch Anime',
+      sourcePath: '/tmp/batch-anime.mkv',
+      sourceUrl: null,
+      sourceType: SOURCE_TYPE_LOCAL,
+    });
+    for (const [videoId, animeId, episode] of [
+      [retainedVideoId, retainedAnimeId, 1],
+      [deletedVideoId, retainedAnimeId, 2],
+      [animeVideoId, deletedAnimeId, 1],
+    ] as const) {
+      linkVideoToAnimeRecord(db, videoId, {
+        animeId,
+        parsedBasename: `batch-${episode}.mkv`,
+        parsedTitle: animeId === retainedAnimeId ? 'Retained Anime' : 'Deleted Anime',
+        parsedSeason: 1,
+        parsedEpisode: episode,
+        parserSource: 'test',
+        parserConfidence: 1,
+        parseMetadataJson: null,
+      });
+    }
+
+    const startedAtMs = 1_700_000_000_000;
+    const deletedSessionId = startSessionRecord(db, retainedVideoId, startedAtMs).sessionId;
+    const retainedSessionId = startSessionRecord(
+      db,
+      retainedVideoId,
+      startedAtMs + 1_000,
+    ).sessionId;
+    const videoSessionId = startSessionRecord(db, deletedVideoId, startedAtMs + 2_000).sessionId;
+    const animeSessionId = startSessionRecord(db, animeVideoId, startedAtMs + 3_000).sessionId;
+    for (const [sessionId, sessionStartedAtMs] of [
+      [deletedSessionId, startedAtMs],
+      [retainedSessionId, startedAtMs + 1_000],
+      [videoSessionId, startedAtMs + 2_000],
+      [animeSessionId, startedAtMs + 3_000],
+    ] as const) {
+      finalizeSessionMetrics(db, sessionId, sessionStartedAtMs);
+    }
+
+    for (const [index, sessionId, videoId, animeId] of [
+      [1, deletedSessionId, retainedVideoId, retainedAnimeId],
+      [2, retainedSessionId, retainedVideoId, retainedAnimeId],
+      [3, videoSessionId, deletedVideoId, retainedAnimeId],
+      [4, animeSessionId, animeVideoId, deletedAnimeId],
+    ] as const) {
+      insertWordOccurrence(db, stmts, {
+        sessionId,
+        videoId,
+        animeId,
+        lineIndex: index,
+        text: '猫日',
+        word: { headword: '猫', word: '猫', reading: 'ねこ' },
+      });
+      insertKanjiOccurrence(db, stmts, {
+        sessionId,
+        videoId,
+        animeId,
+        lineIndex: index + 10,
+        text: '猫日',
+        kanji: '日',
+      });
+    }
+
+    const rollupDay = getLocalEpochDay(db, startedAtMs);
+    const rollupMonth = (
+      db
+        .prepare(
+          `SELECT CAST(strftime('%Y%m', CAST(? AS REAL) / 1000, 'unixepoch', 'localtime') AS INTEGER) AS rollupMonth`,
+        )
+        .get(startedAtMs) as { rollupMonth: number }
+    ).rollupMonth;
+    for (const videoId of [retainedVideoId, deletedVideoId, animeVideoId]) {
+      db.prepare(
+        `INSERT INTO imm_daily_rollups (
+          rollup_day, video_id, total_sessions, total_active_min, total_lines_seen,
+          total_tokens_seen, total_cards, CREATED_DATE, LAST_UPDATE_DATE
+        ) VALUES (?, ?, 99, 99, 99, 99, 99, ?, ?)`,
+      ).run(rollupDay, videoId, startedAtMs, startedAtMs);
+      db.prepare(
+        `INSERT INTO imm_monthly_rollups (
+          rollup_month, video_id, total_sessions, total_active_min, total_lines_seen,
+          total_tokens_seen, total_cards, CREATED_DATE, LAST_UPDATE_DATE
+        ) VALUES (?, ?, 99, 99, 99, 99, 99, ?, ?)`,
+      ).run(rollupMonth, videoId, startedAtMs, startedAtMs);
+    }
+
+    deleteMaintenanceBatch(db, [
+      { kind: 'session', sessionId: deletedSessionId },
+      { kind: 'session', sessionId: videoSessionId },
+      { kind: 'video', videoId: deletedVideoId },
+      { kind: 'video', videoId: animeVideoId },
+      { kind: 'anime', animeId: deletedAnimeId },
+    ]);
+
+    assert.deepEqual(db.prepare('SELECT session_id FROM imm_sessions').all(), [
+      { session_id: retainedSessionId },
+    ]);
+    assert.deepEqual(db.prepare('SELECT video_id FROM imm_videos').all(), [
+      { video_id: retainedVideoId },
+    ]);
+    assert.deepEqual(db.prepare('SELECT anime_id FROM imm_anime').all(), [
+      { anime_id: retainedAnimeId },
+    ]);
+    assert.equal(
+      (
+        db.prepare(`SELECT frequency FROM imm_words WHERE headword = '猫'`).get() as {
+          frequency: number;
+        }
+      ).frequency,
+      1,
+    );
+    assert.equal(
+      (
+        db.prepare(`SELECT frequency FROM imm_kanji WHERE kanji = '日'`).get() as {
+          frequency: number;
+        }
+      ).frequency,
+      1,
+    );
+    assert.deepEqual(
+      db.prepare('SELECT video_id, total_sessions FROM imm_daily_rollups').all() as Array<{
+        video_id: number;
+        total_sessions: number;
+      }>,
+      [{ video_id: retainedVideoId, total_sessions: 1 }],
+    );
+    assert.deepEqual(
+      db.prepare('SELECT video_id, total_sessions FROM imm_monthly_rollups').all() as Array<{
+        video_id: number;
+        total_sessions: number;
+      }>,
+      [{ video_id: retainedVideoId, total_sessions: 1 }],
+    );
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('delete maintenance batch chunks id lists below the SQLite variable limit', () => {
+  const { db, dbPath } = createDb();
+
+  try {
+    const ids = Array.from({ length: 32_767 }, (_, index) => index + 1);
+
+    assert.doesNotThrow(() => {
+      deleteMaintenanceBatch(db, [
+        { kind: 'sessions', sessionIds: ids },
+        ...ids.map((videoId) => ({ kind: 'video' as const, videoId })),
+        ...ids.map((animeId) => ({ kind: 'anime' as const, animeId })),
+      ]);
+    });
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
@@ -0,0 +1,160 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { DeleteMaintenanceScheduler } from './delete-maintenance-scheduler';
+import type { DeleteMaintenanceTask } from './delete-maintenance';
+
+test('scheduler batches same-turn requests and balances busy state', async () => {
+  const tasks: DeleteMaintenanceTask[] = [];
+  const states: string[] = [];
+  const scheduler = new DeleteMaintenanceScheduler({
+    batchWindowMs: 0,
+    runTask: async (task) => {
+      tasks.push(task);
+    },
+    onBusy: () => states.push('busy'),
+    onIdle: () => states.push('idle'),
+  });
+
+  const first = scheduler.enqueue(() => ({ kind: 'session', sessionId: 1 }));
+  const second = scheduler.enqueue(() => ({ kind: 'sessions', sessionIds: [2, 3] }));
+  const third = scheduler.enqueue(() => null);
+  await Promise.all([first, second, third]);
+
+  assert.deepEqual(tasks, [
+    {
+      kind: 'batch',
+      tasks: [
+        { kind: 'session', sessionId: 1 },
+        { kind: 'sessions', sessionIds: [2, 3] },
+      ],
+    },
+  ]);
+  assert.deepEqual(states, ['busy', 'idle']);
+});
+
+test('scheduler rejects enqueue after destruction without entering busy state', async () => {
+  let busyCalls = 0;
+  let runCalls = 0;
+  const scheduler = new DeleteMaintenanceScheduler({
+    batchWindowMs: 0,
+    runTask: async () => {
+      runCalls += 1;
+    },
+    onBusy: () => {
+      busyCalls += 1;
+    },
+    onIdle: () => {},
+  });
+  scheduler.destroy();
+
+  await assert.rejects(
+    scheduler.enqueue(() => ({ kind: 'session', sessionId: 1 })),
+    /shutting down/,
+  );
+  assert.equal(busyCalls, 0);
+  assert.equal(runCalls, 0);
+});
+
+test('scheduler rejects every request in a batch when the maintenance task fails', async () => {
+  const failure = new Error('maintenance failed');
+  const scheduler = new DeleteMaintenanceScheduler({
+    batchWindowMs: 0,
+    runTask: async () => {
+      throw failure;
+    },
+    onBusy: () => {},
+    onIdle: () => {},
+  });
+
+  const first = scheduler.enqueue(() => ({ kind: 'session', sessionId: 1 }));
+  const second = scheduler.enqueue(() => ({ kind: 'session', sessionId: 2 }));
+
+  const results = await Promise.allSettled([first, second]);
+  assert.deepEqual(
+    results.map((result) => (result.status === 'rejected' ? result.reason : null)),
+    [failure, failure],
+  );
+});
+
+test('scheduler rejects only the request whose task resolution fails', async () => {
+  const failure = new Error('resolution failed');
+  const tasks: DeleteMaintenanceTask[] = [];
+  const scheduler = new DeleteMaintenanceScheduler({
+    batchWindowMs: 0,
+    runTask: async (task) => {
+      tasks.push(task);
+    },
+    onBusy: () => {},
+    onIdle: () => {},
+  });
+
+  const failed = scheduler.enqueue(() => {
+    throw failure;
+  });
+  const succeeded = scheduler.enqueue(() => ({ kind: 'session', sessionId: 2 }));
+
+  const results = await Promise.allSettled([failed, succeeded]);
+  assert.equal(results[0]?.status, 'rejected');
+  assert.equal(results[0]?.status === 'rejected' ? results[0].reason : null, failure);
+  assert.equal(results[1]?.status, 'fulfilled');
+  assert.deepEqual(tasks, [{ kind: 'session', sessionId: 2 }]);
+});
+
+test('scheduler does not schedule another drain when the queue is empty', async () => {
+  const originalSetTimeout = globalThis.setTimeout;
+  let timerCalls = 0;
+  globalThis.setTimeout = ((handler: TimerHandler, timeout?: number, ...args: unknown[]) => {
+    timerCalls += 1;
+    return originalSetTimeout(handler, timeout, ...args);
+  }) as typeof setTimeout;
+
+  try {
+    const scheduler = new DeleteMaintenanceScheduler({
+      batchWindowMs: 0,
+      runTask: async () => {},
+      onBusy: () => {},
+      onIdle: () => {},
+    });
+
+    await scheduler.enqueue(() => ({ kind: 'session', sessionId: 1 }));
+    assert.equal(timerCalls, 1);
+  } finally {
+    globalThis.setTimeout = originalSetTimeout;
+  }
+});
+
+test('scheduler serializes batches and rejects requests queued at destruction', async () => {
+  const releases: Array<() => void> = [];
+  let activeTasks = 0;
+  let maxActiveTasks = 0;
+  const scheduler = new DeleteMaintenanceScheduler({
+    batchWindowMs: 0,
+    runTask: async () => {
+      activeTasks += 1;
+      maxActiveTasks = Math.max(maxActiveTasks, activeTasks);
+      await new Promise<void>((resolve) => releases.push(resolve));
+      activeTasks -= 1;
+    },
+    onBusy: () => {},
+    onIdle: () => {},
+  });
+
+  const first = scheduler.enqueue(() => ({ kind: 'session', sessionId: 1 }));
+  const maxPollAttempts = 100;
+  let pollAttempts = 0;
+  while (releases.length === 0 && pollAttempts < maxPollAttempts) {
+    pollAttempts += 1;
+    await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  }
+  assert.ok(
+    releases.length > 0,
+    `runTask did not produce a release after ${maxPollAttempts} polling attempts`,
+  );
+  const queued = scheduler.enqueue(() => ({ kind: 'session', sessionId: 2 }));
+  scheduler.destroy();
+
+  await assert.rejects(queued, /shutting down/);
+  releases[0]?.();
+  await first;
+  assert.equal(maxActiveTasks, 1);
+});
@@ -0,0 +1,105 @@
+import type { DeleteMaintenanceOperation, DeleteMaintenanceTask } from './delete-maintenance';
+
+type ResolveDeleteMaintenanceOperation = () =>
+  | DeleteMaintenanceOperation
+  | null
+  | Promise<DeleteMaintenanceOperation | null>;
+
+interface PendingDeleteMaintenanceRequest {
+  resolveTask: ResolveDeleteMaintenanceOperation;
+  resolve: () => void;
+  reject: (error: unknown) => void;
+}
+
+interface DeleteMaintenanceSchedulerOptions {
+  batchWindowMs: number;
+  runTask: (task: DeleteMaintenanceTask) => Promise<void>;
+  onBusy: () => void;
+  onIdle: () => void;
+}
+
+export class DeleteMaintenanceScheduler {
+  private readonly pendingRequests: PendingDeleteMaintenanceRequest[] = [];
+  private running = false;
+  private drainTimer: ReturnType<typeof setTimeout> | null = null;
+  private pendingTaskCount = 0;
+  private destroyed = false;
+
+  constructor(private readonly options: DeleteMaintenanceSchedulerOptions) {}
+
+  enqueue(resolveTask: ResolveDeleteMaintenanceOperation): Promise<void> {
+    if (this.destroyed) {
+      return Promise.reject(new Error('Immersion tracker is shutting down'));
+    }
+
+    if (this.pendingTaskCount === 0) this.options.onBusy();
+    this.pendingTaskCount += 1;
+
+    const result = new Promise<void>((resolve, reject) => {
+      this.pendingRequests.push({ resolveTask, resolve, reject });
+      this.scheduleDrain();
+    });
+
+    return result.finally(() => {
+      this.pendingTaskCount -= 1;
+      if (this.pendingTaskCount === 0) this.options.onIdle();
+    });
+  }
+
+  destroy(): void {
+    if (this.destroyed) return;
+    this.destroyed = true;
+    if (this.drainTimer) {
+      clearTimeout(this.drainTimer);
+      this.drainTimer = null;
+    }
+    const error = new Error('Immersion tracker is shutting down');
+    for (const request of this.pendingRequests.splice(0)) request.reject(error);
+  }
+
+  private scheduleDrain(): void {
+    if (this.destroyed || this.running || this.drainTimer || this.pendingRequests.length === 0) {
+      return;
+    }
+    this.drainTimer = setTimeout(() => {
+      this.drainTimer = null;
+      void this.drain();
+    }, this.options.batchWindowMs);
+  }
+
+  private async drain(): Promise<void> {
+    if (this.running || this.pendingRequests.length === 0) return;
+    this.running = true;
+    const requests = this.pendingRequests.splice(0);
+    const runnable: Array<{
+      request: PendingDeleteMaintenanceRequest;
+      task: DeleteMaintenanceOperation;
+    }> = [];
+
+    for (const request of requests) {
+      try {
+        const task = await request.resolveTask();
+        if (task) runnable.push({ request, task });
+        else request.resolve();
+      } catch (error) {
+        request.reject(error);
+      }
+    }
+
+    if (runnable.length > 0) {
+      const task: DeleteMaintenanceTask =
+        runnable.length === 1
+          ? runnable[0]!.task
+          : { kind: 'batch', tasks: runnable.map((entry) => entry.task) };
+      try {
+        await this.options.runTask(task);
+        for (const { request } of runnable) request.resolve();
+      } catch (error) {
+        for (const { request } of runnable) request.reject(error);
+      }
+    }
+
+    this.running = false;
+    this.scheduleDrain();
+  }
+}
@@ -0,0 +1,239 @@
+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 {
+  DeleteMaintenanceWorkerRuntime,
+  resolveDeleteMaintenanceWorkerPath,
+} from './delete-maintenance-worker-runtime';
+import { executeDeleteMaintenanceTask } from './delete-maintenance';
+import { startSessionRecord } from './session';
+import { Database } from './sqlite';
+import { applyPragmas, ensureSchema, getOrCreateVideoRecord } from './storage';
+
+type FakeWorkerListener = (value: never) => void;
+
+function createFakeWorker() {
+  const listeners = new Map<string, FakeWorkerListener>();
+  const terminationState = { calls: 0 };
+  const worker = {
+    once(event: string, listener: FakeWorkerListener) {
+      listeners.set(event, listener);
+      return this;
+    },
+    terminate: async () => {
+      terminationState.calls += 1;
+      return 0;
+    },
+  };
+  return { worker, listeners, terminationState };
+}
+
+type FakeWorker = ReturnType<typeof createFakeWorker>['worker'];
+
+test('a delete batch rebuilds lifetime summaries once', () => {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-delete-batch-test-'));
+  const dbPath = path.join(tempDir, 'immersion.sqlite');
+  let db = new Database(dbPath);
+
+  try {
+    applyPragmas(db);
+    ensureSchema(db);
+    const videoId = getOrCreateVideoRecord(db, 'local:/tmp/batch-delete.mkv', {
+      canonicalTitle: 'Batch Delete',
+      sourcePath: '/tmp/batch-delete.mkv',
+      sourceUrl: null,
+      sourceType: 1,
+    });
+    const firstSessionId = startSessionRecord(db, videoId, 1_000).sessionId;
+    const secondSessionId = startSessionRecord(db, videoId, 2_000).sessionId;
+    const deletedVideoId = getOrCreateVideoRecord(db, 'local:/tmp/batch-delete-video.mkv', {
+      canonicalTitle: 'Batch Delete Video',
+      sourcePath: '/tmp/batch-delete-video.mkv',
+      sourceUrl: null,
+      sourceType: 1,
+    });
+    startSessionRecord(db, deletedVideoId, 3_000);
+    db.exec(`
+      CREATE TABLE delete_rebuild_audit (id INTEGER PRIMARY KEY);
+      CREATE TRIGGER count_delete_lifetime_rebuild
+      AFTER UPDATE OF last_rebuilt_ms ON imm_lifetime_global
+      BEGIN
+        INSERT INTO delete_rebuild_audit (id) VALUES (NULL);
+      END;
+    `);
+    db.close();
+
+    executeDeleteMaintenanceTask(dbPath, {
+      kind: 'batch',
+      tasks: [
+        { kind: 'session', sessionId: firstSessionId },
+        { kind: 'video', videoId: deletedVideoId },
+      ],
+    });
+
+    db = new Database(dbPath);
+    const audit = db.prepare('SELECT COUNT(*) AS total FROM delete_rebuild_audit').get() as {
+      total: number;
+    };
+    const retainedSession = db
+      .prepare('SELECT session_id AS sessionId FROM imm_sessions WHERE video_id = ?')
+      .get(videoId) as { sessionId: number } | null;
+    const deletedVideo = db
+      .prepare('SELECT video_id AS videoId FROM imm_videos WHERE video_id = ?')
+      .get(deletedVideoId) as { videoId: number } | null;
+    assert.equal(retainedSession?.sessionId, secondSessionId);
+    assert.equal(deletedVideo, undefined);
+    assert.equal(
+      audit.total,
+      2,
+      'one rebuild performs exactly its reset and final global summary writes',
+    );
+  } finally {
+    try {
+      db.close();
+    } catch {
+      // The setup connection closes before maintenance runs.
+    }
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+});
+
+test(
+  'compiled delete worker removes data through its separate database connection',
+  { skip: resolveDeleteMaintenanceWorkerPath() === null },
+  async () => {
+    const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-delete-worker-test-'));
+    const dbPath = path.join(tempDir, 'immersion.sqlite');
+    const runtime = new DeleteMaintenanceWorkerRuntime();
+    let db = new Database(dbPath);
+
+    try {
+      applyPragmas(db);
+      ensureSchema(db);
+      const videoId = getOrCreateVideoRecord(db, 'local:/tmp/worker-delete.mkv', {
+        canonicalTitle: 'Worker Delete',
+        sourcePath: '/tmp/worker-delete.mkv',
+        sourceUrl: null,
+        sourceType: 1,
+      });
+      const firstSessionId = startSessionRecord(db, videoId, 1_000).sessionId;
+      const secondSessionId = startSessionRecord(db, videoId, 2_000).sessionId;
+      db.close();
+
+      await runtime.run(dbPath, {
+        kind: 'batch',
+        tasks: [
+          { kind: 'session', sessionId: firstSessionId },
+          { kind: 'session', sessionId: secondSessionId },
+        ],
+      });
+
+      db = new Database(dbPath);
+      const row = db
+        .prepare('SELECT COUNT(*) AS total FROM imm_sessions WHERE video_id = ?')
+        .get(videoId) as { total: number };
+      assert.equal(row.total, 0);
+    } finally {
+      runtime.destroy();
+      try {
+        db.close();
+      } catch {
+        // The setup connection is already closed before the worker starts.
+      }
+      fs.rmSync(tempDir, { recursive: true, force: true });
+    }
+  },
+);
+
+test('worker runtime warns before falling back when no emitted worker is available', async () => {
+  const warnings: unknown[][] = [];
+  const fallbackTasks: unknown[] = [];
+  const runtime = new DeleteMaintenanceWorkerRuntime({
+    resolveWorkerPath: () => null,
+    warn: (...args) => warnings.push(args),
+    executeFallback: (_dbPath, task) => fallbackTasks.push(task),
+  });
+
+  await runtime.run('/tmp/fallback.sqlite', { kind: 'session', sessionId: 1 });
+
+  assert.equal(warnings.length, 1);
+  assert.match(String(warnings[0]?.[0]), /worker unavailable/i);
+  assert.deepEqual(fallbackTasks, [{ kind: 'session', sessionId: 1 }]);
+});
+
+test('worker runtime terminates a worker after successful settlement', async () => {
+  const { worker, listeners, terminationState } = createFakeWorker();
+  const runtime = new DeleteMaintenanceWorkerRuntime({
+    resolveWorkerPath: () => '/tmp/delete-worker.js',
+    createWorker: async () => worker,
+  });
+
+  const result = runtime.run('/tmp/test.sqlite', { kind: 'session', sessionId: 1 });
+  await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  listeners.get('message')?.({ ok: true } as never);
+  await result;
+
+  assert.equal(terminationState.calls, 1);
+});
+
+test('worker runtime terminates a worker after failed settlement', async () => {
+  const { worker, listeners, terminationState } = createFakeWorker();
+  const runtime = new DeleteMaintenanceWorkerRuntime({
+    resolveWorkerPath: () => '/tmp/delete-worker.js',
+    createWorker: async () => worker,
+  });
+
+  const result = runtime.run('/tmp/test.sqlite', { kind: 'session', sessionId: 1 });
+  await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  listeners.get('error')?.(new Error('worker failed') as never);
+
+  await assert.rejects(result, /worker failed/);
+  assert.equal(terminationState.calls, 1);
+});
+
+test('worker runtime terminates a worker created after shutdown begins', async () => {
+  const { worker, listeners, terminationState } = createFakeWorker();
+  const createGate: { resolve?: (worker: FakeWorker) => void } = {};
+  const fallbackTasks: unknown[] = [];
+  const runtime = new DeleteMaintenanceWorkerRuntime({
+    resolveWorkerPath: () => '/tmp/delete-worker.js',
+    createWorker: () =>
+      new Promise((resolve) => {
+        createGate.resolve = resolve;
+      }),
+    executeFallback: (_dbPath, task) => fallbackTasks.push(task),
+  });
+
+  const result = runtime.run('/tmp/test.sqlite', { kind: 'session', sessionId: 1 });
+  await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  runtime.destroy();
+  createGate.resolve?.(worker);
+
+  await assert.rejects(result, /shut down/);
+  assert.equal(terminationState.calls, 1);
+  assert.equal(listeners.size, 0);
+  assert.deepEqual(fallbackTasks, []);
+});
+
+test('worker runtime does not fall back when worker creation fails during shutdown', async () => {
+  const createGate: { reject?: (error: Error) => void } = {};
+  const fallbackTasks: unknown[] = [];
+  const runtime = new DeleteMaintenanceWorkerRuntime({
+    resolveWorkerPath: () => '/tmp/delete-worker.js',
+    createWorker: () =>
+      new Promise((_resolve, reject) => {
+        createGate.reject = reject;
+      }),
+    executeFallback: (_dbPath, task) => fallbackTasks.push(task),
+  });
+
+  const result = runtime.run('/tmp/test.sqlite', { kind: 'session', sessionId: 1 });
+  await new Promise<void>((resolve) => setTimeout(resolve, 0));
+  runtime.destroy();
+  createGate.reject?.(new Error('creation failed'));
+
+  await assert.rejects(result, /shut down/);
+  assert.deepEqual(fallbackTasks, []);
+});
@@ -0,0 +1,121 @@
+import fs from 'node:fs';
+import path from 'node:path';
+import { createLogger } from '../../../logger';
+import { executeDeleteMaintenanceTask, type DeleteMaintenanceTask } from './delete-maintenance';
+
+interface DeleteMaintenanceWorkerResponse {
+  ok?: unknown;
+  error?: unknown;
+}
+
+export type RunDeleteMaintenanceTask = (
+  dbPath: string,
+  task: DeleteMaintenanceTask,
+) => Promise<void>;
+
+interface DeleteMaintenanceWorkerHandle {
+  once(event: 'message', listener: (message: DeleteMaintenanceWorkerResponse) => void): this;
+  once(event: 'error', listener: (error: Error) => void): this;
+  once(event: 'exit', listener: (code: number) => void): this;
+  terminate(): Promise<number>;
+}
+
+interface DeleteMaintenanceWorkerRuntimeOptions {
+  resolveWorkerPath?: () => string | null;
+  createWorker?: (
+    workerPath: string,
+    workerData: { dbPath: string; task: DeleteMaintenanceTask },
+  ) => Promise<DeleteMaintenanceWorkerHandle>;
+  executeFallback?: typeof executeDeleteMaintenanceTask;
+  warn?: (message: string, ...meta: unknown[]) => void;
+}
+
+export function resolveDeleteMaintenanceWorkerPath(): string | null {
+  const workerPath = path.join(__dirname, 'delete-maintenance-worker-thread.js');
+  return fs.existsSync(workerPath) ? workerPath : null;
+}
+
+const logger = createLogger('main:immersion-tracker:delete-worker');
+
+export class DeleteMaintenanceWorkerRuntime {
+  private readonly activeWorkers = new Set<DeleteMaintenanceWorkerHandle>();
+  private destroyed = false;
+
+  constructor(private readonly options: DeleteMaintenanceWorkerRuntimeOptions = {}) {}
+
+  async run(dbPath: string, task: DeleteMaintenanceTask): Promise<void> {
+    if (this.destroyed) {
+      throw new Error('Delete maintenance worker is shut down');
+    }
+
+    let worker: DeleteMaintenanceWorkerHandle;
+    try {
+      const workerPath = (this.options.resolveWorkerPath ?? resolveDeleteMaintenanceWorkerPath)();
+      if (!workerPath) throw new Error('Emitted delete-maintenance worker module was not found');
+      const createWorker =
+        this.options.createWorker ??
+        (async (resolvedPath, workerData) => {
+          const { Worker } = await import('node:worker_threads');
+          return new Worker(resolvedPath, { workerData });
+        });
+      worker = await createWorker(workerPath, { dbPath, task });
+    } catch (error) {
+      if (this.destroyed) {
+        throw new Error('Delete maintenance worker is shut down');
+      }
+      (this.options.warn ?? logger.warn)(
+        'Delete maintenance worker unavailable; running maintenance on the current thread',
+        error,
+      );
+      (this.options.executeFallback ?? executeDeleteMaintenanceTask)(dbPath, task);
+      return;
+    }
+
+    if (this.destroyed) {
+      await worker.terminate().catch(() => undefined);
+      throw new Error('Delete maintenance worker is shut down');
+    }
+
+    await new Promise<void>((resolve, reject) => {
+      let settled = false;
+      this.activeWorkers.add(worker);
+
+      const settle = (error?: Error) => {
+        if (settled) return;
+        settled = true;
+        this.activeWorkers.delete(worker);
+        if (error) reject(error);
+        else resolve();
+        void worker.terminate();
+      };
+
+      worker.once('message', (message: DeleteMaintenanceWorkerResponse) => {
+        if (message.ok === true) {
+          settle();
+          return;
+        }
+        const detail = typeof message.error === 'string' ? message.error : 'unknown worker error';
+        settle(new Error(`Delete maintenance failed: ${detail}`));
+      });
+      worker.once('error', (error) => settle(error));
+      worker.once('exit', (code) => {
+        settle(
+          new Error(
+            code === 0
+              ? 'Delete maintenance worker exited without a response'
+              : `Delete maintenance worker exited with code ${code}`,
+          ),
+        );
+      });
+    });
+  }
+
+  destroy(): void {
+    if (this.destroyed) return;
+    this.destroyed = true;
+    for (const worker of this.activeWorkers) {
+      void worker.terminate();
+    }
+    this.activeWorkers.clear();
+  }
+}
@@ -0,0 +1,22 @@
+import { parentPort, workerData } from 'node:worker_threads';
+import { executeDeleteMaintenanceTask, type DeleteMaintenanceTask } from './delete-maintenance';
+
+interface DeleteMaintenanceWorkerData {
+  dbPath: string;
+  task: DeleteMaintenanceTask;
+}
+
+if (!parentPort) {
+  throw new Error('delete maintenance worker missing parent port');
+}
+
+const port = parentPort;
+const request = workerData as DeleteMaintenanceWorkerData;
+
+try {
+  executeDeleteMaintenanceTask(request.dbPath, request.task);
+  port.postMessage({ ok: true });
+} catch (error) {
+  const message = error instanceof Error ? error.message : String(error);
+  port.postMessage({ ok: false, error: message });
+}
@@ -0,0 +1,47 @@
+import { Database } from './sqlite';
+import { applyPragmas } from './storage';
+import { deleteAnime, deleteSession, deleteSessions, deleteVideo } from './query-maintenance';
+import {
+  deleteMaintenanceBatch,
+  type DeleteMaintenanceOperation,
+} from './query-delete-maintenance';
+
+export type { DeleteMaintenanceOperation } from './query-delete-maintenance';
+
+export type DeleteMaintenanceTask =
+  | DeleteMaintenanceOperation
+  | { kind: 'batch'; tasks: DeleteMaintenanceOperation[] };
+
+function executeDeleteMaintenanceOperation(
+  db: InstanceType<typeof Database>,
+  task: DeleteMaintenanceOperation,
+): void {
+  switch (task.kind) {
+    case 'session':
+      deleteSession(db, task.sessionId);
+      return;
+    case 'sessions':
+      deleteSessions(db, task.sessionIds);
+      return;
+    case 'video':
+      deleteVideo(db, task.videoId);
+      return;
+    case 'anime':
+      deleteAnime(db, task.animeId);
+      return;
+  }
+}
+
+export function executeDeleteMaintenanceTask(dbPath: string, task: DeleteMaintenanceTask): void {
+  const db = new Database(dbPath);
+  try {
+    applyPragmas(db);
+    if (task.kind === 'batch') {
+      deleteMaintenanceBatch(db, task.tasks);
+      return;
+    }
+    executeDeleteMaintenanceOperation(db, task);
+  } finally {
+    db.close();
+  }
+}
@@ -0,0 +1,423 @@
+/*
+ * Retroactive removal of animation-burst subtitle lines from the stats database.
+ *
+ * Before the live ingest gate existed, a karaoke OP recorded one line -- and one count
+ * for every word in it -- per animation frame, which is enough to put an OP lyric at the
+ * top of "Top Repeated Words" for good. This module finds those runs in what is already
+ * stored and takes them back down to one line.
+ *
+ * Only timing is available here: the stored text has been stripped of ASS markup, so the
+ * authoring evidence the file-level parser uses (`\t`, `\move`, karaoke timing, a
+ * changing override signature) is long gone. What is left is a run of identical,
+ * contiguous, short-lived lines inside a single session.
+ *
+ * The run has to be as long as the timing-only rule in `subtitle-cue-dedup` demands, but
+ * its short frames may be as long as the animation-frame bound rather than the much
+ * tighter timing-only one. A qualifying run may end with one longer hold, which is a
+ * common karaoke shape. Five or more repeats of the same text, each ending where the next
+ * begins, is already conclusive on its own -- no dialogue does that -- and the tighter
+ * bound would walk straight past the heavier typesetting that motivated this, where
+ * frames sit nearer a quarter of a second. Both bounds are options, so a cautious run can
+ * ask for more, and a dry run always reports before anything is removed.
+ *
+ * Scope: subtitle lines, their word/kanji occurrences, and the `imm_words`/`imm_kanji`
+ * aggregates those occurrences feed. Session telemetry (`lines_seen`, `tokens_seen`) and
+ * the rollups derived from it are left alone; they are cumulative samples taken at record
+ * time, and for sessions whose raw rows have since been pruned they cannot be recomputed.
+ */
+
+import type { DatabaseSync } from './sqlite';
+import {
+  ANIMATION_FRAME_MAX_SECONDS,
+  DUPLICATE_CUE_GAP_TOLERANCE_SECONDS,
+  MIN_STREAM_RESIDUE_FRAMES,
+  MIN_TIMING_ONLY_FRAMES,
+  TIMING_ONLY_FRAME_MAX_SECONDS,
+} from '../subtitle-burst-constants';
+import {
+  applyLexicalRemovals,
+  makePlaceholders,
+  planLexicalRemovalsForLines,
+  toDbTimestamp,
+} from './query-shared';
+import { nowMs } from './time';
+
+const MS_PER_DAY = 86_400_000;
+/** SQLite caps bound parameters per statement; stay well under it. */
+const ID_BATCH_SIZE = 400;
+const DEFAULT_SAMPLE_LIMIT = 20;
+
+export interface DuplicateSubtitleLineCleanupOptions {
+  /** Only consider lines recorded within this many days. Null or omitted = all history. */
+  lookbackDays?: number | null;
+  /** Measure without writing. */
+  dryRun?: boolean;
+  /** Identical contiguous lines needed before a run counts as an animation. */
+  minRunLength?: number;
+  /** Longest a single event may last and still look like an animation frame. */
+  maxFrameSeconds?: number;
+  /** How many of the largest runs to describe in the summary. */
+  sampleLimit?: number;
+}
+
+export interface DuplicateSubtitleLineBurst {
+  sessionId: number;
+  videoId: number;
+  text: string;
+  /** Kept line, extended to cover the whole run. */
+  keptLineId: number;
+  removedLineIds: number[];
+  startMs: number;
+  endMs: number;
+}
+
+export interface DuplicateSubtitleLineSample {
+  videoId: number;
+  videoTitle: string | null;
+  text: string;
+  frames: number;
+  removedLines: number;
+  startMs: number;
+  endMs: number;
+}
+
+export interface DuplicateSubtitleLineCleanupSummary {
+  dryRun: boolean;
+  lookbackDays: number | null;
+  scannedLines: number;
+  burstGroups: number;
+  removedLines: number;
+  removedWordOccurrences: number;
+  removedKanjiOccurrences: number;
+  samples: DuplicateSubtitleLineSample[];
+}
+
+export interface StoredSubtitleLineRow {
+  lineId: number;
+  sessionId: number;
+  videoId: number;
+  text: string;
+  startMs: number;
+  endMs: number;
+}
+
+interface ResolvedBounds {
+  lookbackDays: number | null;
+  minRunLength: number;
+  maxFrameMs: number;
+  /** Shorter runs qualify only when every event sits under this much stricter bound. */
+  residueMinRunLength: number;
+  strictFrameMs: number;
+  gapToleranceMs: number;
+  sampleLimit: number;
+}
+
+function resolveBounds(options: DuplicateSubtitleLineCleanupOptions): ResolvedBounds {
+  const lookbackDays =
+    typeof options.lookbackDays === 'number' && Number.isFinite(options.lookbackDays)
+      ? Math.max(1, Math.floor(options.lookbackDays))
+      : null;
+  const minRunLength =
+    typeof options.minRunLength === 'number' && Number.isFinite(options.minRunLength)
+      ? Math.max(2, Math.floor(options.minRunLength))
+      : MIN_TIMING_ONLY_FRAMES;
+  const maxFrameSeconds =
+    typeof options.maxFrameSeconds === 'number' &&
+    Number.isFinite(options.maxFrameSeconds) &&
+    options.maxFrameSeconds > 0
+      ? options.maxFrameSeconds
+      : ANIMATION_FRAME_MAX_SECONDS;
+  const sampleLimit =
+    typeof options.sampleLimit === 'number' && options.sampleLimit >= 0
+      ? Math.floor(options.sampleLimit)
+      : DEFAULT_SAMPLE_LIMIT;
+  return {
+    lookbackDays,
+    minRunLength,
+    maxFrameMs: Math.round(maxFrameSeconds * 1000),
+    residueMinRunLength: Math.max(MIN_STREAM_RESIDUE_FRAMES, minRunLength - 1),
+    strictFrameMs: Math.round(TIMING_ONLY_FRAME_MAX_SECONDS * 1000),
+    gapToleranceMs: Math.round(DUPLICATE_CUE_GAP_TOLERANCE_SECONDS * 1000),
+    sampleLimit,
+  };
+}
+
+/**
+ * `CREATED_DATE` holds epoch milliseconds on rows this app wrote, but older and synced
+ * rows can carry seconds, so normalize before comparing against the cutoff.
+ */
+const CREATED_MS_SQL = `
+  CASE
+    WHEN sl.CREATED_DATE < 10000000000 THEN sl.CREATED_DATE * 1000
+    ELSE sl.CREATED_DATE
+  END`;
+
+function readCandidateLines(db: DatabaseSync, bounds: ResolvedBounds): StoredSubtitleLineRow[] {
+  const scope =
+    bounds.lookbackDays === null
+      ? ''
+      : `AND sl.CREATED_DATE IS NOT NULL AND ${CREATED_MS_SQL} >= ?`;
+  const params = bounds.lookbackDays === null ? [] : [nowMs() - bounds.lookbackDays * MS_PER_DAY];
+  return db
+    .prepare(
+      `SELECT
+         sl.line_id AS lineId,
+         sl.session_id AS sessionId,
+         sl.video_id AS videoId,
+         sl.text AS text,
+         sl.segment_start_ms AS startMs,
+         sl.segment_end_ms AS endMs
+       FROM imm_subtitle_lines sl
+       WHERE sl.segment_start_ms IS NOT NULL
+         AND sl.segment_end_ms IS NOT NULL
+         ${scope}
+       ORDER BY sl.session_id, sl.video_id, sl.segment_start_ms, sl.line_id`,
+    )
+    .all(...params) as StoredSubtitleLineRow[];
+}
+
+function isBurst(run: StoredSubtitleLineRow[], bounds: ResolvedBounds): boolean {
+  const isShortFrame = (row: StoredSubtitleLineRow): boolean =>
+    row.endMs - row.startMs <= bounds.maxFrameMs;
+  // The residue the live gate leaves behind: it records the first frames of a burst
+  // before the run is long enough to recognise, so one frame fewer than the timing-only
+  // minimum, every one under the strict timing-only bound. No dialogue holds identical
+  // sub-tenth-second lines back to back that many times.
+  if (
+    run.length >= bounds.residueMinRunLength &&
+    run.every((row) => row.endMs - row.startMs <= bounds.strictFrameMs)
+  ) {
+    return true;
+  }
+  if (run.length < bounds.minRunLength) {
+    return false;
+  }
+  if (run.every(isShortFrame)) {
+    return true;
+  }
+  // Karaoke commonly finishes its short animation frames with one long hold. Only the
+  // final event may exceed the frame bound, and the short frames before it must already
+  // meet the minimum run length on their own.
+  return (
+    run.length - 1 >= bounds.minRunLength &&
+    run.slice(0, -1).every(isShortFrame) &&
+    !isShortFrame(run[run.length - 1]!)
+  );
+}
+
+function toBurst(run: StoredSubtitleLineRow[]): DuplicateSubtitleLineBurst {
+  const [first] = run;
+  return {
+    sessionId: first!.sessionId,
+    videoId: first!.videoId,
+    text: first!.text,
+    keptLineId: first!.lineId,
+    removedLineIds: run.slice(1).map((row) => row.lineId),
+    startMs: first!.startMs,
+    endMs: run.reduce((latest, row) => Math.max(latest, row.endMs), first!.endMs),
+  };
+}
+
+/**
+ * Group stored lines into animation runs.
+ *
+ * Rows are bucketed per (session, video, text) before chaining, the way the file-level
+ * dedup buckets cues: dual-line karaoke interleaves two texts frame by frame, and
+ * chaining across the interleave would break every run at length one.
+ *
+ * Runs never cross a session, which is what keeps a rewatch intact: the same episode
+ * watched twice stores the same line twice, and those two belong to different sessions.
+ */
+export function findDuplicateSubtitleLineBursts(
+  rows: readonly StoredSubtitleLineRow[],
+  options: DuplicateSubtitleLineCleanupOptions = {},
+): DuplicateSubtitleLineBurst[] {
+  const bounds = resolveBounds(options);
+
+  // Insertion order preserves the query's startMs ordering within each bucket.
+  const rowsByKey = new Map<string, StoredSubtitleLineRow[]>();
+  for (const row of rows) {
+    const key = `${row.sessionId}|${row.videoId}|${row.text}`;
+    const bucket = rowsByKey.get(key);
+    if (bucket) {
+      bucket.push(row);
+    } else {
+      rowsByKey.set(key, [row]);
+    }
+  }
+
+  const bursts: DuplicateSubtitleLineBurst[] = [];
+  for (const bucket of rowsByKey.values()) {
+    if (bucket.length < 2) {
+      continue;
+    }
+
+    let run: StoredSubtitleLineRow[] = [];
+    let chainEndMs = 0;
+
+    const closeRun = (): void => {
+      if (run.length > 1 && isBurst(run, bounds)) {
+        bursts.push(toBurst(run));
+      }
+      run = [];
+    };
+
+    for (const row of bucket) {
+      if (run.length > 0 && row.startMs <= chainEndMs + bounds.gapToleranceMs) {
+        run.push(row);
+        chainEndMs = Math.max(chainEndMs, row.endMs);
+        continue;
+      }
+      closeRun();
+      run = [row];
+      chainEndMs = row.endMs;
+    }
+    closeRun();
+  }
+
+  return bursts;
+}
+
+function chunk<T>(values: T[], size: number): T[][] {
+  const chunks: T[][] = [];
+  for (let i = 0; i < values.length; i += size) {
+    chunks.push(values.slice(i, i + size));
+  }
+  return chunks;
+}
+
+function buildSamples(
+  db: DatabaseSync,
+  bursts: DuplicateSubtitleLineBurst[],
+  sampleLimit: number,
+): DuplicateSubtitleLineSample[] {
+  if (sampleLimit === 0 || bursts.length === 0) {
+    return [];
+  }
+  const largest = [...bursts]
+    .sort((a, b) => b.removedLineIds.length - a.removedLineIds.length)
+    .slice(0, sampleLimit);
+  const videoIds = [...new Set(largest.map((burst) => burst.videoId))];
+  const titles = new Map<number, string>();
+  for (const batch of chunk(videoIds, ID_BATCH_SIZE)) {
+    const rows = db
+      .prepare(
+        `SELECT video_id AS videoId, canonical_title AS title
+         FROM imm_videos
+         WHERE video_id IN (${makePlaceholders(batch)})`,
+      )
+      .all(...batch) as Array<{ videoId: number; title: string | null }>;
+    for (const row of rows) {
+      if (row.title) titles.set(row.videoId, row.title);
+    }
+  }
+
+  return largest.map((burst) => ({
+    videoId: burst.videoId,
+    videoTitle: titles.get(burst.videoId) ?? null,
+    text: burst.text,
+    frames: burst.removedLineIds.length + 1,
+    removedLines: burst.removedLineIds.length,
+    startMs: burst.startMs,
+    endMs: burst.endMs,
+  }));
+}
+
+function sumRemovedOccurrences(
+  db: DatabaseSync,
+  table: 'imm_word_line_occurrences' | 'imm_kanji_line_occurrences',
+  lineIds: number[],
+): number {
+  let total = 0;
+  for (const batch of chunk(lineIds, ID_BATCH_SIZE)) {
+    const row = db
+      .prepare(
+        `SELECT COALESCE(SUM(occurrence_count), 0) AS total
+         FROM ${table}
+         WHERE line_id IN (${makePlaceholders(batch)})`,
+      )
+      .get(...batch) as { total: number } | null;
+    total += row?.total ?? 0;
+  }
+  return total;
+}
+
+function applyBursts(db: DatabaseSync, bursts: DuplicateSubtitleLineBurst[]): void {
+  const removedLineIds = bursts.flatMap((burst) => burst.removedLineIds);
+  const currentMs = toDbTimestamp(nowMs());
+
+  db.exec('BEGIN IMMEDIATE');
+  try {
+    for (const batch of chunk(removedLineIds, ID_BATCH_SIZE)) {
+      const placeholders = makePlaceholders(batch);
+      // Measured before the delete, applied after it: `applyLexicalRemovals` checks the
+      // surviving occurrences to decide whether a zeroed count really means the word is
+      // gone, so the rows it inspects have to be the post-delete ones.
+      const plan = planLexicalRemovalsForLines(db, batch);
+      db.prepare(`DELETE FROM imm_word_line_occurrences WHERE line_id IN (${placeholders})`).run(
+        ...batch,
+      );
+      db.prepare(`DELETE FROM imm_kanji_line_occurrences WHERE line_id IN (${placeholders})`).run(
+        ...batch,
+      );
+      db.prepare(`DELETE FROM imm_subtitle_lines WHERE line_id IN (${placeholders})`).run(...batch);
+      applyLexicalRemovals(db, plan);
+    }
+
+    const extendStmt = db.prepare(
+      `UPDATE imm_subtitle_lines
+       SET segment_end_ms = ?, LAST_UPDATE_DATE = ?
+       WHERE line_id = ? AND (segment_end_ms IS NULL OR segment_end_ms < ?)`,
+    );
+    for (const burst of bursts) {
+      extendStmt.run(burst.endMs, currentMs, burst.keptLineId, burst.endMs);
+    }
+    db.exec('COMMIT');
+  } catch (error) {
+    try {
+      db.exec('ROLLBACK');
+    } catch {
+      // Surface the transaction failure, not the rollback's.
+    }
+    throw error;
+  }
+}
+
+/**
+ * Collapse stored animation bursts down to one line each.
+ *
+ * A dry run measures exactly what an apply would remove, using the same scan, so the
+ * numbers shown in a confirmation prompt are the numbers that will happen.
+ */
+export function cleanupDuplicateSubtitleLines(
+  db: DatabaseSync,
+  options: DuplicateSubtitleLineCleanupOptions = {},
+): DuplicateSubtitleLineCleanupSummary {
+  const bounds = resolveBounds(options);
+  const dryRun = options.dryRun === true;
+  const rows = readCandidateLines(db, bounds);
+  const bursts = findDuplicateSubtitleLineBursts(rows, options);
+  const removedLineIds = bursts.flatMap((burst) => burst.removedLineIds);
+
+  const summary: DuplicateSubtitleLineCleanupSummary = {
+    dryRun,
+    lookbackDays: bounds.lookbackDays,
+    scannedLines: rows.length,
+    burstGroups: bursts.length,
+    removedLines: removedLineIds.length,
+    removedWordOccurrences: sumRemovedOccurrences(db, 'imm_word_line_occurrences', removedLineIds),
+    removedKanjiOccurrences: sumRemovedOccurrences(
+      db,
+      'imm_kanji_line_occurrences',
+      removedLineIds,
+    ),
+    samples: buildSamples(db, bursts, bounds.sampleLimit),
+  };
+
+  if (dryRun || removedLineIds.length === 0) {
+    return summary;
+  }
+
+  applyBursts(db, bursts);
+  return summary;
+}
@@ -0,0 +1,196 @@
+import type { DatabaseSync } from './sqlite';
+import { rebuildLifetimeSummariesInTransaction } from './lifetime';
+import { getRollupGroupsForSessions, refreshRollupsForGroupsInTransaction } from './maintenance';
+import {
+  applyLexicalRemovals,
+  cleanupUnusedCoverArtBlobHash,
+  deleteSessionsByIds,
+  forEachIdChunk,
+  makePlaceholders,
+  planLexicalRemovalsForSessions,
+  SQLITE_ID_CHUNK_SIZE,
+  type LexicalRemovalPlan,
+} from './query-shared';
+
+export type DeleteMaintenanceOperation =
+  | { kind: 'session'; sessionId: number }
+  | { kind: 'sessions'; sessionIds: number[] }
+  | { kind: 'video'; videoId: number }
+  | { kind: 'anime'; animeId: number };
+
+function addOperationTargets(
+  operations: DeleteMaintenanceOperation[],
+  sessionIds: Set<number>,
+  videoIds: Set<number>,
+  animeIds: Set<number>,
+): void {
+  for (const operation of operations) {
+    switch (operation.kind) {
+      case 'session':
+        sessionIds.add(operation.sessionId);
+        break;
+      case 'sessions':
+        for (const sessionId of operation.sessionIds) sessionIds.add(sessionId);
+        break;
+      case 'video':
+        videoIds.add(operation.videoId);
+        break;
+      case 'anime':
+        animeIds.add(operation.animeId);
+        break;
+    }
+  }
+}
+
+function selectIds(
+  db: DatabaseSync,
+  buildSql: (placeholders: string) => string,
+  params: number[],
+  column: string,
+): number[] {
+  if (params.length === 0) return [];
+  const ids: number[] = [];
+  forEachIdChunk(params, (chunk) => {
+    const rows = db.prepare(buildSql(makePlaceholders(chunk))).all(...chunk) as Array<
+      Record<string, number>
+    >;
+    for (const row of rows) ids.push(row[column]!);
+  });
+  return ids;
+}
+
+function planLexicalRemovalsInChunks(db: DatabaseSync, sessionIds: number[]): LexicalRemovalPlan {
+  const combined: LexicalRemovalPlan = { words: [], kanji: [] };
+  const merge = (target: LexicalRemovalPlan['words'], source: LexicalRemovalPlan['words']) => {
+    const byId = new Map(target.map((entry) => [entry.id, entry]));
+    for (const entry of source) {
+      const existing = byId.get(entry.id);
+      if (!existing) {
+        const added = { ...entry };
+        target.push(added);
+        byId.set(entry.id, added);
+        continue;
+      }
+      existing.removedFrequency += entry.removedFrequency;
+      if (
+        entry.removedFirstSeenMs !== null &&
+        (existing.removedFirstSeenMs === null ||
+          entry.removedFirstSeenMs < existing.removedFirstSeenMs)
+      ) {
+        existing.removedFirstSeenMs = entry.removedFirstSeenMs;
+      }
+      if (
+        entry.removedLastSeenMs !== null &&
+        (existing.removedLastSeenMs === null ||
+          entry.removedLastSeenMs > existing.removedLastSeenMs)
+      ) {
+        existing.removedLastSeenMs = entry.removedLastSeenMs;
+      }
+    }
+  };
+
+  forEachIdChunk(sessionIds, (chunk) => {
+    const plan = planLexicalRemovalsForSessions(db, chunk);
+    merge(combined.words, plan.words);
+    merge(combined.kanji, plan.kanji);
+  });
+  return combined;
+}
+
+export function deleteMaintenanceBatch(
+  db: DatabaseSync,
+  operations: DeleteMaintenanceOperation[],
+): void {
+  if (operations.length === 0) return;
+
+  db.exec('BEGIN IMMEDIATE');
+  try {
+    const sessionIds = new Set<number>();
+    const videoIds = new Set<number>();
+    const animeIds = new Set<number>();
+    addOperationTargets(operations, sessionIds, videoIds, animeIds);
+
+    const animeIdList = [...animeIds];
+    for (const videoId of selectIds(
+      db,
+      (placeholders) => `SELECT video_id FROM imm_videos WHERE anime_id IN (${placeholders})`,
+      animeIdList,
+      'video_id',
+    )) {
+      videoIds.add(videoId);
+    }
+
+    const videoIdList = [...videoIds];
+    for (const sessionId of selectIds(
+      db,
+      (placeholders) => `SELECT session_id FROM imm_sessions WHERE video_id IN (${placeholders})`,
+      videoIdList,
+      'session_id',
+    )) {
+      sessionIds.add(sessionId);
+    }
+
+    const sessionIdList = [...sessionIds];
+    const lexicalRemovals = planLexicalRemovalsInChunks(db, sessionIdList);
+    const affectedRollupGroups = sessionIdList
+      .flatMap((_, index) =>
+        index % SQLITE_ID_CHUNK_SIZE === 0
+          ? getRollupGroupsForSessions(db, sessionIdList.slice(index, index + SQLITE_ID_CHUNK_SIZE))
+          : [],
+      )
+      .filter((group) => !videoIds.has(group.videoId));
+    const coverBlobHashes = new Set<string>();
+    if (videoIdList.length > 0) {
+      forEachIdChunk(videoIdList, (chunk) => {
+        const placeholders = makePlaceholders(chunk);
+        const artRows = db
+          .prepare(
+            `SELECT cover_blob_hash AS coverBlobHash
+             FROM imm_media_art
+             WHERE video_id IN (${placeholders}) AND cover_blob_hash IS NOT NULL`,
+          )
+          .all(...chunk) as Array<{ coverBlobHash: string }>;
+        for (const row of artRows) coverBlobHashes.add(row.coverBlobHash);
+      });
+
+      deleteSessionsByIds(db, sessionIdList);
+      forEachIdChunk(videoIdList, (chunk) => {
+        const placeholders = makePlaceholders(chunk);
+        db.prepare(`DELETE FROM imm_subtitle_lines WHERE video_id IN (${placeholders})`).run(
+          ...chunk,
+        );
+        db.prepare(`DELETE FROM imm_daily_rollups WHERE video_id IN (${placeholders})`).run(
+          ...chunk,
+        );
+        db.prepare(`DELETE FROM imm_monthly_rollups WHERE video_id IN (${placeholders})`).run(
+          ...chunk,
+        );
+        db.prepare(`DELETE FROM imm_media_art WHERE video_id IN (${placeholders})`).run(...chunk);
+        db.prepare(`DELETE FROM imm_videos WHERE video_id IN (${placeholders})`).run(...chunk);
+      });
+    } else {
+      deleteSessionsByIds(db, sessionIdList);
+    }
+
+    for (const coverBlobHash of coverBlobHashes) {
+      cleanupUnusedCoverArtBlobHash(db, coverBlobHash);
+    }
+    if (animeIdList.length > 0) {
+      forEachIdChunk(animeIdList, (chunk) => {
+        const placeholders = makePlaceholders(chunk);
+        db.prepare(`DELETE FROM imm_lifetime_anime WHERE anime_id IN (${placeholders})`).run(
+          ...chunk,
+        );
+        db.prepare(`DELETE FROM imm_anime WHERE anime_id IN (${placeholders})`).run(...chunk);
+      });
+    }
+
+    applyLexicalRemovals(db, lexicalRemovals);
+    rebuildLifetimeSummariesInTransaction(db);
+    refreshRollupsForGroupsInTransaction(db, affectedRollupGroups);
+    db.exec('COMMIT');
+  } catch (error) {
+    db.exec('ROLLBACK');
+    throw error;
+  }
+}
@@ -80,6 +80,14 @@ export function makePlaceholders(values: number[]): string {
  return values.map(() => '?').join(',');
 }

+export const SQLITE_ID_CHUNK_SIZE = 1_000;
+
+export function forEachIdChunk(ids: number[], callback: (chunk: number[]) => void): void {
+  for (let start = 0; start < ids.length; start += SQLITE_ID_CHUNK_SIZE) {
+    callback(ids.slice(start, start + SQLITE_ID_CHUNK_SIZE));
+  }
+}
+
 export function resolvedCoverBlobExpr(mediaAlias: string, blobStoreAlias: string): string {
  return `COALESCE(${blobStoreAlias}.cover_blob, CASE WHEN ${mediaAlias}.cover_blob_hash IS NULL THEN ${mediaAlias}.cover_blob ELSE NULL END)`;
 }
@@ -268,6 +276,19 @@ export function planLexicalRemovalsForSessions(
  return planLexicalRemovals(db, `sl.session_id IN (${makePlaceholders(sessionIds)})`, sessionIds);
 }

+/**
+ * Measure what deleting these individual subtitle lines removes from the vocabulary
+ * tables. Used by the duplicate-line cleanup, which drops animation frames out of the
+ * middle of sessions that otherwise stay intact.
+ */
+export function planLexicalRemovalsForLines(
+  db: DatabaseSync,
+  lineIds: number[],
+): LexicalRemovalPlan {
+  if (lineIds.length === 0) return EMPTY_LEXICAL_REMOVAL_PLAN;
+  return planLexicalRemovals(db, `sl.line_id IN (${makePlaceholders(lineIds)})`, lineIds);
+}
+
 /** Measure what deleting these videos removes from the vocabulary tables. */
 export function planLexicalRemovalsForVideos(
  db: DatabaseSync,
@@ -490,17 +511,19 @@ export function deleteSessionsByIds(db: DatabaseSync, sessionIds: number[]): voi
    return;
  }

-  const placeholders = makePlaceholders(sessionIds);
-  db.prepare(`DELETE FROM imm_subtitle_lines WHERE session_id IN (${placeholders})`).run(
-    ...sessionIds,
-  );
-  db.prepare(`DELETE FROM imm_session_telemetry WHERE session_id IN (${placeholders})`).run(
-    ...sessionIds,
-  );
-  db.prepare(`DELETE FROM imm_session_events WHERE session_id IN (${placeholders})`).run(
-    ...sessionIds,
-  );
-  db.prepare(`DELETE FROM imm_sessions WHERE session_id IN (${placeholders})`).run(...sessionIds);
+  forEachIdChunk(sessionIds, (chunk) => {
+    const placeholders = makePlaceholders(chunk);
+    db.prepare(`DELETE FROM imm_subtitle_lines WHERE session_id IN (${placeholders})`).run(
+      ...chunk,
+    );
+    db.prepare(`DELETE FROM imm_session_telemetry WHERE session_id IN (${placeholders})`).run(
+      ...chunk,
+    );
+    db.prepare(`DELETE FROM imm_session_events WHERE session_id IN (${placeholders})`).run(
+      ...chunk,
+    );
+    db.prepare(`DELETE FROM imm_sessions WHERE session_id IN (${placeholders})`).run(...chunk);
+  });
 }

 export function toDbMs(ms: number | bigint): bigint {
@@ -1,6 +1,7 @@
 import electron from 'electron';
 import type { BrowserWindow as ElectronBrowserWindow, IpcMainEvent } from 'electron';
 import type {
+  ChangelogSnapshot,
  CompiledSessionBinding,
  ControllerConfigUpdate,
  PlaylistBrowserMutationResult,
@@ -122,6 +123,7 @@ export interface IpcServiceDeps {
  removeCharacterDictionaryManagedEntry?: (mediaId: number) => Promise<unknown>;
  moveCharacterDictionaryManagedEntry?: (mediaId: number, direction: 1 | -1) => Promise<unknown>;
  appendClipboardVideoToQueue: () => { ok: boolean; message: string };
+  getChangelogSnapshot?: (options?: { refresh?: boolean }) => Promise<ChangelogSnapshot>;
  getPlaylistBrowserSnapshot: () => Promise<PlaylistBrowserSnapshot>;
  appendPlaylistBrowserFile: (filePath: string) => Promise<PlaylistBrowserMutationResult>;
  playPlaylistBrowserIndex: (index: number) => Promise<PlaylistBrowserMutationResult>;
@@ -297,6 +299,7 @@ export interface IpcDepsRuntimeOptions {
  removeCharacterDictionaryManagedEntry?: (mediaId: number) => Promise<unknown>;
  moveCharacterDictionaryManagedEntry?: (mediaId: number, direction: 1 | -1) => Promise<unknown>;
  appendClipboardVideoToQueue: () => { ok: boolean; message: string };
+  getChangelogSnapshot?: (options?: { refresh?: boolean }) => Promise<ChangelogSnapshot>;
  getPlaylistBrowserSnapshot: () => Promise<PlaylistBrowserSnapshot>;
  appendPlaylistBrowserFile: (filePath: string) => Promise<PlaylistBrowserMutationResult>;
  playPlaylistBrowserIndex: (index: number) => Promise<PlaylistBrowserMutationResult>;
@@ -418,6 +421,7 @@ export function createIpcDepsRuntime(options: IpcDepsRuntimeOptions): IpcService
        entries: [],
      })),
    appendClipboardVideoToQueue: options.appendClipboardVideoToQueue,
+    getChangelogSnapshot: options.getChangelogSnapshot,
    getPlaylistBrowserSnapshot: options.getPlaylistBrowserSnapshot,
    appendPlaylistBrowserFile: options.appendPlaylistBrowserFile,
    playPlaylistBrowserIndex: options.playPlaylistBrowserIndex,
@@ -820,6 +824,17 @@ export function registerIpcHandlers(deps: IpcServiceDeps, ipc: IpcMainRegistrar
    return deps.appendClipboardVideoToQueue();
  });

+  ipc.handle(IPC_CHANNELS.request.getChangelogSnapshot, async (_event, payload: unknown) => {
+    const refresh =
+      typeof payload === 'object' && payload !== null && 'refresh' in payload
+        ? (payload as { refresh?: unknown }).refresh === true
+        : false;
+    if (!deps.getChangelogSnapshot) {
+      throw new Error('Changelog service is unavailable.');
+    }
+    return await deps.getChangelogSnapshot({ refresh });
+  });
+
  ipc.handle(IPC_CHANNELS.request.getPlaylistBrowserSnapshot, async () => {
    return await deps.getPlaylistBrowserSnapshot();
  });
@@ -5,6 +5,7 @@ import {
  buildSentenceSearchOptions,
  enrichSessionsWithKnownWordMetrics,
  parseBooleanQuery,
+  parseDuplicateLineCleanupBody,
  parseExcludedWordsBody,
  parseIntQuery,
 } from './route-support.js';
@@ -40,6 +41,19 @@ export function registerStatsLibraryRoutes(
    return c.json(statsJson('setExcludedWords', { ok: true }));
  });

+  // Collapse animation bursts older versions recorded frame by frame. `dryRun` measures
+  // the same scan without writing, so the confirmation the user sees is the real cost.
+  app.post('/api/stats/maintenance/duplicate-lines', async (c) => {
+    const contentType = c.req.header('content-type')?.split(';', 1)[0]?.trim().toLowerCase();
+    if (contentType !== 'application/json') return c.body(null, 415);
+    const body = await c.req.json().catch(() => null);
+    const options = parseDuplicateLineCleanupBody(body);
+    if (!options) return c.body(null, 400);
+    const { dryRun, lookbackDays } = options;
+    const result = await tracker.cleanupDuplicateSubtitleLines({ dryRun, lookbackDays });
+    return c.json(statsJson('duplicateLineCleanup', result));
+  });
+
  app.get('/api/stats/vocabulary/occurrences', async (c) => {
    const headword = (c.req.query('headword') ?? '').trim();
    const word = (c.req.query('word') ?? '').trim();
@@ -88,6 +88,35 @@ export function parseExcludedWordsBody(body: unknown): StatsExcludedWord[] | nul
  return words;
 }

+/**
+ * Read a duplicate-line cleanup request. An explicit object with no lookback scans all
+ * history. Invalid bodies and invalid windows are rejected instead of broadening scope.
+ */
+export function parseDuplicateLineCleanupBody(body: unknown): {
+  dryRun: boolean;
+  lookbackDays: number | null;
+} | null {
+  if (!body || typeof body !== 'object' || Array.isArray(body)) {
+    return null;
+  }
+  const source = body as Record<string, unknown>;
+  if (source.dryRun !== undefined && typeof source.dryRun !== 'boolean') {
+    return null;
+  }
+  const rawLookback = source.lookbackDays;
+  if (
+    rawLookback !== undefined &&
+    rawLookback !== null &&
+    (typeof rawLookback !== 'number' || !Number.isFinite(rawLookback) || rawLookback < 1)
+  ) {
+    return null;
+  }
+  return {
+    dryRun: source.dryRun === true,
+    lookbackDays: typeof rawLookback === 'number' ? Math.floor(rawLookback) : null,
+  };
+}
+
 export function loadKnownWordsSet(cachePath: string | undefined): Set<string> | null {
  if (!cachePath || !existsSync(cachePath)) return null;
  try {
@@ -0,0 +1,48 @@
+/*
+ * Thresholds that decide when a run of repeated subtitle events is one animation.
+ *
+ * Three consumers have to agree on these numbers or the same karaoke line is one cue in
+ * the sidebar and two hundred in the stats: the file-level cue dedup
+ * (`subtitle-cue-dedup`), the live gate that decides what immersion stats record
+ * (`subtitle-line-dedup-gate`), and the retroactive database cleanup
+ * (`immersion-tracker/duplicate-line-cleanup`).
+ */
+
+/**
+ * Back-to-back frames of the same animation are authored flush against each other; a
+ * tiny tolerance absorbs the centisecond rounding of the ASS timestamp format.
+ */
+export const DUPLICATE_CUE_GAP_TOLERANCE_SECONDS = 0.05;
+
+/**
+ * A burst is a *sequence*. Two adjacent events are two events, not an animation --
+ * characters do repeat each other, and a repeated line can legitimately be short.
+ */
+export const MIN_BURST_EVENTS = 3;
+
+/**
+ * Real dialogue holds on screen for about a second, so a run with a couple of much
+ * shorter events among them looks like frames. Used only alongside authoring evidence.
+ */
+export const ANIMATION_FRAME_MAX_SECONDS = 0.3;
+
+/** A karaoke run usually ends on a long "hold" frame, so not every event is short. */
+export const MIN_TAGGED_BURST_FRAMES = 2;
+
+/**
+ * SRT and VTT carry no authoring metadata at all, so timing is the only signal available
+ * -- which makes it the easiest one to get wrong. ASS->SRT conversion leaves frames at
+ * ~0.04s, well under any real utterance, and a burst leaves many of them behind. Both
+ * bounds are deliberately far stricter than the ASS path: a run of ordinary short lines
+ * (`えっ` traded between characters) must not clear them.
+ */
+export const TIMING_ONLY_FRAME_MAX_SECONDS = 0.1;
+export const MIN_TIMING_ONLY_FRAMES = 5;
+
+/**
+ * The live gate can only recognise a streaming burst from the inside, so it records the
+ * first `MIN_TIMING_ONLY_FRAMES - 1` frames before it starts dropping. That stored
+ * residue is one frame short of the timing-only minimum, and the retroactive cleanup
+ * accepts it only when every event also sits under the strict timing-only frame bound.
+ */
+export const MIN_STREAM_RESIDUE_FRAMES = MIN_TIMING_ONLY_FRAMES - 1;
@@ -0,0 +1,180 @@
+/*
+ * Duplicate/animation-burst collapsing for parsed subtitle cues.
+ *
+ * Split out of the cue parser so the parsing rules and the "is this run one animation?"
+ * heuristics can be read -- and tested -- on their own. The parser owns the cue shape;
+ * this module only decides which cues survive.
+ */
+
+import { hasAssTemporalOverride, isAnimatedAssEffectKind } from './ass-text';
+import {
+  ANIMATION_FRAME_MAX_SECONDS,
+  DUPLICATE_CUE_GAP_TOLERANCE_SECONDS,
+  MIN_BURST_EVENTS,
+  MIN_TAGGED_BURST_FRAMES,
+  MIN_TIMING_ONLY_FRAMES,
+  TIMING_ONLY_FRAME_MAX_SECONDS,
+} from './subtitle-burst-constants';
+import type {
+  AnnotatedSubtitleCue,
+  SubtitleCue,
+  SubtitleSourceFormat,
+} from './subtitle-cue-parser';
+
+function cueKey(cue: SubtitleCue): string {
+  return `${cue.startTime}|${cue.endTime}|${cue.text}`;
+}
+
+/**
+ * Identical text over an identical span is redundant however it was authored -- most
+ * often a layered ASS event stacking a shadow copy under the visible one.
+ */
+function collapseExactDuplicates(cues: AnnotatedSubtitleCue[]): AnnotatedSubtitleCue[] {
+  const seen = new Set<string>();
+  return cues.filter((cue) => {
+    const key = cueKey(cue);
+    if (seen.has(key)) {
+      return false;
+    }
+    seen.add(key);
+    return true;
+  });
+}
+
+function countFramesShorterThan(run: AnnotatedSubtitleCue[], maxSeconds: number): number {
+  return run.filter((cue) => cue.endTime - cue.startTime < maxSeconds).length;
+}
+
+/**
+ * Evidence that a run of ASS events is one animation rather than several authored lines.
+ * A static tag says nothing on its own -- three events sharing one `\clip(...)` are three
+ * signs -- so the tag has to be temporal by nature (`\t`, `\move`, karaoke timing, or
+ * anything wrapped in `\t(...)`), an animated `Effect` column, or a value that actually
+ * changes from event to event, which is how per-frame typesetting is authored.
+ */
+export function hasAssAnimationEvidence(run: AnnotatedSubtitleCue[]): boolean {
+  if (run.every((cue) => hasAssTemporalOverride(cue.overrides))) {
+    return true;
+  }
+  if (run.every((cue) => isAnimatedAssEffectKind(cue.effectKind))) {
+    return true;
+  }
+
+  const [first] = run;
+  const everyEventTypeset = run.every((cue) => cue.overrides.length > 0);
+  const signatureChanges = run.some((cue) => cue.overrideSignature !== first!.overrideSignature);
+  return everyEventTypeset && signatureChanges;
+}
+
+export function isAnimationBurst(
+  run: AnnotatedSubtitleCue[],
+  format: SubtitleSourceFormat,
+): boolean {
+  if (run.length < MIN_BURST_EVENTS) {
+    return false;
+  }
+
+  if (format === 'srt') {
+    return (
+      run.length >= MIN_TIMING_ONLY_FRAMES &&
+      countFramesShorterThan(run, TIMING_ONLY_FRAME_MAX_SECONDS) === run.length
+    );
+  }
+
+  if (countFramesShorterThan(run, ANIMATION_FRAME_MAX_SECONDS) < MIN_TAGGED_BURST_FRAMES) {
+    return false;
+  }
+
+  // One animation belongs to one styled, one named source line. Two characters trading
+  // the same short word are two styles or two actors, and never merge.
+  const [first] = run;
+  if (run.some((cue) => cue.style !== first!.style || cue.name !== first!.name)) {
+    return false;
+  }
+
+  return hasAssAnimationEvidence(run);
+}
+
+/**
+ * Karaoke and sign typesetting emits one Dialogue event per animation frame, all carrying
+ * the same visible text over a contiguous span. Collapse each such run into a single cue.
+ *
+ * Only runs that look like animation collapse. Two ordinary lines that happen to repeat
+ * -- several characters each saying `おはよう` in turn, a positioned sign redrawn with a
+ * different fade -- stay separate, because merging them would destroy real mineable lines.
+ */
+function collapseAnimationBursts(
+  cues: AnnotatedSubtitleCue[],
+  format: SubtitleSourceFormat,
+): AnnotatedSubtitleCue[] {
+  const indicesByText = new Map<string, number[]>();
+  cues.forEach((cue, index) => {
+    const bucket = indicesByText.get(cue.text);
+    if (bucket) {
+      bucket.push(index);
+    } else {
+      indicesByText.set(cue.text, [index]);
+    }
+  });
+
+  const dropped = new Set<number>();
+  const extendedEnd = new Map<number, number>();
+
+  for (const indices of indicesByText.values()) {
+    if (indices.length < MIN_BURST_EVENTS) {
+      continue;
+    }
+
+    let runStart = 0;
+    while (runStart < indices.length) {
+      let runEnd = runStart;
+      let chainEnd = cues[indices[runStart]!]!.endTime;
+
+      while (runEnd + 1 < indices.length) {
+        const next = cues[indices[runEnd + 1]!]!;
+        if (next.startTime > chainEnd + DUPLICATE_CUE_GAP_TOLERANCE_SECONDS) {
+          break;
+        }
+        chainEnd = Math.max(chainEnd, next.endTime);
+        runEnd += 1;
+      }
+
+      const run = indices.slice(runStart, runEnd + 1).map((index) => cues[index]!);
+      if (isAnimationBurst(run, format)) {
+        for (let i = runStart + 1; i <= runEnd; i += 1) {
+          dropped.add(indices[i]!);
+        }
+        extendedEnd.set(indices[runStart]!, chainEnd);
+      }
+
+      runStart = runEnd + 1;
+    }
+  }
+
+  if (dropped.size === 0) {
+    return cues;
+  }
+
+  const merged: AnnotatedSubtitleCue[] = [];
+  cues.forEach((cue, index) => {
+    if (dropped.has(index)) {
+      return;
+    }
+    const end = extendedEnd.get(index);
+    merged.push(end !== undefined && end > cue.endTime ? { ...cue, endTime: end } : cue);
+  });
+  return merged;
+}
+
+/**
+ * Collapse redundant cues. Input must already be sorted by non-decreasing `startTime`,
+ * ties broken by `endTime` then source `order` -- burst detection chains events by
+ * comparing each one against the running end of the events before it, so an unsorted
+ * list breaks runs apart and leaves the frames behind.
+ */
+export function mergeDuplicateCues(
+  cues: AnnotatedSubtitleCue[],
+  format: SubtitleSourceFormat,
+): AnnotatedSubtitleCue[] {
+  return collapseAnimationBursts(collapseExactDuplicates(cues), format);
+}
@@ -91,6 +91,17 @@ test('parseSrtCues skips malformed timing lines gracefully', () => {
  assert.equal(cues[0]!.text, '有効');
 });

+test('parseSubtitleCues strips complete brace blocks from SRT and VTT text', () => {
+  const content = ['1', '00:00:01,000 --> 00:00:02,000', '彼は{謎}と言った', ''].join('\n');
+
+  for (const filename of ['test.srt', 'test.vtt']) {
+    const cues = parseSubtitleCues(content, filename);
+
+    assert.equal(cues.length, 1, filename);
+    assert.equal(cues[0]!.text, '彼はと言った', filename);
+  }
+});
+
 test('parseAssCues parses basic ASS dialogue lines', () => {
  const content = [
    '[Script Info]',
@@ -137,7 +148,9 @@ test('parseAssCues handles text containing commas', () => {
  assert.equal(cues[0]!.text, 'はい、そうです、ね');
 });

-test('parseAssCues handles \\N line breaks', () => {
+test('parseAssCues decodes \\N line breaks into real newlines', () => {
+  // ASS is decoded once, here at ingestion, so cue text matches what mpv hands over for
+  // the same line played live.
  const content = [
    '[Events]',
    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
@@ -146,7 +159,7 @@ test('parseAssCues handles \\N line breaks', () => {

  const cues = parseAssCues(content);

-  assert.equal(cues[0]!.text, '一行目\\N二行目');
+  assert.equal(cues[0]!.text, '一行目\n二行目');
 });

 test('parseAssCues strips HTML-like markup while preserving ASS line breaks', () => {
@@ -158,7 +171,46 @@ test('parseAssCues strips HTML-like markup while preserving ASS line breaks', ()

  const cues = parseAssCues(content);

-  assert.equal(cues[0]!.text, '一行目\\N二行目');
+  assert.equal(cues[0]!.text, '一行目\n二行目');
+});
+
+test('parseAssCues drops vector drawing runs enabled by \\p', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:00:01.00,0:00:04.00,Default,,0,0,0,,{\\an5\\pos(730,1042)\\p1\\blur1}m 20 0 b 10 0 0 10 0 20 b 0 31 10 40 20 40 {\\p0}',
+    'Dialogue: 0,0:00:05.00,0:00:08.00,Default,,0,0,0,,これは字幕',
+  ].join('\n');
+
+  const cues = parseAssCues(content);
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.text, 'これは字幕');
+});
+
+test('parseAssCues keeps text that follows a \\p0 reset on the same line', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:04.00,Default,,0,0,0,,{\\p1}m 0 0 l 10 10{\\p0}本文{\\p1}m 5 5 l 6 6{\\p0}続き',
+  ].join('\n');
+
+  const cues = parseAssCues(content);
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.text, '本文続き');
+});
+
+test('parseAssCues leaves \\pos untouched when no drawing mode is active', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:04.00,Default,,0,0,0,,{\\pos(960,1068)\\bord3}位置指定',
+  ].join('\n');
+
+  const cues = parseAssCues(content);
+
+  assert.equal(cues[0]!.text, '位置指定');
 });

 test('parseAssCues returns empty for content without Events section', () => {
@@ -258,6 +310,344 @@ test('parseSubtitleCues returns cues sorted by start time', () => {
  assert.equal(cues[1]!.text, '二番目');
 });

+test('parseSubtitleCues collapses per-frame karaoke duplicates into one cue', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}過ぎ去ってしまう瞬間を',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}過ぎ去ってしまう瞬間を',
+    'Dialogue: 0,0:00:01.09,0:00:03.55,OP_JP,,0,0,0,,{\\clip(m 3 3)}過ぎ去ってしまう瞬間を',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.startTime, 1.0);
+  assert.equal(cues[0]!.endTime, 3.55);
+  assert.equal(cues[0]!.text, '過ぎ去ってしまう瞬間を');
+});
+
+test('parseSubtitleCues keeps back-to-back plain dialogue repeats separate', () => {
+  // Several characters greeting in turn: distinct utterances that happen to abut.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:04:05.67,0:04:06.82,Dial_JP,,0,0,0,,おはよう',
+    'Dialogue: 0,0:04:06.82,0:04:07.56,Dial_JP,,0,0,0,,おはよう',
+    'Dialogue: 0,0:04:07.56,0:04:08.78,Dial_JP,,0,0,0,,おはよう',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+  assert.equal(cues[0]!.endTime, 246.82);
+  assert.equal(cues[2]!.startTime, 247.56);
+});
+
+test('parseSubtitleCues collapses exact duplicate cues even without effect tags', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:04.00,Default,,0,0,0,,重なった行',
+    'Dialogue: 1,0:00:01.00,0:00:04.00,Default,,0,0,0,,重なった行',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+});
+
+test('parseSubtitleCues collapses tag-less animation frames in converted SRT', () => {
+  // ASS -> SRT conversion drops override tags, so only the ~0.04s frame timing remains.
+  const lines = ['1', '00:00:07,870 --> 00:00:07,910', 'Kaguya Wants to be Confessed to', ''];
+  for (let i = 1; i < 8; i++) {
+    const start = 7910 + (i - 1) * 40;
+    const end = start + 40;
+    const at = (ms: number) =>
+      `00:00:0${Math.floor(ms / 1000)},${String(ms % 1000).padStart(3, '0')}`;
+    lines.push(String(i + 1), `${at(start)} --> ${at(end)}`, 'Kaguya Wants to be Confessed to', '');
+  }
+
+  const cues = parseSubtitleCues(lines.join('\n'), 'test.srt');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.startTime, 7.87);
+});
+
+test('parseSubtitleCues keeps identical lines that recur far apart', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:02.00,Default,,0,0,0,,なんで',
+    'Dialogue: 0,0:05:00.00,0:05:01.00,Default,,0,0,0,,なんで',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.equal(cues[0]!.startTime, 1.0);
+  assert.equal(cues[1]!.startTime, 300.0);
+});
+
+test('parseSubtitleCues keeps two positioned signs that repeat the same text', () => {
+  // Both carry override tags, but `\pos` and `\fad` are static placement, not animation.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:01:00.00,0:01:03.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(200,200)}第一話',
+    'Dialogue: 0,0:01:03.00,0:01:06.00,Sign,,0,0,0,,{\\pos(960,900)\\fad(200,200)}第一話',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.equal(cues[1]!.startTime, 63.0);
+});
+
+test('parseSubtitleCues keeps a run of ordinary positioned lines separate', () => {
+  // Three events is a sequence, but none of them runs at animation-frame speed.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:01:00.00,0:01:02.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(100,100)}止まれ',
+    'Dialogue: 0,0:01:02.00,0:01:04.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(100,100)}止まれ',
+    'Dialogue: 0,0:01:04.00,0:01:06.00,Sign,,0,0,0,,{\\pos(960,120)\\fad(100,100)}止まれ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues keeps a short repeated SRT pair without burst evidence', () => {
+  const content = [
+    '1',
+    '00:00:01,000 --> 00:00:01,200',
+    'えっ',
+    '',
+    '2',
+    '00:00:01,200 --> 00:00:01,400',
+    'えっ',
+    '',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.srt');
+
+  assert.equal(cues.length, 2);
+});
+
+test('parseSubtitleCues collapses a burst marked only by the Effect column', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,Karaoke,歌詞',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,Karaoke,歌詞',
+    'Dialogue: 0,0:00:01.09,0:00:03.55,OP_JP,,0,0,0,Karaoke,歌詞',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.endTime, 3.55);
+});
+
+test('parseSubtitleCues keeps a second karaoke burst that starts after a gap', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}リフレイン',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}リフレイン',
+    'Dialogue: 0,0:00:01.09,0:00:03.00,OP_JP,,0,0,0,,{\\clip(m 3 3)}リフレイン',
+    'Dialogue: 0,0:00:20.00,0:00:20.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}リフレイン',
+    'Dialogue: 0,0:00:20.05,0:00:20.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}リフレイン',
+    'Dialogue: 0,0:00:20.09,0:00:22.00,OP_JP,,0,0,0,,{\\clip(m 3 3)}リフレイン',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.equal(cues[0]!.endTime, 3.0);
+  assert.equal(cues[1]!.startTime, 20.0);
+  assert.equal(cues[1]!.endTime, 22.0);
+});
+
+test('parseSubtitleCues does not merge a burst into unrelated dialogue between frames', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,OP_JP,,0,0,0,,{\\clip(m 1 1)}歌詞',
+    'Dialogue: 0,0:00:01.02,0:00:03.00,Dial_JP,,0,0,0,,別のセリフ',
+    'Dialogue: 0,0:00:01.05,0:00:01.09,OP_JP,,0,0,0,,{\\clip(m 2 2)}歌詞',
+    'Dialogue: 0,0:00:01.09,0:00:03.55,OP_JP,,0,0,0,,{\\clip(m 3 3)}歌詞',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 2);
+  assert.deepEqual(
+    cues.map((cue) => cue.text),
+    ['歌詞', '別のセリフ'],
+  );
+  assert.equal(cues[0]!.endTime, 3.55);
+});
+
+test('parseSubtitleCues keeps rapid ASS lines from different actors separate', () => {
+  // Three 200ms `えっ` reactions traded between characters. Fast, adjacent and identical,
+  // but authored as three lines: different styles and different actors.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Dial_A,アリス,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Dial_B,ボブ,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Dial_C,キャロル,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues reads the speaker column when it is spelled Actor', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Actor, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Dial_JP,アリス,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Dial_JP,ボブ,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Dial_JP,キャロル,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues does not treat a custom Effect name as animation', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Sign,,0,0,0,scrolling-credit,制作',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Sign,,0,0,0,scrolling-credit,制作',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Sign,,0,0,0,scrolling-credit,制作',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues keeps rapid ASS lines that share a style but not an actor', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Dial_JP,アリス,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Dial_JP,ボブ,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Dial_JP,キャロル,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues keeps untagged rapid ASS repeats separate', () => {
+  // No overrides at all: timing-only evidence is an SRT/VTT fallback and must not apply
+  // to ASS, where the absence of typesetting is itself evidence of plain dialogue.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.05,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.05,0:00:01.10,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.10,0:00:01.15,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.15,0:00:01.20,Dial_JP,,0,0,0,,えっ',
+    'Dialogue: 0,0:00:01.20,0:00:01.25,Dial_JP,,0,0,0,,えっ',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 5);
+});
+
+test('parseSubtitleCues keeps repeated signs sharing one static clip', () => {
+  // `\clip` is a static shape for the event. Three events with the identical clip were
+  // typeset the same way, so none of them is a frame of the others.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Sign,,0,0,0,,{\\clip(0,0,100,100)}注意',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Sign,,0,0,0,,{\\clip(0,0,100,100)}注意',
+    'Dialogue: 0,0:00:01.40,0:00:01.60,Sign,,0,0,0,,{\\clip(0,0,100,100)}注意',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 3);
+});
+
+test('parseSubtitleCues collapses a sign animated through \\t', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Sign,,0,0,0,,{\\pos(10,10)\\t(0,200,\\frz30)}回る',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Sign,,0,0,0,,{\\pos(10,10)\\t(0,200,\\frz30)}回る',
+    'Dialogue: 0,0:00:01.40,0:00:03.00,Sign,,0,0,0,,{\\pos(10,10)\\t(0,200,\\frz30)}回る',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.endTime, 3.0);
+});
+
+test('parseSubtitleCues keeps a short repeated SRT run above the frame threshold', () => {
+  // Five contiguous 200ms cues: a sequence, but nowhere near animation-frame speed.
+  const lines: string[] = [];
+  for (let i = 0; i < 5; i++) {
+    const start = 1000 + i * 200;
+    const at = (ms: number) =>
+      `00:00:0${Math.floor(ms / 1000)},${String(ms % 1000).padStart(3, '0')}`;
+    lines.push(String(i + 1), `${at(start)} --> ${at(start + 200)}`, 'えっ', '');
+  }
+
+  const cues = parseSubtitleCues(lines.join('\n'), 'test.srt');
+
+  assert.equal(cues.length, 5);
+});
+
+test('parseSubtitleCues keeps a short SRT frame run below the minimum length', () => {
+  // Four 40ms frames: frame-speed, but too few to tell an animation from an artefact.
+  const lines: string[] = [];
+  for (let i = 0; i < 4; i++) {
+    const start = 7870 + i * 40;
+    const at = (ms: number) =>
+      `00:00:0${Math.floor(ms / 1000)},${String(ms % 1000).padStart(3, '0')}`;
+    lines.push(String(i + 1), `${at(start)} --> ${at(start + 40)}`, 'タイトル', '');
+  }
+
+  const cues = parseSubtitleCues(lines.join('\n'), 'test.srt');
+
+  assert.equal(cues.length, 4);
+});
+
+test('parseSubtitleCues applies ASS burst rules to ASS content behind an .srt filename', () => {
+  // The extension lies, so the SRT parser finds nothing and the content-sniffing fallback
+  // takes over -- which has to carry the `ass` source format with it, or the far stricter
+  // timing-only thresholds would let this karaoke burst through as three cues.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:01.20,Karaoke,,0,0,0,,{\\k20}歌詞',
+    'Dialogue: 0,0:00:01.20,0:00:01.40,Karaoke,,0,0,0,,{\\k20}歌詞',
+    'Dialogue: 0,0:00:01.40,0:00:03.00,Karaoke,,0,0,0,,{\\k20}歌詞',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.srt');
+
+  assert.equal(cues.length, 1);
+  assert.equal(cues[0]!.startTime, 1.0);
+  assert.equal(cues[0]!.endTime, 3.0);
+  assert.equal(cues[0]!.text, '歌詞');
+});
+
 test('parseSubtitleCues detects subtitle formats from remote URLs', () => {
  const assContent = [
    '[Events]',
@@ -1,9 +1,46 @@
+import {
+  assOverrideSignature,
+  assToPlainText,
+  collectAssOverrideCommands,
+  parseAssEffectField,
+  type AssEffectKind,
+  type AssOverrideCommand,
+} from './ass-text';
+import { mergeDuplicateCues } from './subtitle-cue-dedup';
+
 export interface SubtitleCue {
  startTime: number;
  endTime: number;
  text: string;
 }

+/**
+ * Everything the parser knows about a source event, shared only with the dedup engine.
+ * Deduplication needs the authoring context -- which style the line belongs to, which
+ * override commands it carries, whether the `Effect` column was set -- to tell a karaoke
+ * burst apart from two characters saying the same word in turn. None of it is meaningful
+ * outside the parser, so the public API stays `{startTime, endTime, text}`.
+ */
+export interface AnnotatedSubtitleCue extends SubtitleCue {
+  /** Text exactly as authored, override blocks and all. */
+  rawText: string;
+  style: string;
+  layer: number;
+  /** ASS `Name`/`Actor` column. */
+  name: string;
+  /** ASS `Effect` column, verbatim. */
+  effect: string;
+  effectKind: AssEffectKind;
+  /** Override commands found in `{...}` blocks, with their arguments. */
+  overrides: readonly AssOverrideCommand[];
+  /** Canonical form of `overrides`, for spotting values that change across a run. */
+  overrideSignature: string;
+  /** Position in the source file, so sorting by time stays deterministic across layers. */
+  order: number;
+}
+
+export type SubtitleSourceFormat = 'ass' | 'srt';
+
 const HTML_SUBTITLE_TAG_PATTERN = /<\/?[A-Za-z][^>\n]*>/g;

 const SRT_TIMING_PATTERN =
@@ -23,12 +60,21 @@ function parseTimestamp(
  );
 }

+/**
+ * The single ASS decode for the file path: cues leave the parser as plain text with real
+ * line breaks, matching what mpv hands over for the same line played live. No layer
+ * downstream decodes ASS again.
+ */
 function sanitizeSubtitleCueText(text: string): string {
-  return text.replace(ASS_OVERRIDE_TAG_PATTERN, '').replace(HTML_SUBTITLE_TAG_PATTERN, '').trim();
+  return assToPlainText(text, '\n').replace(HTML_SUBTITLE_TAG_PATTERN, '').trim();
 }

-export function parseSrtCues(content: string): SubtitleCue[] {
-  const cues: SubtitleCue[] = [];
+function toPublicCues(cues: AnnotatedSubtitleCue[]): SubtitleCue[] {
+  return cues.map(({ startTime, endTime, text }) => ({ startTime, endTime, text }));
+}
+
+function parseAnnotatedSrtCues(content: string): AnnotatedSubtitleCue[] {
+  const cues: AnnotatedSubtitleCue[] = [];
  const lines = content.split(/\r?\n/);
  let i = 0;

@@ -60,20 +106,39 @@ export function parseSrtCues(content: string): SubtitleCue[] {
      i += 1;
    }

-    const text = sanitizeSubtitleCueText(textLines.join('\n'));
+    const rawText = textLines.join('\n');
+    const text = sanitizeSubtitleCueText(rawText);
    if (text) {
-      cues.push({ startTime, endTime, text });
+      cues.push({
+        startTime,
+        endTime,
+        text,
+        rawText,
+        style: '',
+        layer: 0,
+        name: '',
+        effect: '',
+        effectKind: 'none',
+        // SRT and VTT carry no authoring metadata, and the dedup engine never reads
+        // overrides for those formats -- collecting them would be parsing for nobody.
+        overrides: [],
+        overrideSignature: '',
+        order: cues.length,
+      });
    }
  }

  return cues;
 }

-const ASS_OVERRIDE_TAG_PATTERN = /\{[^}]*\}/g;
+export function parseSrtCues(content: string): SubtitleCue[] {
+  return toPublicCues(parseAnnotatedSrtCues(content));
+}

 const ASS_TIMING_PATTERN = /^(\d+):(\d{2}):(\d{2})\.(\d{1,2})$/;
 const ASS_FORMAT_PREFIX = 'Format:';
 const ASS_DIALOGUE_PREFIX = 'Dialogue:';
+const ASS_NAME_FIELD_ALIASES = ['name', 'actor'];

 function parseAssTimestamp(raw: string): number | null {
  const match = ASS_TIMING_PATTERN.exec(raw.trim());
@@ -87,13 +152,43 @@ function parseAssTimestamp(raw: string): number | null {
  return hours * 3600 + minutes * 60 + seconds + centiseconds / 100;
 }

-export function parseAssCues(content: string): SubtitleCue[] {
-  const cues: SubtitleCue[] = [];
+function readField(fields: string[], index: number): string {
+  return index >= 0 && index < fields.length ? fields[index]!.trim() : '';
+}
+
+function findFieldIndex(formatFields: string[], aliases: string[]): number {
+  for (const alias of aliases) {
+    const index = formatFields.indexOf(alias);
+    if (index >= 0) {
+      return index;
+    }
+  }
+  return -1;
+}
+
+function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
+  const cues: AnnotatedSubtitleCue[] = [];
  const lines = content.split(/\r?\n/);
  let inEventsSection = false;
-  let startFieldIndex = -1;
-  let endFieldIndex = -1;
-  let textFieldIndex = -1;
+  const fieldIndex = {
+    start: -1,
+    end: -1,
+    text: -1,
+    style: -1,
+    layer: -1,
+    name: -1,
+    effect: -1,
+  };
+
+  const resetFieldIndex = () => {
+    fieldIndex.start = -1;
+    fieldIndex.end = -1;
+    fieldIndex.text = -1;
+    fieldIndex.style = -1;
+    fieldIndex.layer = -1;
+    fieldIndex.name = -1;
+    fieldIndex.effect = -1;
+  };

  for (const line of lines) {
    const trimmed = line.trim();
@@ -101,9 +196,7 @@ export function parseAssCues(content: string): SubtitleCue[] {
    if (trimmed.startsWith('[') && trimmed.endsWith(']')) {
      inEventsSection = trimmed.toLowerCase() === '[events]';
      if (!inEventsSection) {
-        startFieldIndex = -1;
-        endFieldIndex = -1;
-        textFieldIndex = -1;
+        resetFieldIndex();
      }
      continue;
    }
@@ -117,9 +210,15 @@ export function parseAssCues(content: string): SubtitleCue[] {
        .slice(ASS_FORMAT_PREFIX.length)
        .split(',')
        .map((field) => field.trim().toLowerCase());
-      startFieldIndex = formatFields.indexOf('start');
-      endFieldIndex = formatFields.indexOf('end');
-      textFieldIndex = formatFields.indexOf('text');
+      fieldIndex.start = formatFields.indexOf('start');
+      fieldIndex.end = formatFields.indexOf('end');
+      fieldIndex.text = formatFields.indexOf('text');
+      fieldIndex.style = formatFields.indexOf('style');
+      fieldIndex.layer = formatFields.indexOf('layer');
+      // Aegisub writes the speaker column as `Actor`; the v4+ spec calls it `Name`.
+      // Missing it costs the burst check its speaker guard, so both spellings count.
+      fieldIndex.name = findFieldIndex(formatFields, ASS_NAME_FIELD_ALIASES);
+      fieldIndex.effect = formatFields.indexOf('effect');
      continue;
    }

@@ -127,34 +226,57 @@ export function parseAssCues(content: string): SubtitleCue[] {
      continue;
    }

-    if (startFieldIndex < 0 || endFieldIndex < 0 || textFieldIndex < 0) {
+    if (fieldIndex.start < 0 || fieldIndex.end < 0 || fieldIndex.text < 0) {
      continue;
    }

    const fields = trimmed.slice(ASS_DIALOGUE_PREFIX.length).split(',');
    if (
-      startFieldIndex >= fields.length ||
-      endFieldIndex >= fields.length ||
-      textFieldIndex >= fields.length
+      fieldIndex.start >= fields.length ||
+      fieldIndex.end >= fields.length ||
+      fieldIndex.text >= fields.length
    ) {
      continue;
    }

-    const startTime = parseAssTimestamp(fields[startFieldIndex]!);
-    const endTime = parseAssTimestamp(fields[endFieldIndex]!);
+    const startTime = parseAssTimestamp(fields[fieldIndex.start]!);
+    const endTime = parseAssTimestamp(fields[fieldIndex.end]!);
    if (startTime === null || endTime === null) {
      continue;
    }

-    const text = sanitizeSubtitleCueText(fields.slice(textFieldIndex).join(','));
-    if (text) {
-      cues.push({ startTime, endTime, text });
+    const rawText = fields.slice(fieldIndex.text).join(',');
+    const text = sanitizeSubtitleCueText(rawText);
+    if (!text) {
+      continue;
    }
+
+    const effect = readField(fields, fieldIndex.effect);
+    const layer = Number(readField(fields, fieldIndex.layer));
+    const overrides = collectAssOverrideCommands(rawText);
+    cues.push({
+      startTime,
+      endTime,
+      text,
+      rawText,
+      style: readField(fields, fieldIndex.style),
+      layer: Number.isFinite(layer) ? layer : 0,
+      name: readField(fields, fieldIndex.name),
+      effect,
+      effectKind: parseAssEffectField(effect),
+      overrides,
+      overrideSignature: assOverrideSignature(overrides),
+      order: cues.length,
+    });
  }

  return cues;
 }

+export function parseAssCues(content: string): SubtitleCue[] {
+  return toPublicCues(parseAnnotatedAssCues(content));
+}
+
 function detectSubtitleFormat(source: string): 'srt' | 'vtt' | 'ass' | 'ssa' | null {
  const [normalizedSource = source] =
    (() => {
@@ -173,27 +295,31 @@ function detectSubtitleFormat(source: string): 'srt' | 'vtt' | 'ass' | 'ssa' | n

 export function parseSubtitleCues(content: string, filename: string): SubtitleCue[] {
  const format = detectSubtitleFormat(filename);
-  let cues: SubtitleCue[];
+  let cues: AnnotatedSubtitleCue[];
+  let sourceFormat: SubtitleSourceFormat = 'srt';

  switch (format) {
    case 'srt':
    case 'vtt':
-      cues = parseSrtCues(content);
+      cues = parseAnnotatedSrtCues(content);
      break;
    case 'ass':
    case 'ssa':
-      cues = parseAssCues(content);
+      cues = parseAnnotatedAssCues(content);
+      sourceFormat = 'ass';
      break;
    default:
      cues = [];
  }

  if (cues.length === 0) {
-    const assCues = parseAssCues(content);
-    const srtCues = parseSrtCues(content);
-    cues = assCues.length >= srtCues.length ? assCues : srtCues;
+    const assCues = parseAnnotatedAssCues(content);
+    const srtCues = parseAnnotatedSrtCues(content);
+    const preferAss = assCues.length >= srtCues.length;
+    cues = preferAss ? assCues : srtCues;
+    sourceFormat = preferAss && assCues.length > 0 ? 'ass' : 'srt';
  }

-  cues.sort((a, b) => a.startTime - b.startTime);
-  return cues;
+  cues.sort((a, b) => a.startTime - b.startTime || a.endTime - b.endTime || a.order - b.order);
+  return toPublicCues(mergeDuplicateCues(cues, sourceFormat));
 }
@@ -0,0 +1,204 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { createSubtitleLineDedupGate } from './subtitle-line-dedup-gate';
+import type { SubtitleCue } from '../../types';
+
+function karaokeFrames(text: string, start: number, frames: number, frameSeconds: number) {
+  return Array.from({ length: frames }, (_, index) => ({
+    text,
+    startSec: start + index * frameSeconds,
+    endSec: start + (index + 1) * frameSeconds,
+  }));
+}
+
+test('parsed cues drop the frames the sidebar already collapsed', () => {
+  // What `mergeDuplicateCues` leaves behind for a karaoke run: one cue over the run.
+  const cues: SubtitleCue[] = [
+    { startTime: 10, endTime: 14, text: '飛び上がる' },
+    { startTime: 14, endTime: 16, text: 'もしも' },
+  ];
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  const recorded = karaokeFrames('飛び上がる', 10, 40, 0.04).filter((sample) =>
+    gate.shouldRecord(sample),
+  );
+
+  assert.equal(recorded.length, 1);
+  assert.equal(recorded[0]!.startSec, 10);
+  assert.equal(gate.shouldRecord({ text: 'もしも', startSec: 14, endSec: 16 }), true);
+});
+
+test('parsed cues keep separate lines that merely repeat', () => {
+  const cues: SubtitleCue[] = [
+    { startTime: 3, endTime: 3.4, text: 'えっ' },
+    { startTime: 3.4, endTime: 3.9, text: 'えっ' },
+    { startTime: 3.9, endTime: 4.5, text: 'えっ' },
+  ];
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  const recorded = cues.filter((cue) =>
+    gate.shouldRecord({ text: cue.text, startSec: cue.startTime, endSec: cue.endTime }),
+  );
+
+  assert.equal(recorded.length, 3);
+});
+
+test('parsed cues outrank the streaming heuristic for short repeated cues', () => {
+  // Long enough to trip the timing-only rule, but the parser saw these with full
+  // lookahead and kept them, so every one of them is a line the sidebar shows.
+  const cues: SubtitleCue[] = Array.from({ length: 8 }, (_, index) => ({
+    startTime: 3 + index * 0.08,
+    endTime: 3 + (index + 1) * 0.08,
+    text: 'えっ',
+  }));
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  const recorded = cues.filter((cue) =>
+    gate.shouldRecord({ text: cue.text, startSec: cue.startTime, endSec: cue.endTime }),
+  );
+
+  assert.equal(recorded.length, 8);
+});
+
+test('parsed cues preserve legitimately separate cues only 40ms apart', () => {
+  const cues: SubtitleCue[] = Array.from({ length: 8 }, (_, index) => ({
+    startTime: 3 + index * 0.04,
+    endTime: 3 + (index + 1) * 0.04,
+    text: 'えっ',
+  }));
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  const recorded = cues.filter((cue) =>
+    gate.shouldRecord({ text: cue.text, startSec: cue.startTime, endSec: cue.endTime }),
+  );
+
+  assert.equal(recorded.length, 8);
+});
+
+test('a line whose timing does not match any cue still records', () => {
+  // A shifted track, an embedded sub nobody parsed: no match, no drop.
+  const cues: SubtitleCue[] = [{ startTime: 10, endTime: 14, text: '飛び上がる' }];
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  assert.equal(gate.shouldRecord({ text: '飛び上がる', startSec: 42, endSec: 44 }), true);
+});
+
+test('shifted parsed text falls back to streaming burst detection', () => {
+  const cues: SubtitleCue[] = [{ startTime: 10, endTime: 14, text: '飛び上がる' }];
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  const recorded = karaokeFrames('飛び上がる', 42, 40, 0.04).filter((sample) =>
+    gate.shouldRecord(sample),
+  );
+
+  assert.equal(recorded.length, 4);
+});
+
+test('replacing the parsed cue source forgets a streaming run', () => {
+  let cues: SubtitleCue[] = [];
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  karaokeFrames('飛び上がる', 42, 20, 0.04).forEach((sample) => gate.shouldRecord(sample));
+  // A new source publishes its own cue list; the old run must not carry over.
+  cues = [{ startTime: 100, endTime: 104, text: 'もしも' }];
+
+  assert.equal(gate.shouldRecord({ text: '飛び上がる', startSec: 42.8, endSec: 42.84 }), true);
+});
+
+test('without parsed cues a long run of identical short frames stops recording', () => {
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  const recorded = karaokeFrames('ひとしずく', 0, 200, 0.04).filter((sample) =>
+    gate.shouldRecord(sample),
+  );
+
+  assert.equal(recorded.length, 4);
+});
+
+test('without parsed cues interleaved dual-line karaoke stops recording per line', () => {
+  // Fansub OPs typically run two typeset lines at once -- kanji and romaji -- and mpv
+  // reports their frames interleaved. Each line must build its own run.
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  const samples = Array.from({ length: 40 }, (_, index) => {
+    const start = 10 + Math.floor(index / 2) * 0.06;
+    return index % 2 === 0
+      ? { text: '歌詞', startSec: start, endSec: start + 0.06 }
+      : { text: 'kashi', startSec: start + 0.001, endSec: start + 0.061 };
+  });
+  const recorded = samples.filter((sample) => gate.shouldRecord(sample));
+
+  assert.equal(recorded.filter((sample) => sample.text === '歌詞').length, 4);
+  assert.equal(recorded.filter((sample) => sample.text === 'kashi').length, 4);
+});
+
+test('interleaved dialogue between two speakers keeps recording', () => {
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  // Two characters trading normal-length lines back and forth.
+  const samples = Array.from({ length: 12 }, (_, index) => {
+    const start = 5 + index * 0.7;
+    return {
+      text: index % 2 === 0 ? 'えっ' : 'なに',
+      startSec: start,
+      endSec: start + 0.7,
+    };
+  });
+  const recorded = samples.filter((sample) => gate.shouldRecord(sample));
+
+  assert.equal(recorded.length, 12);
+});
+
+test('without parsed cues ordinary repeated dialogue keeps recording', () => {
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  // Six contiguous `えっ`, each held for a normal beat rather than an animation frame.
+  const recorded = karaokeFrames('えっ', 0, 6, 0.6).filter((sample) => gate.shouldRecord(sample));
+
+  assert.equal(recorded.length, 6);
+});
+
+test('the same event offered twice does not advance the run', () => {
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  // mpv fires the timing handler once for `sub-start` and once for `sub-end`.
+  for (let i = 0; i < 8; i += 1) {
+    assert.equal(gate.shouldRecord({ text: '待って', startSec: 5, endSec: 5.05 }), true);
+  }
+});
+
+test('a gap between frames starts a new run', () => {
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  const first = karaokeFrames('もし', 0, 6, 0.04).filter((sample) => gate.shouldRecord(sample));
+  const second = karaokeFrames('もし', 30, 6, 0.04).filter((sample) => gate.shouldRecord(sample));
+
+  assert.equal(first.length, 4);
+  assert.equal(second.length, 4);
+});
+
+test('reset forgets the streaming run', () => {
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => null });
+
+  karaokeFrames('もし', 0, 20, 0.04).forEach((sample) => gate.shouldRecord(sample));
+  gate.reset();
+
+  assert.equal(gate.shouldRecord({ text: 'もし', startSec: 0.8, endSec: 0.84 }), true);
+});
+
+test('reset ignores stale parsed cues until the source publishes a new cue list', () => {
+  let cues: SubtitleCue[] = [{ startTime: 10, endTime: 14, text: '飛び上がる' }];
+  const gate = createSubtitleLineDedupGate({ getParsedCues: () => cues });
+
+  assert.equal(gate.shouldRecord({ text: '飛び上がる', startSec: 10, endSec: 10.04 }), true);
+  gate.reset();
+
+  const recordedWithStaleCues = karaokeFrames('飛び上がる', 10.04, 8, 0.04).filter((sample) =>
+    gate.shouldRecord(sample),
+  );
+  assert.equal(recordedWithStaleCues.length, 4);
+
+  cues = [{ startTime: 20, endTime: 24, text: '飛び上がる' }];
+  assert.equal(gate.shouldRecord({ text: '飛び上がる', startSec: 20, endSec: 20.04 }), true);
+  assert.equal(gate.shouldRecord({ text: '飛び上がる', startSec: 20.04, endSec: 20.08 }), false);
+});
@@ -0,0 +1,241 @@
+/*
+ * Decides which live mpv subtitle lines reach the immersion stats.
+ *
+ * The sidebar reads a parsed subtitle file, so it can collapse an animation burst with
+ * full lookahead (`subtitle-cue-dedup`). Stats are fed from mpv's `sub-start`/`sub-end`
+ * properties instead -- one event per animation frame, each with its own start time --
+ * so without a gate a karaoke OP counts its lyrics once per frame and buries every real
+ * word in the vocabulary charts.
+ *
+ * Two layers, in order:
+ *
+ * 1. When the active source has been parsed, its cue list has *already* been collapsed.
+ *    A live line that lands inside a surviving cue of the same text, but after that
+ *    cue's start, is a frame the sidebar merged away, so stats drop it too. This is the
+ *    layer that keeps the two views consistent by construction.
+ * 2. Otherwise (embedded track nobody parsed, a source whose timings mpv has shifted)
+ *    fall back to timing alone. No authoring metadata is available live -- mpv delivers
+ *    `sub-text-ass` after `sub-start`/`sub-end`, so any ASS text read here belongs to the
+ *    previous event -- which puts this layer in the same position as the SRT path in
+ *    `subtitle-cue-dedup`, and it uses that path's deliberately strict bounds.
+ */
+
+import { normalizePlainSubtitleText } from './ass-text';
+import {
+  DUPLICATE_CUE_GAP_TOLERANCE_SECONDS,
+  MIN_TIMING_ONLY_FRAMES,
+  TIMING_ONLY_FRAME_MAX_SECONDS,
+} from './subtitle-burst-constants';
+import type { SubtitleCue } from './subtitle-cue-parser';
+
+export interface SubtitleLineSample {
+  text: string;
+  startSec: number;
+  endSec: number;
+}
+
+export interface SubtitleLineDedupGateDeps {
+  /** Cues for the active source, already collapsed by the parser. */
+  getParsedCues: () => readonly SubtitleCue[] | null | undefined;
+}
+
+export interface SubtitleLineDedupGate {
+  /** False when this line is an animation frame of a line already recorded. */
+  shouldRecord: (sample: SubtitleLineSample) => boolean;
+  /** Forget run state and ignore the current cue list until its source is replaced. */
+  reset: () => void;
+}
+
+interface CueSpan {
+  startTime: number;
+  endTime: number;
+}
+
+interface StreamingRunState {
+  startMs: number;
+  chainEndSec: number;
+  /** Contiguous identical short frames seen so far, including the recorded first one. */
+  frames: number;
+}
+
+/**
+ * Dual-line karaoke interleaves two texts frame by frame, so runs are tracked per text.
+ * Dead runs are pruned as playback moves past them; the cap only matters after a
+ * backward seek leaves runs whose ends sit ahead of the new position.
+ */
+const MAX_ACTIVE_STREAMING_RUNS = 32;
+
+/** Exact cue identity, separate from the looser tolerance used to chain adjacent frames. */
+const CUE_START_IDENTITY_TOLERANCE_SECONDS = 0.005;
+
+function normalizeLineText(text: string): string {
+  return normalizePlainSubtitleText(text, { collapseLineBreaks: true });
+}
+
+function buildSpansByText(cues: readonly SubtitleCue[]): Map<string, CueSpan[]> {
+  const spansByText = new Map<string, CueSpan[]>();
+  for (const cue of cues) {
+    const key = normalizeLineText(cue.text);
+    if (!key) continue;
+    const span = { startTime: cue.startTime, endTime: cue.endTime };
+    const existing = spansByText.get(key);
+    if (existing) {
+      existing.push(span);
+    } else {
+      spansByText.set(key, [span]);
+    }
+  }
+  return spansByText;
+}
+
+/**
+ * A frame the parser merged away: the same text, starting inside a surviving cue but
+ * after it began.
+ *
+ * Starting a cue always wins over falling inside one. The first frame of a collapsed run
+ * starts *at* the merged cue, and a line the parser deliberately kept separate -- three
+ * characters trading `えっ` back to back -- begins exactly where the one before it ends.
+ */
+function isMergedAwayFrame(spans: readonly CueSpan[], startSec: number): boolean | null {
+  const coveringSpans = spans.filter(
+    (span) =>
+      startSec >= span.startTime - CUE_START_IDENTITY_TOLERANCE_SECONDS &&
+      startSec <= span.endTime + CUE_START_IDENTITY_TOLERANCE_SECONDS,
+  );
+  if (coveringSpans.length === 0) {
+    return null;
+  }
+  const startsOwnCue = spans.some(
+    (span) => Math.abs(startSec - span.startTime) <= CUE_START_IDENTITY_TOLERANCE_SECONDS,
+  );
+  if (startsOwnCue) {
+    return false;
+  }
+  return coveringSpans.some(
+    (span) =>
+      startSec > span.startTime + CUE_START_IDENTITY_TOLERANCE_SECONDS &&
+      startSec <= span.endTime + CUE_START_IDENTITY_TOLERANCE_SECONDS,
+  );
+}
+
+export function createSubtitleLineDedupGate(
+  deps: SubtitleLineDedupGateDeps,
+): SubtitleLineDedupGate {
+  let indexedCues: readonly SubtitleCue[] | null | undefined;
+  let ignoredCuesAfterReset: readonly SubtitleCue[] | null | undefined;
+  let spansByText: Map<string, CueSpan[]> = new Map();
+  const runs = new Map<string, StreamingRunState>();
+
+  const lookupSpans = (text: string): CueSpan[] | null => {
+    const cues = deps.getParsedCues() ?? null;
+    if (ignoredCuesAfterReset !== undefined) {
+      if (cues === ignoredCuesAfterReset) {
+        return null;
+      }
+      ignoredCuesAfterReset = undefined;
+    }
+    if (cues !== indexedCues) {
+      indexedCues = cues;
+      spansByText = cues?.length ? buildSpansByText(cues) : new Map();
+      runs.clear();
+    }
+    return spansByText.get(text) ?? null;
+  };
+
+  /**
+   * A run this sample cannot continue is a run no later sample can continue either --
+   * continuation needs a start inside the running end plus tolerance, and starts only
+   * move forward outside of seeks.
+   */
+  const pruneDeadRuns = (startSec: number): void => {
+    for (const [text, state] of runs) {
+      if (state.chainEndSec + DUPLICATE_CUE_GAP_TOLERANCE_SECONDS < startSec) {
+        runs.delete(text);
+      }
+    }
+  };
+
+  /**
+   * Timing-only burst detection over a stream. Without lookahead the run can only be
+   * recognised from the inside, so the first frames of a burst are recorded and the rest
+   * dropped -- an OP costs a handful of counted lines instead of several hundred.
+   */
+  const advanceStreamingRun = (text: string, sample: SubtitleLineSample): boolean => {
+    pruneDeadRuns(sample.startSec);
+    const startMs = Math.round(sample.startSec * 1000);
+    const run = runs.get(text);
+    // mpv reports `sub-start` and `sub-end` separately, so one event can be offered
+    // twice. The same start is the same frame, never the next one in a run.
+    if (run && run.startMs === startMs) {
+      run.chainEndSec = Math.max(run.chainEndSec, sample.endSec);
+      return run.frames < MIN_TIMING_ONLY_FRAMES;
+    }
+
+    const isShortFrame = sample.endSec - sample.startSec < TIMING_ONLY_FRAME_MAX_SECONDS;
+    // Frames are authored flush against each other, but typesetters do overlap them, so
+    // the chain only requires forward progress that stays inside the running end.
+    const continuesRun =
+      run !== undefined &&
+      isShortFrame &&
+      startMs > run.startMs &&
+      sample.startSec <= run.chainEndSec + DUPLICATE_CUE_GAP_TOLERANCE_SECONDS;
+
+    if (continuesRun && run) {
+      run.startMs = startMs;
+      run.chainEndSec = Math.max(run.chainEndSec, sample.endSec);
+      run.frames += 1;
+      return run.frames < MIN_TIMING_ONLY_FRAMES;
+    }
+
+    const fresh: StreamingRunState = {
+      startMs,
+      chainEndSec: sample.endSec,
+      frames: isShortFrame ? 1 : 0,
+    };
+    runs.set(text, fresh);
+    if (runs.size > MAX_ACTIVE_STREAMING_RUNS) {
+      let oldestText: string | undefined;
+      let oldestEnd = Infinity;
+      for (const [runText, state] of runs) {
+        if (runText !== text && state.chainEndSec < oldestEnd) {
+          oldestEnd = state.chainEndSec;
+          oldestText = runText;
+        }
+      }
+      if (oldestText !== undefined) {
+        runs.delete(oldestText);
+      }
+    }
+    return fresh.frames < MIN_TIMING_ONLY_FRAMES;
+  };
+
+  return {
+    shouldRecord: (sample) => {
+      const text = normalizeLineText(sample.text);
+      if (!text) {
+        return true;
+      }
+
+      // The parsed cue list has the final say wherever it covers this line. Falling
+      // through to the streaming heuristic would let it drop cues the parser looked at
+      // with full lookahead and deliberately kept apart, which is the disagreement
+      // between sidebar and stats this gate exists to prevent.
+      const spans = lookupSpans(text);
+      if (spans) {
+        const mergedAway = isMergedAwayFrame(spans, sample.startSec);
+        if (mergedAway !== null) {
+          runs.delete(text);
+          return !mergedAway;
+        }
+      }
+
+      return advanceStreamingRun(text, sample);
+    },
+    reset: () => {
+      runs.clear();
+      ignoredCuesAfterReset = deps.getParsedCues() ?? null;
+      indexedCues = undefined;
+      spansByText = new Map();
+    },
+  };
+}
@@ -115,6 +115,21 @@ test('subtitle processing does not emit plain payload for cached lines', async (
  assert.deepEqual(emitted, [{ text: '字幕', tokens: [] }]);
 });

+test('text that normalizes to nothing is never cached', () => {
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: () => {},
+  });
+
+  // Two different inputs both reduce to an empty key; sharing one entry would serve the
+  // first one's tokens for the second.
+  controller.preCacheTokenization('  ', { text: '  ', tokens: [] });
+
+  assert.equal(controller.hasCachedSubtitle('  '), false);
+  assert.equal(controller.hasCachedSubtitle('\\n'), false);
+  assert.equal(controller.consumeCachedSubtitle('\\n'), null);
+});
+
 test('subtitle processing shows plain line while tokenization is still pending', async () => {
  const emitted: SubtitleData[] = [];
  let resolveTokenization: ((value: SubtitleData | null) => void) | undefined;
@@ -539,3 +554,125 @@ test('default cache limit covers a full-length title without evicting', () => {
  assert.equal(controller.hasCachedSubtitle('line-0'), true);
  assert.equal(controller.hasCachedSubtitle('line-1999'), true);
 });
+
+test('onSubtitleChange reports whether processing was scheduled', async () => {
+  const emitted: SubtitleData[] = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  // New text schedules work, so an emit (and anything gated on it) will follow.
+  assert.equal(controller.onSubtitleChange('字幕'), true);
+  await flushMicrotasks();
+
+  // A repeat emits nothing, so callers must not wait on an emit that is never
+  // coming (subtitle prefetching would stay paused for the rest of the cue).
+  const emittedCount = emitted.length;
+  assert.equal(controller.onSubtitleChange('字幕'), false);
+  await flushMicrotasks();
+  assert.equal(emitted.length, emittedCount);
+});
+
+test('refreshCurrentSubtitle reports the empty-text emit that an in-flight run will deliver', async () => {
+  const emitted: SubtitleData[] = [];
+  let resolveFirst: ((value: SubtitleData | null) => void) | undefined;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      if (text === '字幕') {
+        return await new Promise<SubtitleData | null>((resolve) => {
+          resolveFirst = resolve;
+        });
+      }
+      return { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('字幕');
+  await flushMicrotasks();
+
+  // Clearing the subtitle while tokenization is in flight: the running loop
+  // picks the empty text up and emits it, so callers gated on that emit (the
+  // prefetch pause) must be told one is coming.
+  assert.equal(controller.refreshCurrentSubtitle(''), true);
+
+  resolveFirst?.({ text: '字幕', tokens: [] });
+  await flushMicrotasks();
+  await flushMicrotasks();
+  // '字幕' is the provisional plain emit the in-flight run already made before
+  // the refresh; '' is the emit the refresh promised.
+  assert.deepEqual(
+    emitted.map((payload) => payload.text),
+    ['字幕', ''],
+  );
+});
+
+test('onProcessingSettled fires once after the queue drains, including runs that emit nothing', async () => {
+  const events: string[] = [];
+  let resolveFirst: ((value: SubtitleData | null) => void) | undefined;
+  let tokenizationFails = false;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      if (tokenizationFails) {
+        return null;
+      }
+      if (text === '一行目') {
+        return await new Promise<SubtitleData | null>((resolve) => {
+          resolveFirst = resolve;
+        });
+      }
+      return { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => events.push(`emit:${payload.text}`),
+    onProcessingSettled: () => events.push('settled'),
+  });
+
+  controller.onSubtitleChange('一行目');
+  await flushMicrotasks();
+  // A second line arrives before the first finishes: the controller still has
+  // work, so it must not report itself settled between the two.
+  controller.onSubtitleChange('二行目');
+  resolveFirst?.({ text: '一行目', tokens: [] });
+  await flushMicrotasks();
+  await flushMicrotasks();
+
+  assert.deepEqual(events, ['emit:一行目', 'emit:二行目', 'emit:二行目', 'settled']);
+
+  // Tokenization failure on a line already shown plain: nothing is emitted, and
+  // the settle signal is the only way a caller learns the work is over.
+  events.length = 0;
+  tokenizationFails = true;
+  controller.invalidateTokenizationCache();
+  assert.equal(controller.refreshCurrentSubtitle('二行目'), true);
+  await flushMicrotasks();
+  await flushMicrotasks();
+  assert.deepEqual(events, ['settled']);
+});
+
+test('notePlainSubtitleEmitted suppresses the controller repeat of a payload already shown', async () => {
+  const emitted: SubtitleData[] = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  // Autoplay priming paints the plain line itself, then asks for tokenization.
+  controller.notePlainSubtitleEmitted('字幕');
+  controller.refreshCurrentSubtitle('字幕');
+  await flushMicrotasks();
+
+  assert.deepEqual(emitted, [{ text: '字幕', tokens: [] }]);
+});
+
+test('refreshCurrentSubtitle reports no emit for empty text when nothing is running', async () => {
+  const emitted: SubtitleData[] = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  assert.equal(controller.refreshCurrentSubtitle(''), false);
+  await flushMicrotasks();
+  assert.deepEqual(emitted, []);
+});
@@ -1,8 +1,17 @@
 import type { SubtitleData } from '../../types';
+import { normalizePlainSubtitleText } from './ass-text';

 export interface SubtitleProcessingControllerDeps {
  tokenizeSubtitle: (text: string) => Promise<SubtitleData | null>;
  emitSubtitle: (payload: SubtitleData) => void;
+  /**
+   * Fires when the controller runs out of work: every scheduled line has been
+   * processed, whether it ended in an emit, a suppressed duplicate, or a
+   * tokenizer failure. Callers that hold a resource for the duration of
+   * processing (prefetch pausing) release it here rather than on an emit,
+   * which is not guaranteed to happen.
+   */
+  onProcessingSettled?: () => void;
  logDebug?: (message: string) => void;
  now?: () => number;
  cacheLimit?: number;
@@ -17,16 +26,37 @@ export interface SubtitleProcessingControllerDeps {
 export const DEFAULT_SUBTITLE_TOKENIZATION_CACHE_LIMIT = 2500;

 export interface SubtitleProcessingController {
-  onSubtitleChange: (text: string) => void;
-  refreshCurrentSubtitle: (textOverride?: string) => void;
+  /**
+   * Returns whether processing is now scheduled or already in flight for this
+   * event. A false return means the controller is idle and will do nothing, so
+   * onProcessingSettled will not fire; callers that pause work for the duration
+   * of processing (such as subtitle prefetching) must release it themselves.
+   */
+  onSubtitleChange: (text: string) => boolean;
+  /** Same contract as onSubtitleChange: whether processing is pending. */
+  refreshCurrentSubtitle: (textOverride?: string) => boolean;
+  /**
+   * Records that this exact text has already been shown plain by someone else
+   * (autoplay priming paints its first frame before scheduling tokenization),
+   * so the controller does not repeat that payload on its way to the tokenized
+   * one.
+   */
+  notePlainSubtitleEmitted: (text: string) => void;
  invalidateTokenizationCache: () => void;
  preCacheTokenization: (text: string, data: SubtitleData) => void;
  consumeCachedSubtitle: (text: string) => SubtitleData | null;
  hasCachedSubtitle: (text: string) => boolean;
 }

+/**
+ * Prefetched cues and live mpv text are both already decoded from ASS, so the key only
+ * has to settle whitespace for one authored line to resolve to one entry.
+ *
+ * An empty key is not a line: it is whatever normalization reduced to nothing. Callers
+ * must skip the cache for it rather than let every such input share one entry.
+ */
 export function normalizeSubtitleCacheKey(text: string): string {
-  return text.replace(/\r\n/g, '\n').replace(/\\N/g, '\n').replace(/\\n/g, '\n').trim();
+  return normalizePlainSubtitleText(text);
 }

 export function createSubtitleProcessingController(
@@ -50,6 +80,9 @@ export function createSubtitleProcessingController(

  const getCachedTokenization = (text: string): SubtitleData | null => {
    const cacheKey = normalizeSubtitleCacheKey(text);
+    if (!cacheKey) {
+      return null;
+    }
    const cached = tokenizationCache.get(cacheKey);
    if (!cached) {
      return null;
@@ -61,7 +94,11 @@ export function createSubtitleProcessingController(
  };

  const setCachedTokenization = (text: string, payload: SubtitleData): void => {
-    tokenizationCache.set(normalizeSubtitleCacheKey(text), payload);
+    const cacheKey = normalizeSubtitleCacheKey(text);
+    if (!cacheKey) {
+      return;
+    }
+    tokenizationCache.set(cacheKey, payload);
    while (tokenizationCache.size > SUBTITLE_TOKENIZATION_CACHE_LIMIT) {
      const firstKey = tokenizationCache.keys().next().value;
      if (firstKey !== undefined) {
@@ -164,14 +201,20 @@ export function createSubtitleProcessingController(
          (latestText.trim() && cacheGeneration !== lastEmittedGeneration)
        ) {
          processLatest();
+          return;
        }
+        // Nothing left to do: signal completion even when this run emitted
+        // nothing (suppressed duplicate, tokenizer failure), or callers waiting
+        // on the controller would wait forever.
+        deps.onProcessingSettled?.();
      });
  };

  return {
    onSubtitleChange: (text: string) => {
      if (text === latestText) {
-        return;
+        // A run already in flight for this text will still emit for it.
+        return processing;
      }
      latestText = text;
      if (
@@ -183,21 +226,28 @@ export function createSubtitleProcessingController(
        lastPlainEmittedText = text;
      }
      processLatest();
+      return true;
    },
    refreshCurrentSubtitle: (textOverride?: string) => {
      if (typeof textOverride === 'string') {
        latestText = textOverride;
      }
      if (!latestText.trim()) {
-        return;
+        // A run in flight will pick this up and emit the empty subtitle, so
+        // the caller is still waiting on an emit.
+        return processing;
      }
-      if (
-        processing ||
-        (latestText === lastEmittedText && cacheGeneration === lastEmittedGeneration)
-      ) {
-        return;
+      if (processing) {
+        return true;
+      }
+      if (latestText === lastEmittedText && cacheGeneration === lastEmittedGeneration) {
+        return false;
      }
      processLatest();
+      return true;
+    },
+    notePlainSubtitleEmitted: (text: string) => {
+      lastPlainEmittedText = text;
    },
    invalidateTokenizationCache: () => {
      tokenizationCache.clear();
@@ -219,7 +269,8 @@ export function createSubtitleProcessingController(
      return cached;
    },
    hasCachedSubtitle: (text: string) => {
-      return tokenizationCache.has(normalizeSubtitleCacheKey(text));
+      const cacheKey = normalizeSubtitleCacheKey(text);
+      return cacheKey.length > 0 && tokenizationCache.has(cacheKey);
    },
  };
 }
@@ -1651,9 +1651,11 @@ test('tokenizeSubtitle clears JLPT level from standalone Yomitan particle token'
  assert.equal(result.tokens?.[0]?.jlptLevel, undefined);
 });

-test('tokenizeSubtitle returns null tokens for empty normalized text', async () => {
+test('tokenizeSubtitle returns the normalized text when it comes out empty', async () => {
+  // Handing back the original would push whatever normalization dropped into app state
+  // as if it were subtitle text.
  const result = await tokenizeSubtitle(' \\n  ', makeDeps());
-  assert.deepEqual(result, { text: ' \\n  ', tokens: null });
+  assert.deepEqual(result, { text: '', tokens: null });
 });

 test('tokenizeSubtitle normalizes newlines before Yomitan parse request', async () => {
@@ -2934,44 +2936,12 @@ test('tokenizeSubtitle preserves Yomitan compound token when MeCab components ar
                return [];
              }

-              if (script.includes('parseText')) {
-                return [
-                  {
-                    source: 'scanning-parser',
-                    index: 0,
-                    content: [
-                      [
-                        {
-                          text: '取り組んで',
-                          reading: 'とりくんで',
-                          headwords: [[{ term: '取り組む' }]],
-                        },
-                      ],
-                      [
-                        {
-                          text: 'もらいます',
-                          reading: 'もらいます',
-                          headwords: [[{ term: 'もらう' }]],
-                        },
-                      ],
-                    ],
-                  },
-                ];
-              }
-
              return [
                {
-                  surface: '取り',
-                  reading: 'とり',
-                  headword: '取る',
+                  surface: '取り組んで',
+                  reading: 'とりくんで',
+                  headword: '取り組む',
                  startPos: 0,
-                  endPos: 2,
-                },
-                {
-                  surface: '組んで',
-                  reading: 'くんで',
-                  headword: '組む',
-                  startPos: 2,
                  endPos: 5,
                },
                {
@@ -27,6 +27,7 @@ import {
 } from './tokenizer/yomitan-parser-runtime';
 import type { YomitanTermFrequency } from './tokenizer/yomitan-parser-runtime';
 import { isKanaChar } from './tokenizer/token-classification';
+import { normalizePlainSubtitleText } from './ass-text';

 const logger = createLogger('main:tokenizer');

@@ -70,6 +71,7 @@ export interface TokenizerServiceDeps {
  getNameMatchImagesEnabled?: () => boolean;
  getCharacterNameImage?: (term: string) => CharacterNameImage | null;
  getCurrentCharacterDictionaryMediaId?: () => number | null;
+  getCharacterNameCandidates?: () => { key: string; forms: string[] } | null;
  getFrequencyDictionaryEnabled?: () => boolean;
  getFrequencyDictionaryMatchMode?: () => FrequencyDictionaryMatchMode;
  getFrequencyRank?: FrequencyDictionaryLookup;
@@ -106,6 +108,7 @@ export interface TokenizerDepsRuntimeOptions {
  getNameMatchImagesEnabled?: () => boolean;
  getCharacterNameImage?: (term: string) => CharacterNameImage | null;
  getCurrentCharacterDictionaryMediaId?: () => number | null;
+  getCharacterNameCandidates?: () => { key: string; forms: string[] } | null;
  getFrequencyDictionaryEnabled?: () => boolean;
  getFrequencyDictionaryMatchMode?: () => FrequencyDictionaryMatchMode;
  getFrequencyRank?: FrequencyDictionaryLookup;
@@ -266,6 +269,7 @@ export function createTokenizerDepsRuntime(
    getNameMatchImagesEnabled: options.getNameMatchImagesEnabled,
    getCharacterNameImage: options.getCharacterNameImage,
    getCurrentCharacterDictionaryMediaId: options.getCurrentCharacterDictionaryMediaId,
+    getCharacterNameCandidates: options.getCharacterNameCandidates,
    getFrequencyDictionaryEnabled: options.getFrequencyDictionaryEnabled,
    getFrequencyDictionaryMatchMode: options.getFrequencyDictionaryMatchMode ?? (() => 'headword'),
    getFrequencyRank: options.getFrequencyRank,
@@ -716,15 +720,30 @@ function getAnnotationOptions(deps: TokenizerServiceDeps): TokenizerAnnotationOp
  };
 }

+// Per-line stage durations for the pipeline debug log; every field is filled in
+// by the stage that awaits the corresponding work.
+interface TokenizationStageTimings {
+  scanMs?: number;
+  mecabMs?: number;
+  frequencyMs?: number;
+  annotateMs?: number;
+}
+
 async function parseWithYomitanInternalParser(
  text: string,
  deps: TokenizerServiceDeps,
  options: TokenizerAnnotationOptions,
+  stageTimings?: TokenizationStageTimings,
 ): Promise<MergedToken[] | null> {
+  const scanStartedAtMs = Date.now();
  const selectedTokens = await requestYomitanScanTokens(text, deps, logger, {
    includeNameMatchMetadata: options.nameMatchEnabled,
    currentCharacterDictionaryMediaId: deps.getCurrentCharacterDictionaryMediaId?.() ?? null,
+    nameCandidates: deps.getCharacterNameCandidates?.() ?? null,
  });
+  if (stageTimings) {
+    stageTimings.scanMs = Date.now() - scanStartedAtMs;
+  }
  if (!selectedTokens || selectedTokens.length === 0) {
    return null;
  }
@@ -757,6 +776,7 @@ async function parseWithYomitanInternalParser(

  const frequencyRankPromise: Promise<YomitanFrequencyIndex> = options.frequencyEnabled
    ? (async () => {
+        const frequencyStartedAtMs = Date.now();
        const frequencyMatchMode = options.frequencyMatchMode;
        const termReadingList = buildYomitanFrequencyTermReadingList(
          normalizedSelectedTokens,
@@ -767,12 +787,17 @@ async function parseWithYomitanInternalParser(
          deps,
          logger,
        );
-        return buildYomitanFrequencyIndex(yomitanFrequencies);
+        const frequencyIndex = buildYomitanFrequencyIndex(yomitanFrequencies);
+        if (stageTimings) {
+          stageTimings.frequencyMs = Date.now() - frequencyStartedAtMs;
+        }
+        return frequencyIndex;
      })()
    : Promise.resolve({ byPair: new Map(), byTerm: new Map() });

  const mecabEnrichmentPromise: Promise<MergedToken[]> = needsMecabPosEnrichment(options)
    ? (async () => {
+        const mecabStartedAtMs = Date.now();
        try {
          const mecabTokens = await deps.tokenizeWithMecab(text);
          const enrichTokensWithMecab = deps.enrichTokensWithMecab ?? enrichTokensWithMecabAsync;
@@ -786,6 +811,10 @@ async function parseWithYomitanInternalParser(
            `textLength=${text.length}`,
          );
          return normalizedSelectedTokens;
+        } finally {
+          if (stageTimings) {
+            stageTimings.mecabMs = Date.now() - mecabStartedAtMs;
+          }
        }
      })()
    : Promise.resolve(normalizedSelectedTokens);
@@ -858,14 +887,14 @@ export async function tokenizeSubtitle(
  text: string,
  deps: TokenizerServiceDeps,
 ): Promise<SubtitleData> {
-  const displayText = text
-    .replace(/\r\n/g, '\n')
-    .replace(/\\N/g, '\n')
-    .replace(/\\n/g, '\n')
-    .trim();
+  const displayText = normalizePlainSubtitleText(text);

+  // ASS decoding already happened upstream (cue parser for files, mpv for live text), so
+  // all this drops is whitespace -- but a whitespace-only line still normalizes to empty.
+  // Return the normalized form anyway: handing back the original would put a blank line
+  // into application state as if it were subtitle text.
  if (!displayText) {
-    return { text, tokens: null };
+    return { text: displayText, tokens: null };
  }

  const tokenizeText = displayText
@@ -876,15 +905,35 @@ export async function tokenizeSubtitle(
  const annotationOptions = getAnnotationOptions(deps);
  annotationOptions.sourceText = tokenizeText;

-  const yomitanTokens = await parseWithYomitanInternalParser(tokenizeText, deps, annotationOptions);
+  const stageTimings: TokenizationStageTimings = {};
+  const startedAtMs = Date.now();
+  const logStageTimings = (tokenCount: number): void => {
+    logger.debug(
+      `Subtitle tokenization stages; textLength=${tokenizeText.length}, tokenCount=${tokenCount}, ` +
+        `scanMs=${stageTimings.scanMs ?? '-'}, mecabMs=${stageTimings.mecabMs ?? '-'}, ` +
+        `frequencyMs=${stageTimings.frequencyMs ?? '-'}, annotateMs=${stageTimings.annotateMs ?? '-'}, ` +
+        `totalMs=${Date.now() - startedAtMs}`,
+    );
+  };
+
+  const yomitanTokens = await parseWithYomitanInternalParser(
+    tokenizeText,
+    deps,
+    annotationOptions,
+    stageTimings,
+  );
  if (yomitanTokens && yomitanTokens.length > 0) {
+    const annotateStartedAtMs = Date.now();
    const annotatedTokens = await applyAnnotationStage(yomitanTokens, deps, annotationOptions);
+    stageTimings.annotateMs = Date.now() - annotateStartedAtMs;
    const renderedTokens = applyCharacterNameImages(annotatedTokens, deps, annotationOptions);
+    logStageTimings(renderedTokens.length);
    return {
      text: displayText,
      tokens: renderedTokens.length > 0 ? renderedTokens : null,
    };
  }

+  logStageTimings(0);
  return { text: displayText, tokens: null };
 }
@@ -0,0 +1,4 @@
+// Title prefix of the dictionaries SubMiner generates per media. Lives on its
+// own because both the main process and the injected scan runtime match on it,
+// and the injected fragments interpolate it into their own source.
+export const CHARACTER_DICTIONARY_TITLE_PREFIX = 'SubMiner Character Dictionary';
@@ -366,8 +366,11 @@ export function createReplayMessageStore(messages: GoldenRecordedMessage[]): Rep
  };
 }

-async function runInjectedScriptInVm(script: string, store: ReplayMessageStore): Promise<unknown> {
-  return await vm.runInNewContext(script, {
+// One persistent context per fixture, matching the real parser window: the
+// scan runtime installs itself once into globalThis and later per-line call
+// scripts reuse it.
+function createInjectedScriptVm(store: ReplayMessageStore): (script: string) => Promise<unknown> {
+  const context = vm.createContext({
    chrome: {
      runtime: {
        lastError: null,
@@ -393,6 +396,7 @@ async function runInjectedScriptInVm(script: string, store: ReplayMessageStore):
    Set,
    String,
  });
+  return async (script: string) => await vm.runInContext(script, context);
 }

 export function createReplayTokenizerDeps(fixture: GoldenFixture): TokenizerServiceDeps {
@@ -400,13 +404,14 @@ export function createReplayTokenizerDeps(fixture: GoldenFixture): TokenizerServ
  const scriptResults = new Map(
    fixture.recording.scripts.map((entry) => [entry.sha256, entry] as const),
  );
+  const runInjectedScriptInVm = createInjectedScriptVm(store);

  const parserWindow = {
    isDestroyed: () => false,
    webContents: {
      executeJavaScript: async (script: string) => {
        try {
-          return await runInjectedScriptInVm(script, store);
+          return await runInjectedScriptInVm(script);
        } catch (vmError) {
          const recorded = scriptResults.get(hashInjectedScript(script));
          if (recorded) {
@@ -8,6 +8,7 @@ import {
  isKanaChar,
  isKanaOnlyText,
  isTokenPos2Excluded,
+  normalizeKana,
 } from './token-classification';

 const POS1_EXCLUSIONS = new Set(['助詞']);
@@ -29,6 +30,26 @@ function makeNoun(surface: string): MergedToken {
  };
 }

+test('kana normalization folds halfwidth kana, composing the voiced pairs', () => {
+  // ｶ + ﾞ is two code points for one character: without composing them, a
+  // halfwidth word counts as longer than the reading that spells it, which
+  // disqualifies the reading from known-word matching.
+  assert.equal(normalizeKana('ｶﾞｸ'), normalizeKana('ガク'));
+  assert.equal(normalizeKana('ﾊﾟﾝ'), normalizeKana('パン'));
+  assert.equal(normalizeKana('ﾐﾅﾄ'), 'みなと');
+  assert.ok(isKanaOnlyText('ｶﾞｸ'));
+});
+
+test('kana normalization leaves characters other than halfwidth kana alone', () => {
+  // The composition is scoped to the halfwidth runs: applied to the whole
+  // string, NFKC would also rewrite these into something the dictionary, the
+  // known-word list, and the frequency data were never keyed on.
+  assert.equal(normalizeKana('①ｶﾞ'), '①が');
+  assert.equal(normalizeKana('Ａｶﾞ'), 'Ａが');
+  assert.equal(normalizeKana('㍑ｶﾞ'), '㍑が');
+  assert.equal(normalizeKana('ﬁｶﾞ'), 'ﬁが');
+});
+
 test('kana classification excludes the katakana-hiragana double hyphen', () => {
  assert.equal(isKanaChar('゠'), false);
  assert.equal(isKanaOnlyText('゠'), false);
@@ -4,8 +4,20 @@ const KATAKANA_TO_HIRAGANA_OFFSET = 0x60;
 const KATAKANA_CODEPOINT_START = 0x30a1;
 const KATAKANA_CODEPOINT_END = 0x30f6;

+// No `u` flag: the range is entirely BMP so it changes nothing here, and
+// Bun's unicode-mode matcher mis-handles this class next to certain ligatures.
+const HALFWIDTH_KANA_RUN = /[\uff66-\uff9f]+/g;
+
+// NFKC over the halfwidth kana only, never the whole string: it composes the
+// voiced pairs (ｶ + ﾞ) into single characters so ｶﾞｸ compares equal to ガク
+// instead of counting one character longer than the word it spells, but run
+// over everything it would also rewrite unrelated text (① → 1, ㍑ → リットル).
+function composeHalfwidthKana(text: string): string {
+  return text.replace(HALFWIDTH_KANA_RUN, (run) => run.normalize('NFKC'));
+}
+
 export function normalizeKana(text: string): string {
-  const raw = text.trim();
+  const raw = composeHalfwidthKana(text).trim();
  if (!raw) {
    return '';
  }
@@ -0,0 +1,150 @@
+// Dictionary classification for the injected scan runtime: which dictionaries
+// an entry came from, and whether it is a SubMiner character entry for the
+// media being watched. Both walk nested entry data, so both are memoized on the
+// entry object by the runtime that hosts them.
+
+import { CHARACTER_DICTIONARY_TITLE_PREFIX } from './character-dictionary-title';
+
+// The prefix is interpolated into generated regex source, so metacharacters in
+// it would change what the pattern matches (or fail to compile).
+const ESCAPED_TITLE_PREFIX = CHARACTER_DICTIONARY_TITLE_PREFIX.replace(
+  /[.*+?^${}()|[\]\\]/g,
+  '\\$&',
+);
+const TITLE_MEDIA_ID_PATTERN = ESCAPED_TITLE_PREFIX + String.raw`[^\d]*(?:AniList\s*)?(\d+)`;
+
+export const YOMITAN_DICTIONARY_CLASSIFICATION_HELPERS = String.raw`
+      function normalizeWordClasses(headword) {
+          if (!Array.isArray(headword?.wordClasses)) { return undefined; }
+          const classes = headword.wordClasses.filter((wordClass) => typeof wordClass === "string" && wordClass.trim().length > 0);
+          return classes.length > 0 ? classes : undefined;
+        }
+        function appendDictionaryNames(target, value) {
+          if (!value || typeof value !== 'object') {
+            return;
+          }
+          const candidates = [
+            value.dictionary,
+            value.dictionaryName,
+            value.name,
+            value.title,
+            value.dictionaryTitle,
+            value.dictionaryAlias
+          ];
+          for (const candidate of candidates) {
+            if (typeof candidate === 'string' && candidate.trim().length > 0) {
+              target.push(candidate.trim());
+            }
+          }
+        }
+        // Memoized on the entry object: termsFind results are cached across
+        // lines, so the same entries come back for every repeated lookup, and
+        // each one is classified several times per scan (name pre-pass,
+        // headword preference, every retry window).
+        function getDictionaryEntryNames(entry) {
+          if (!entry || typeof entry !== 'object') { return []; }
+          const cached = dictionaryEntryNamesCache.get(entry);
+          if (cached !== undefined) { return cached; }
+          const names = [];
+          appendDictionaryNames(names, entry);
+          for (const definition of entry?.definitions || []) {
+            appendDictionaryNames(names, definition);
+          }
+          for (const frequency of entry?.frequencies || []) {
+            appendDictionaryNames(names, frequency);
+          }
+          for (const pronunciation of entry?.pronunciations || []) {
+            appendDictionaryNames(names, pronunciation);
+          }
+          dictionaryEntryNamesCache.set(entry, names);
+          return names;
+        }
+        // Cached per scan rather than per runtime: the answer depends on
+        // includeNameMatchMetadata, which is a per-call parameter.
+        const nameDictionaryEntryCache = new WeakMap();
+        function isNameDictionaryEntry(entry) {
+          if (!includeNameMatchMetadata || !entry || typeof entry !== 'object') {
+            return false;
+          }
+          const cached = nameDictionaryEntryCache.get(entry);
+          if (cached !== undefined) { return cached; }
+          const isName = getDictionaryEntryNames(entry).some((name) => name.startsWith(${JSON.stringify(CHARACTER_DICTIONARY_TITLE_PREFIX)}));
+          nameDictionaryEntryCache.set(entry, isName);
+          return isName;
+        }
+        const TITLE_MEDIA_ID_REGEX = new RegExp(${JSON.stringify(TITLE_MEDIA_ID_PATTERN)}, 'i');
+        function parseSubMinerMediaIdFromString(value) {
+          const imageMatch = value.match(/\bimg\/m(\d+)-/i);
+          if (imageMatch) {
+            const parsed = Number.parseInt(imageMatch[1], 10);
+            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
+          }
+          const titleMatch = value.match(TITLE_MEDIA_ID_REGEX);
+          if (titleMatch) {
+            const parsed = Number.parseInt(titleMatch[1], 10);
+            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
+          }
+          return null;
+        }
+        function parseSubMinerMediaIdCandidate(value) {
+          if (typeof value === 'number' && Number.isSafeInteger(value) && value > 0) {
+            return value;
+          }
+          if (typeof value === 'string' && /^\d+$/.test(value.trim())) {
+            const parsed = Number.parseInt(value.trim(), 10);
+            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
+          }
+          return null;
+        }
+        function collectSubMinerMediaIds(value, target) {
+          if (typeof value === 'string') {
+            const parsed = parseSubMinerMediaIdFromString(value);
+            if (parsed !== null) { target.add(parsed); }
+            return;
+          }
+          if (!value || typeof value !== 'object') {
+            return;
+          }
+          if (Array.isArray(value)) {
+            for (const item of value) { collectSubMinerMediaIds(item, target); }
+            return;
+          }
+          const mediaIdCandidates = [
+            value.subminerMediaId,
+            value.subMinerMediaId,
+            value.characterDictionaryMediaId,
+            value.data?.subminerMediaId,
+            value.data?.subMinerMediaId,
+            value.data?.characterDictionaryMediaId
+          ];
+          for (const candidate of mediaIdCandidates) {
+            const parsed = parseSubMinerMediaIdCandidate(candidate);
+            if (parsed !== null) { target.add(parsed); }
+          }
+          for (const child of Object.values(value)) {
+            collectSubMinerMediaIds(child, target);
+          }
+        }
+        // Walking an entry collects media ids from every nested value, so this
+        // is the most expensive classification step; memoized on the entry for
+        // the same reason as the dictionary names above.
+        function getSubMinerMediaIds(entry) {
+          if (!entry || typeof entry !== 'object') { return EMPTY_MEDIA_ID_SET; }
+          const cached = subMinerMediaIdsCache.get(entry);
+          if (cached !== undefined) { return cached; }
+          const mediaIds = new Set();
+          collectSubMinerMediaIds(entry, mediaIds);
+          subMinerMediaIdsCache.set(entry, mediaIds);
+          return mediaIds;
+        }
+        function isCurrentMediaNameDictionaryEntry(entry) {
+          if (!isNameDictionaryEntry(entry)) {
+            return false;
+          }
+          if (currentCharacterDictionaryMediaId === null) {
+            return true;
+          }
+          const mediaIds = getSubMinerMediaIds(entry);
+          return mediaIds.size === 0 || mediaIds.has(currentCharacterDictionaryMediaId);
+      }
+`;
@@ -0,0 +1,135 @@
+// Frequency-rank resolution for the injected scan runtime: reads the many
+// shapes a Yomitan frequency entry can take and picks the best rank for a
+// headword, honouring per-dictionary priority and occurrence-vs-rank mode.
+
+export const YOMITAN_FREQUENCY_HELPERS = String.raw`
+      function parsePositiveFrequencyNumber(value) {
+        if (typeof value === 'number' && Number.isFinite(value) && value > 0) {
+          return Math.max(1, Math.floor(value));
+        }
+        if (typeof value === 'string') {
+          const numericMatch = value.trim().match(/[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?/)?.[0];
+          if (!numericMatch) { return null; }
+          const parsed = Number.parseFloat(numericMatch);
+          if (!Number.isFinite(parsed) || parsed <= 0) { return null; }
+          return Math.max(1, Math.floor(parsed));
+        }
+        if (Array.isArray(value)) {
+          for (const item of value) {
+            const parsed = parsePositiveFrequencyNumber(item);
+            if (parsed !== null) { return parsed; }
+          }
+        }
+        return null;
+      }
+      function parseDisplayFrequencyNumber(value) {
+        if (typeof value === 'string') {
+          const leadingDigits = value.trim().match(/^\d+/)?.[0];
+          if (!leadingDigits) { return null; }
+          const parsed = Number.parseInt(leadingDigits, 10);
+          return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
+        }
+        return parsePositiveFrequencyNumber(value);
+      }
+      function getFrequencyDictionaryName(frequency) {
+        const candidates = [
+          frequency?.dictionary,
+          frequency?.dictionaryName,
+          frequency?.name,
+          frequency?.title,
+          frequency?.dictionaryTitle,
+          frequency?.dictionaryAlias
+        ];
+        for (const candidate of candidates) {
+          if (typeof candidate === 'string' && candidate.trim().length > 0) {
+            return candidate.trim();
+          }
+        }
+        return null;
+      }
+      function getBestFrequencyRank(dictionaryEntry, headwordIndex, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
+        let best = null;
+        const headwordCount = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords.length : 0;
+        for (const frequency of dictionaryEntry?.frequencies || []) {
+          if (!frequency || typeof frequency !== 'object') { continue; }
+          const frequencyHeadwordIndex = frequency.headwordIndex;
+          if (typeof frequencyHeadwordIndex === 'number') {
+            if (frequencyHeadwordIndex !== headwordIndex) { continue; }
+          } else if (headwordCount > 1) {
+            continue;
+          }
+          const dictionary = getFrequencyDictionaryName(frequency);
+          if (!dictionary) { continue; }
+          if (dictionaryFrequencyModeByName[dictionary] === 'occurrence-based') { continue; }
+          const rank =
+            parseDisplayFrequencyNumber(frequency.displayValue) ??
+            parsePositiveFrequencyNumber(frequency.frequency);
+          if (rank === null) { continue; }
+          const priorityRaw = dictionaryPriorityByName[dictionary];
+          const fallbackPriority =
+            typeof frequency.dictionaryIndex === 'number' && Number.isFinite(frequency.dictionaryIndex)
+              ? Math.max(0, Math.floor(frequency.dictionaryIndex))
+              : Number.MAX_SAFE_INTEGER;
+          const priority =
+            typeof priorityRaw === 'number' && Number.isFinite(priorityRaw)
+              ? Math.max(0, Math.floor(priorityRaw))
+              : fallbackPriority;
+          if (best === null || priority < best.priority || (priority === best.priority && rank < best.rank)) {
+            best = { priority, rank };
+          }
+        }
+        return best?.rank ?? null;
+      }
+      function hasExactSource(headword, token, requirePrimary) {
+        for (const src of headword?.sources || []) {
+          if (src.originalText !== token) { continue; }
+          if (requirePrimary && !src.isPrimary) { continue; }
+          if (src.matchType !== 'exact') { continue; }
+          return true;
+        }
+        return false;
+      }
+      function collectExactHeadwordMatches(dictionaryEntries, token, requirePrimary) {
+        const matches = [];
+        for (const dictionaryEntry of dictionaryEntries || []) {
+          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
+          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
+            const headword = headwords[headwordIndex];
+            if (!hasExactSource(headword, token, requirePrimary)) { continue; }
+            matches.push({ dictionaryEntry, headword, headwordIndex });
+          }
+        }
+        return matches;
+      }
+      function sameHeadword(match, preferredMatch) {
+        if (!match || !preferredMatch) {
+          return false;
+        }
+        if (match.headword?.term !== preferredMatch.headword?.term) {
+          return false;
+        }
+        const matchReading = typeof match.headword?.reading === 'string' ? match.headword.reading : '';
+        const preferredReading =
+          typeof preferredMatch.headword?.reading === 'string' ? preferredMatch.headword.reading : '';
+        if (!matchReading || !preferredReading) {
+          return true;
+        }
+        return matchReading === preferredReading;
+      }
+      function getBestFrequencyRankForMatches(matches, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
+        let best = null;
+        for (const match of matches) {
+          const rank = getBestFrequencyRank(
+            match.dictionaryEntry,
+            match.headwordIndex,
+            dictionaryPriorityByName,
+            dictionaryFrequencyModeByName
+          );
+          if (rank === null) { continue; }
+          if (best === null || rank < best) {
+            best = rank;
+          }
+        }
+        return best;
+      }
+`;
@@ -0,0 +1,170 @@
+// Furigana distribution for the injected scan runtime: splits a headword and
+// its reading into the segments a token carries, including the inflected case
+// where the matched source text differs from the dictionary form.
+
+export const YOMITAN_FURIGANA_HELPERS = String.raw`
+      function createFuriganaSegment(text, reading) { return {text, reading}; }
+      function getSegmentReadingContribution(segment) {
+        if (typeof segment.reading === "string" && segment.reading.length > 0) { return segment.reading; }
+        const segmentText = typeof segment.text === "string" ? segment.text : "";
+        const isKanaOnly = segmentText.length > 0 && [...segmentText].every((char) => isCodePointKana(char.codePointAt(0)));
+        return isKanaOnly ? convertHalfwidthKanaToKatakana(segmentText) : "";
+      }
+      function getProlongedHiragana(previousCharacter) {
+        switch (previousCharacter) {
+          case "あ": case "か": case "が": case "さ": case "ざ": case "た": case "だ": case "な": case "は": case "ば": case "ぱ": case "ま": case "や": case "ら": case "わ": case "ぁ": case "ゃ": case "ゎ": return "あ";
+          case "い": case "き": case "ぎ": case "し": case "じ": case "ち": case "ぢ": case "に": case "ひ": case "び": case "ぴ": case "み": case "り": case "ぃ": return "い";
+          case "う": case "く": case "ぐ": case "す": case "ず": case "つ": case "づ": case "ぬ": case "ふ": case "ぶ": case "ぷ": case "む": case "ゆ": case "る": case "ぅ": case "ゅ": return "う";
+          case "え": case "け": case "げ": case "せ": case "ぜ": case "て": case "で": case "ね": case "へ": case "べ": case "ぺ": case "め": case "れ": case "ぇ": return "え";
+          case "お": case "こ": case "ご": case "そ": case "ぞ": case "と": case "ど": case "の": case "ほ": case "ぼ": case "ぽ": case "も": case "よ": case "ろ": case "を": case "ぉ": case "ょ": return "う";
+          default: return null;
+        }
+      }
+      function getFuriganaKanaSegments(text, reading) {
+        const newSegments = [];
+        let start = 0;
+        let state = (reading[0] === text[0]);
+        for (let i = 1; i < text.length; ++i) {
+          const newState = (reading[i] === text[i]);
+          if (state === newState) { continue; }
+          newSegments.push(createFuriganaSegment(text.substring(start, i), state ? '' : reading.substring(start, i)));
+          state = newState;
+          start = i;
+        }
+        newSegments.push(createFuriganaSegment(text.substring(start), state ? '' : reading.substring(start)));
+        return newSegments;
+      }
+      function convertKatakanaToHiragana(text, keepProlongedSoundMarks = false) {
+        let result = '';
+        const offset = (HIRAGANA_CONVERSION_RANGE[0] - KATAKANA_CONVERSION_RANGE[0]);
+        for (let char of text) {
+          const codePoint = char.codePointAt(0);
+          switch (codePoint) {
+            case KATAKANA_SMALL_KA_CODE_POINT:
+            case KATAKANA_SMALL_KE_CODE_POINT:
+              break;
+            case KANA_PROLONGED_SOUND_MARK_CODE_POINT:
+            case HALFWIDTH_KANA_PROLONGED_SOUND_MARK_CODE_POINT:
+              char = "ー";
+              if (!keepProlongedSoundMarks && result.length > 0) {
+                const char2 = getProlongedHiragana(result[result.length - 1]);
+                if (char2 !== null) { char = char2; }
+              }
+              break;
+            default:
+              if (isCodePointInRange(codePoint, KATAKANA_CONVERSION_RANGE)) {
+                char = String.fromCodePoint(codePoint + offset);
+                break;
+              }
+              // Halfwidth katakana folds too, or a name written that way would
+              // match neither a candidate form nor its own reading.
+              const halfwidthHiragana = convertHalfwidthKanaCodePointToHiragana(codePoint);
+              if (halfwidthHiragana !== null) { char = halfwidthHiragana; }
+              break;
+          }
+          result += char;
+        }
+        return result;
+      }
+      function segmentizeFurigana(reading, readingNormalized, groups, groupsStart) {
+        const groupCount = groups.length - groupsStart;
+        if (groupCount <= 0) { return reading.length === 0 ? [] : null; }
+        const group = groups[groupsStart];
+        const {isKana, text} = group;
+        if (isKana) {
+          if (group.textNormalized !== null && readingNormalized.startsWith(group.textNormalized)) {
+            const segments = segmentizeFurigana(reading.substring(text.length), readingNormalized.substring(text.length), groups, groupsStart + 1);
+            if (segments !== null) {
+              if (reading.startsWith(text)) { segments.unshift(createFuriganaSegment(text, '')); }
+              else { segments.unshift(...getFuriganaKanaSegments(text, reading)); }
+              return segments;
+            }
+          }
+          return null;
+        }
+        let result = null;
+        for (let i = reading.length; i >= text.length; --i) {
+          const segments = segmentizeFurigana(reading.substring(i), readingNormalized.substring(i), groups, groupsStart + 1);
+          if (segments !== null) {
+            if (result !== null) { return null; }
+            segments.unshift(createFuriganaSegment(text, reading.substring(0, i)));
+            result = segments;
+          }
+          if (groupCount === 1) { break; }
+        }
+        return result;
+      }
+      function distributeFurigana(term, reading) {
+        if (reading === term) { return [createFuriganaSegment(term, '')]; }
+        const groups = [];
+        let groupPre = null;
+        let isKanaPre = null;
+        for (const c of term) {
+          const isKana = isCodePointKana(c.codePointAt(0));
+          if (isKana === isKanaPre) { groupPre.text += c; }
+          else {
+            groupPre = {isKana, text: c, textNormalized: null};
+            groups.push(groupPre);
+            isKanaPre = isKana;
+          }
+        }
+        for (const group of groups) {
+          if (group.isKana) { group.textNormalized = convertKatakanaToHiragana(group.text); }
+        }
+        const segments = segmentizeFurigana(reading, convertKatakanaToHiragana(reading), groups, 0);
+        return segments !== null ? segments : [createFuriganaSegment(term, reading)];
+      }
+      function getStemLength(text1, text2) {
+        const minLength = Math.min(text1.length, text2.length);
+        if (minLength === 0) { return 0; }
+        let i = 0;
+        while (true) {
+          const char1 = text1.codePointAt(i);
+          const char2 = text2.codePointAt(i);
+          if (char1 !== char2) { break; }
+          const charLength = String.fromCodePoint(char1).length;
+          i += charLength;
+          if (i >= minLength) {
+            if (i > minLength) { i -= charLength; }
+            break;
+          }
+        }
+        return i;
+      }
+      function distributeFuriganaInflected(term, reading, source) {
+        const termNormalized = convertKatakanaToHiragana(term);
+        const readingNormalized = convertKatakanaToHiragana(reading);
+        const sourceNormalized = convertKatakanaToHiragana(source);
+        let mainText = term;
+        let stemLength = getStemLength(termNormalized, sourceNormalized);
+        const readingStemLength = getStemLength(readingNormalized, sourceNormalized);
+        if (readingStemLength > 0 && readingStemLength >= stemLength) {
+          mainText = reading;
+          stemLength = readingStemLength;
+          reading = source.substring(0, stemLength) + reading.substring(stemLength);
+        }
+        const segments = [];
+        if (stemLength > 0) {
+          mainText = source.substring(0, stemLength) + mainText.substring(stemLength);
+          const segments2 = distributeFurigana(mainText, reading);
+          let consumed = 0;
+          for (const segment of segments2) {
+            const start = consumed;
+            consumed += segment.text.length;
+            if (consumed < stemLength) { segments.push(segment); }
+            else if (consumed === stemLength) { segments.push(segment); break; }
+            else {
+              if (start < stemLength) { segments.push(createFuriganaSegment(mainText.substring(start, stemLength), '')); }
+              break;
+            }
+          }
+        }
+        if (stemLength < source.length) {
+          const remainder = source.substring(stemLength);
+          const last = segments[segments.length - 1];
+          if (last && last.reading.length === 0) { last.text += remainder; }
+          else { segments.push(createFuriganaSegment(remainder, '')); }
+        }
+        return segments;
+      }
+`;
@@ -0,0 +1,45 @@
+// Kana classification and normalization for the injected scan runtime: the
+// code-point ranges the walk tests every character against, and the folds that
+// let halfwidth and katakana spellings compare equal to their dictionary form.
+import { HAN_CODE_POINT_RANGES } from '../../text/han-code-points';
+
+export const YOMITAN_KANA_HELPERS = String.raw`
+      const HIRAGANA_CONVERSION_RANGE = [0x3041, 0x3096];
+      const KATAKANA_CONVERSION_RANGE = [0x30a1, 0x30f6];
+      const KANA_PROLONGED_SOUND_MARK_CODE_POINT = 0x30fc;
+      const KATAKANA_SMALL_KA_CODE_POINT = 0x30f5;
+      const KATAKANA_SMALL_KE_CODE_POINT = 0x30f6;
+      const KANA_RANGES = [[0x3040, 0x309f], [0x30a0, 0x30ff], [0xff66, 0xff9f]];
+      const HALFWIDTH_KATAKANA_RANGE = [0xff66, 0xff9d];
+      const HALFWIDTH_KANA_PROLONGED_SOUND_MARK_CODE_POINT = 0xff70;
+      // Folded one code point to one, so every index into a normalized string
+      // still lines up with the original text — the name-candidate prefilter
+      // and the furigana stem matching both index back into it. The standalone
+      // voiced marks (ﾞ ﾟ) have no one-character equivalent and stay as they are.
+      const HALFWIDTH_KATAKANA_TO_HIRAGANA = "をぁぃぅぇぉゃゅょっーあいうえおかきくけこさしすせそたちつてとなにぬねのはひふへほまみむめもやゆよらりるれろわん";
+      function convertHalfwidthKanaCodePointToHiragana(codePoint) {
+        if (codePoint < HALFWIDTH_KATAKANA_RANGE[0] || codePoint > HALFWIDTH_KATAKANA_RANGE[1]) { return null; }
+        return HALFWIDTH_KATAKANA_TO_HIRAGANA[codePoint - HALFWIDTH_KATAKANA_RANGE[0]] || null;
+      }
+      // Halfwidth katakana is kana here but not to the rest of the pipeline
+      // (known-word matching and frequency lookups only fold fullwidth), so a
+      // reading taken from halfwidth text is written the way the fullwidth
+      // katakana path already writes it. NFKC rather than the per-code-point
+      // table: this is the one place where nothing indexes back into the
+      // result, so a voiced pair (ｶ + ﾞ) can compose into the single ガ it
+      // means instead of leaving a stray combining mark in the reading. Scoped
+      // to the halfwidth runs, because NFKC over everything else rewrites
+      // characters that have nothing to do with kana (① → 1, ㍑ → リットル).
+      function convertHalfwidthKanaToKatakana(text) {
+        return text.replace(/[ｦ-ﾟ]+/g, (run) => run.normalize("NFKC"));
+      }
+      // Han ranges come from the shared table so the scan walk and the character
+      // dictionary agree on what a kanji is (supplementary planes included).
+      // Halfwidth katakana counts as Japanese text: a name written that way has
+      // to reach the greedy pre-pass, which has its own handling for it.
+      const JAPANESE_RANGES = [[0x3040, 0x30ff], [0xff66, 0xff9f], ...${JSON.stringify(HAN_CODE_POINT_RANGES)}];
+      function isCodePointInRange(codePoint, range) { return codePoint >= range[0] && codePoint <= range[1]; }
+      function isCodePointInRanges(codePoint, ranges) { return ranges.some((range) => isCodePointInRange(codePoint, range)); }
+      function isCodePointKana(codePoint) { return isCodePointInRanges(codePoint, KANA_RANGES); }
+      function isCodePointJapanese(codePoint) { return isCodePointInRanges(codePoint, JAPANESE_RANGES); }
+`;
@@ -0,0 +1,79 @@
+// Match selection for the injected scan runtime: picks the headword a position
+// tokenizes to, and the longest name or generic match in a window, which is how
+// the greedy name pre-pass decides what to reserve.
+
+export const YOMITAN_MATCH_SELECTION_HELPERS = String.raw`
+      function findLongestNameMatch(dictionaryEntries, textWindow) {
+        let best = null;
+        for (const dictionaryEntry of dictionaryEntries || []) {
+          if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
+          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
+          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
+            const headword = headwords[headwordIndex];
+            for (const src of headword?.sources || []) {
+              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
+              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
+              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
+              if (best === null || originalText.length > best.sourceLength) {
+                best = { dictionaryEntry, headword, headwordIndex, sourceLength: originalText.length };
+              }
+            }
+          }
+        }
+        return best;
+      }
+      function findLongestGenericMatchLength(dictionaryEntries, textWindow) {
+        let best = 0;
+        for (const dictionaryEntry of dictionaryEntries || []) {
+          if (isNameDictionaryEntry(dictionaryEntry)) { continue; }
+          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
+          for (const headword of headwords) {
+            for (const src of headword?.sources || []) {
+              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
+              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
+              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
+              if (originalText.length > best) { best = originalText.length; }
+            }
+          }
+        }
+        return best;
+      }
+      function getPreferredHeadword(dictionaryEntries, token, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
+        const currentMediaDictionaryEntries =
+          currentCharacterDictionaryMediaId === null
+            ? (dictionaryEntries || [])
+            : (dictionaryEntries || []).filter((entry) => {
+                if (!isNameDictionaryEntry(entry)) { return true; }
+                return isCurrentMediaNameDictionaryEntry(entry);
+              });
+        const exactPrimaryMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, true);
+        let matchedNameDictionary = false;
+        if (includeNameMatchMetadata) {
+          // Every match already comes from currentMediaDictionaryEntries, so
+          // classifying its own entry is enough.
+          for (const match of exactPrimaryMatches) {
+            if (!isCurrentMediaNameDictionaryEntry(match.dictionaryEntry)) { continue; }
+            matchedNameDictionary = true;
+            break;
+          }
+        }
+        const preferredMatch = exactPrimaryMatches[0];
+        if (preferredMatch) {
+          const exactFrequencyMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, false)
+            .filter((match) => sameHeadword(match, preferredMatch));
+          return {
+            term: preferredMatch.headword.term,
+            reading: preferredMatch.headword.reading,
+            wordClasses: normalizeWordClasses(preferredMatch.headword),
+            isNameMatch:
+              matchedNameDictionary || isCurrentMediaNameDictionaryEntry(preferredMatch.dictionaryEntry),
+            frequencyRank: getBestFrequencyRankForMatches(
+              exactFrequencyMatches.length > 0 ? exactFrequencyMatches : exactPrimaryMatches,
+              dictionaryPriorityByName,
+              dictionaryFrequencyModeByName
+            )
+          };
+        }
+        return null;
+      }
+`;
@@ -3,6 +3,14 @@ import * as fs from 'fs';
 import * as http from 'http';
 import * as path from 'path';
 import { selectYomitanParseTokens } from './parser-selection-stage';
+import {
+  buildYomitanScanCallScript,
+  buildYomitanScanNameCandidatesScript,
+  CHARACTER_DICTIONARY_TITLE_PREFIX,
+  YOMITAN_SCAN_RUNTIME_INSTALL_SCRIPT,
+  YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL,
+  type YomitanFrequencyMode,
+} from './yomitan-scan-runtime-script';

 interface LoggerLike {
  error: (message: string, ...args: unknown[]) => void;
@@ -22,8 +30,6 @@ interface YomitanParserRuntimeDeps {
  createYomitanExtensionWindow?: (pageName: string) => Promise<BrowserWindow | null>;
 }

-type YomitanFrequencyMode = 'occurrence-based' | 'rank-based';
-
 export interface YomitanDictionaryInfo {
  title: string;
  revision?: string | number;
@@ -74,13 +80,19 @@ export interface YomitanAddNoteResult {
 }

 const DEFAULT_YOMITAN_SCAN_LENGTH = 40;
-const CHARACTER_DICTIONARY_TITLE_PREFIX = 'SubMiner Character Dictionary';
 const yomitanProfileMetadataByWindow = new WeakMap<BrowserWindow, YomitanProfileMetadata>();
 const yomitanProfileDiagnosticsLoggedByWindow = new WeakSet<BrowserWindow>();
 const yomitanFrequencyCacheByWindow = new WeakMap<
  BrowserWindow,
  Map<string, YomitanTermFrequency[]>
 >();
+// Epoch passed with every scan request; the in-window termsFind cache clears
+// itself when the epoch changes (dictionary imports, settings changes).
+const yomitanScanCacheEpochByWindow = new WeakMap<BrowserWindow, number>();
+
+function getYomitanScanCacheEpoch(window: BrowserWindow): number {
+  return yomitanScanCacheEpochByWindow.get(window) ?? 0;
+}

 function isObject(value: unknown): value is Record<string, unknown> {
  return Boolean(value && typeof value === 'object');
@@ -99,6 +111,7 @@ function isScanTokenArray(value: unknown): value is YomitanScanToken[] {
        typeof entry.startPos === 'number' &&
        typeof entry.endPos === 'number' &&
        (entry.isNameMatch === undefined || typeof entry.isNameMatch === 'boolean') &&
+        (entry.isUnparsedRun === undefined || typeof entry.isUnparsedRun === 'boolean') &&
        (entry.frequencyRank === undefined || typeof entry.frequencyRank === 'number') &&
        (entry.wordClasses === undefined ||
          (Array.isArray(entry.wordClasses) &&
@@ -107,13 +120,9 @@ function isScanTokenArray(value: unknown): value is YomitanScanToken[] {
  );
 }

-function scanTokenSpanKey(token: YomitanScanToken): string {
-  return `${token.startPos}:${token.endPos}:${token.surface}`;
-}
-
 // Maps a parse-selected token to the scanner-token shape carried out of the
-// parser runtime. Shared by both selectYomitanParseTokens fallback paths so the
-// projected fields stay in sync as the shape changes.
+// parser runtime, used by the parseText fallback path when the in-window
+// scanner is unavailable.
 function toYomitanScanToken(token: {
  surface: string;
  reading: string;
@@ -132,66 +141,6 @@ function toYomitanScanToken(token: {
  };
 }

-// parseText segmentation is authoritative (it emits filler chunks for text the
-// termsFind scanner skips), but only the termsFind scanner carries annotation
-// metadata (isNameMatch, frequencyRank, headwordReading, wordClasses). Graft
-// scanner tokens onto the parseText segmentation per matching span so one
-// unmatched chunk degrades only itself instead of dropping the whole line's
-// metadata.
-//
-// Exception: character-name tokens. The greedy name scan can re-segment text
-// around a name (e.g. とヨータ → と + ヨータ instead of とヨー + タ), so
-// parseText segmentation cannot be authoritative there. Each name span is
-// expanded until it aligns with token boundaries in both segmentations, then
-// the parse tokens inside are replaced with the scanner tokens.
-function mergeScannerTokensIntoParseTokens(
-  parseScanTokens: YomitanScanToken[],
-  scannerTokens: YomitanScanToken[],
-): YomitanScanToken[] {
-  const scannerTokensBySpan = new Map<string, YomitanScanToken>();
-  for (const token of scannerTokens) {
-    scannerTokensBySpan.set(scanTokenSpanKey(token), token);
-  }
-  const graftedTokens = parseScanTokens.map(
-    (token) => scannerTokensBySpan.get(scanTokenSpanKey(token)) ?? token,
-  );
-
-  const nameTokens = scannerTokens.filter((token) => token.isNameMatch === true);
-  if (nameTokens.length === 0) {
-    return graftedTokens;
-  }
-
-  const regions = nameTokens.map((token) => ({ start: token.startPos, end: token.endPos }));
-  const allTokens = [...parseScanTokens, ...scannerTokens];
-  let expanded = true;
-  while (expanded) {
-    expanded = false;
-    for (const region of regions) {
-      for (const token of allTokens) {
-        const overlaps = token.startPos < region.end && token.endPos > region.start;
-        const extendsBeyond = token.startPos < region.start || token.endPos > region.end;
-        if (overlaps && extendsBeyond) {
-          region.start = Math.min(region.start, token.startPos);
-          region.end = Math.max(region.end, token.endPos);
-          expanded = true;
-        }
-      }
-    }
-  }
-
-  const isInsideNameRegion = (token: YomitanScanToken): boolean =>
-    regions.some((region) => token.startPos >= region.start && token.endPos <= region.end);
-
-  const merged = graftedTokens.filter((token) => !isInsideNameRegion(token));
-  for (const token of scannerTokens) {
-    if (isInsideNameRegion(token)) {
-      merged.push(token);
-    }
-  }
-  merged.sort((a, b) => a.startPos - b.startPos || a.endPos - b.endPos);
-  return merged;
-}
-
 function makeTermReadingCacheKey(term: string, reading: string | null): string {
  return `${term}\u0000${reading ?? ''}`;
 }
@@ -208,6 +157,7 @@ function getWindowFrequencyCache(window: BrowserWindow): Map<string, YomitanTerm
 function clearWindowCaches(window: BrowserWindow): void {
  yomitanProfileMetadataByWindow.delete(window);
  yomitanFrequencyCacheByWindow.delete(window);
+  yomitanScanCacheEpochByWindow.set(window, getYomitanScanCacheEpoch(window) + 1);
 }
 export function clearYomitanParserCachesForWindow(window: BrowserWindow): void {
  clearWindowCaches(window);
@@ -704,6 +654,10 @@ async function ensureYomitanParserWindow(
      if (readyPromise) {
        await readyPromise;
      }
+      // Eagerly install the scan runtime so the first subtitle line does not
+      // pay the install round trip; failures fall back to the per-request
+      // install-and-retry path.
+      await installYomitanScanRuntime(parserWindow).catch(() => {});

      return true;
    } catch (err) {
@@ -877,668 +831,42 @@ async function serveDictionaryZipOnce<T>(
  }
 }

-const YOMITAN_SCANNING_HELPERS = String.raw`
-      const HIRAGANA_CONVERSION_RANGE = [0x3041, 0x3096];
-      const KATAKANA_CONVERSION_RANGE = [0x30a1, 0x30f6];
-      const KANA_PROLONGED_SOUND_MARK_CODE_POINT = 0x30fc;
-      const KATAKANA_SMALL_KA_CODE_POINT = 0x30f5;
-      const KATAKANA_SMALL_KE_CODE_POINT = 0x30f6;
-      const KANA_RANGES = [[0x3040, 0x309f], [0x30a0, 0x30ff]];
-      const JAPANESE_RANGES = [[0x3040, 0x30ff], [0x3400, 0x9fff]];
-      function isCodePointInRange(codePoint, range) { return codePoint >= range[0] && codePoint <= range[1]; }
-      function isCodePointInRanges(codePoint, ranges) { return ranges.some((range) => isCodePointInRange(codePoint, range)); }
-      function isCodePointKana(codePoint) { return isCodePointInRanges(codePoint, KANA_RANGES); }
-      function isCodePointJapanese(codePoint) { return isCodePointInRanges(codePoint, JAPANESE_RANGES); }
-      function createFuriganaSegment(text, reading) { return {text, reading}; }
-      function getSegmentReadingContribution(segment) {
-        if (typeof segment.reading === "string" && segment.reading.length > 0) { return segment.reading; }
-        const segmentText = typeof segment.text === "string" ? segment.text : "";
-        const isKanaOnly = segmentText.length > 0 && [...segmentText].every((char) => isCodePointKana(char.codePointAt(0)));
-        return isKanaOnly ? segmentText : "";
-      }
-      function getProlongedHiragana(previousCharacter) {
-        switch (previousCharacter) {
-          case "あ": case "か": case "が": case "さ": case "ざ": case "た": case "だ": case "な": case "は": case "ば": case "ぱ": case "ま": case "や": case "ら": case "わ": case "ぁ": case "ゃ": case "ゎ": return "あ";
-          case "い": case "き": case "ぎ": case "し": case "じ": case "ち": case "ぢ": case "に": case "ひ": case "び": case "ぴ": case "み": case "り": case "ぃ": return "い";
-          case "う": case "く": case "ぐ": case "す": case "ず": case "つ": case "づ": case "ぬ": case "ふ": case "ぶ": case "ぷ": case "む": case "ゆ": case "る": case "ぅ": case "ゅ": return "う";
-          case "え": case "け": case "げ": case "せ": case "ぜ": case "て": case "で": case "ね": case "へ": case "べ": case "ぺ": case "め": case "れ": case "ぇ": return "え";
-          case "お": case "こ": case "ご": case "そ": case "ぞ": case "と": case "ど": case "の": case "ほ": case "ぼ": case "ぽ": case "も": case "よ": case "ろ": case "を": case "ぉ": case "ょ": return "う";
-          default: return null;
-        }
-      }
-      function getFuriganaKanaSegments(text, reading) {
-        const newSegments = [];
-        let start = 0;
-        let state = (reading[0] === text[0]);
-        for (let i = 1; i < text.length; ++i) {
-          const newState = (reading[i] === text[i]);
-          if (state === newState) { continue; }
-          newSegments.push(createFuriganaSegment(text.substring(start, i), state ? '' : reading.substring(start, i)));
-          state = newState;
-          start = i;
-        }
-        newSegments.push(createFuriganaSegment(text.substring(start), state ? '' : reading.substring(start)));
-        return newSegments;
-      }
-      function convertKatakanaToHiragana(text, keepProlongedSoundMarks = false) {
-        let result = '';
-        const offset = (HIRAGANA_CONVERSION_RANGE[0] - KATAKANA_CONVERSION_RANGE[0]);
-        for (let char of text) {
-          const codePoint = char.codePointAt(0);
-          switch (codePoint) {
-            case KATAKANA_SMALL_KA_CODE_POINT:
-            case KATAKANA_SMALL_KE_CODE_POINT:
-              break;
-            case KANA_PROLONGED_SOUND_MARK_CODE_POINT:
-              if (!keepProlongedSoundMarks && result.length > 0) {
-                const char2 = getProlongedHiragana(result[result.length - 1]);
-                if (char2 !== null) { char = char2; }
-              }
-              break;
-            default:
-              if (isCodePointInRange(codePoint, KATAKANA_CONVERSION_RANGE)) {
-                char = String.fromCodePoint(codePoint + offset);
-              }
-              break;
-          }
-          result += char;
-        }
-        return result;
-      }
-      function segmentizeFurigana(reading, readingNormalized, groups, groupsStart) {
-        const groupCount = groups.length - groupsStart;
-        if (groupCount <= 0) { return reading.length === 0 ? [] : null; }
-        const group = groups[groupsStart];
-        const {isKana, text} = group;
-        if (isKana) {
-          if (group.textNormalized !== null && readingNormalized.startsWith(group.textNormalized)) {
-            const segments = segmentizeFurigana(reading.substring(text.length), readingNormalized.substring(text.length), groups, groupsStart + 1);
-            if (segments !== null) {
-              if (reading.startsWith(text)) { segments.unshift(createFuriganaSegment(text, '')); }
-              else { segments.unshift(...getFuriganaKanaSegments(text, reading)); }
-              return segments;
-            }
-          }
-          return null;
-        }
-        let result = null;
-        for (let i = reading.length; i >= text.length; --i) {
-          const segments = segmentizeFurigana(reading.substring(i), readingNormalized.substring(i), groups, groupsStart + 1);
-          if (segments !== null) {
-            if (result !== null) { return null; }
-            segments.unshift(createFuriganaSegment(text, reading.substring(0, i)));
-            result = segments;
-          }
-          if (groupCount === 1) { break; }
-        }
-        return result;
-      }
-      function distributeFurigana(term, reading) {
-        if (reading === term) { return [createFuriganaSegment(term, '')]; }
-        const groups = [];
-        let groupPre = null;
-        let isKanaPre = null;
-        for (const c of term) {
-          const isKana = isCodePointKana(c.codePointAt(0));
-          if (isKana === isKanaPre) { groupPre.text += c; }
-          else {
-            groupPre = {isKana, text: c, textNormalized: null};
-            groups.push(groupPre);
-            isKanaPre = isKana;
-          }
-        }
-        for (const group of groups) {
-          if (group.isKana) { group.textNormalized = convertKatakanaToHiragana(group.text); }
-        }
-        const segments = segmentizeFurigana(reading, convertKatakanaToHiragana(reading), groups, 0);
-        return segments !== null ? segments : [createFuriganaSegment(term, reading)];
-      }
-      function getStemLength(text1, text2) {
-        const minLength = Math.min(text1.length, text2.length);
-        if (minLength === 0) { return 0; }
-        let i = 0;
-        while (true) {
-          const char1 = text1.codePointAt(i);
-          const char2 = text2.codePointAt(i);
-          if (char1 !== char2) { break; }
-          const charLength = String.fromCodePoint(char1).length;
-          i += charLength;
-          if (i >= minLength) {
-            if (i > minLength) { i -= charLength; }
-            break;
-          }
-        }
-        return i;
-      }
-      function distributeFuriganaInflected(term, reading, source) {
-        const termNormalized = convertKatakanaToHiragana(term);
-        const readingNormalized = convertKatakanaToHiragana(reading);
-        const sourceNormalized = convertKatakanaToHiragana(source);
-        let mainText = term;
-        let stemLength = getStemLength(termNormalized, sourceNormalized);
-        const readingStemLength = getStemLength(readingNormalized, sourceNormalized);
-        if (readingStemLength > 0 && readingStemLength >= stemLength) {
-          mainText = reading;
-          stemLength = readingStemLength;
-          reading = source.substring(0, stemLength) + reading.substring(stemLength);
-        }
-        const segments = [];
-        if (stemLength > 0) {
-          mainText = source.substring(0, stemLength) + mainText.substring(stemLength);
-          const segments2 = distributeFurigana(mainText, reading);
-          let consumed = 0;
-          for (const segment of segments2) {
-            const start = consumed;
-            consumed += segment.text.length;
-            if (consumed < stemLength) { segments.push(segment); }
-            else if (consumed === stemLength) { segments.push(segment); break; }
-            else {
-              if (start < stemLength) { segments.push(createFuriganaSegment(mainText.substring(start, stemLength), '')); }
-              break;
-            }
-          }
-        }
-        if (stemLength < source.length) {
-          const remainder = source.substring(stemLength);
-          const last = segments[segments.length - 1];
-          if (last && last.reading.length === 0) { last.text += remainder; }
-          else { segments.push(createFuriganaSegment(remainder, '')); }
-        }
-        return segments;
-      }
-      function parsePositiveFrequencyNumber(value) {
-        if (typeof value === 'number' && Number.isFinite(value) && value > 0) {
-          return Math.max(1, Math.floor(value));
-        }
-        if (typeof value === 'string') {
-          const numericMatch = value.trim().match(/[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?/)?.[0];
-          if (!numericMatch) { return null; }
-          const parsed = Number.parseFloat(numericMatch);
-          if (!Number.isFinite(parsed) || parsed <= 0) { return null; }
-          return Math.max(1, Math.floor(parsed));
-        }
-        if (Array.isArray(value)) {
-          for (const item of value) {
-            const parsed = parsePositiveFrequencyNumber(item);
-            if (parsed !== null) { return parsed; }
-          }
-        }
-        return null;
-      }
-      function parseDisplayFrequencyNumber(value) {
-        if (typeof value === 'string') {
-          const leadingDigits = value.trim().match(/^\d+/)?.[0];
-          if (!leadingDigits) { return null; }
-          const parsed = Number.parseInt(leadingDigits, 10);
-          return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
-        }
-        return parsePositiveFrequencyNumber(value);
-      }
-      function getFrequencyDictionaryName(frequency) {
-        const candidates = [
-          frequency?.dictionary,
-          frequency?.dictionaryName,
-          frequency?.name,
-          frequency?.title,
-          frequency?.dictionaryTitle,
-          frequency?.dictionaryAlias
-        ];
-        for (const candidate of candidates) {
-          if (typeof candidate === 'string' && candidate.trim().length > 0) {
-            return candidate.trim();
-          }
-        }
-        return null;
-      }
-      function getBestFrequencyRank(dictionaryEntry, headwordIndex, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
-        let best = null;
-        const headwordCount = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords.length : 0;
-        for (const frequency of dictionaryEntry?.frequencies || []) {
-          if (!frequency || typeof frequency !== 'object') { continue; }
-          const frequencyHeadwordIndex = frequency.headwordIndex;
-          if (typeof frequencyHeadwordIndex === 'number') {
-            if (frequencyHeadwordIndex !== headwordIndex) { continue; }
-          } else if (headwordCount > 1) {
-            continue;
-          }
-          const dictionary = getFrequencyDictionaryName(frequency);
-          if (!dictionary) { continue; }
-          if (dictionaryFrequencyModeByName[dictionary] === 'occurrence-based') { continue; }
-          const rank =
-            parseDisplayFrequencyNumber(frequency.displayValue) ??
-            parsePositiveFrequencyNumber(frequency.frequency);
-          if (rank === null) { continue; }
-          const priorityRaw = dictionaryPriorityByName[dictionary];
-          const fallbackPriority =
-            typeof frequency.dictionaryIndex === 'number' && Number.isFinite(frequency.dictionaryIndex)
-              ? Math.max(0, Math.floor(frequency.dictionaryIndex))
-              : Number.MAX_SAFE_INTEGER;
-          const priority =
-            typeof priorityRaw === 'number' && Number.isFinite(priorityRaw)
-              ? Math.max(0, Math.floor(priorityRaw))
-              : fallbackPriority;
-          if (best === null || priority < best.priority || (priority === best.priority && rank < best.rank)) {
-            best = { priority, rank };
-          }
-        }
-        return best?.rank ?? null;
-      }
-      function hasExactSource(headword, token, requirePrimary) {
-        for (const src of headword.sources || []) {
-          if (src.originalText !== token) { continue; }
-          if (requirePrimary && !src.isPrimary) { continue; }
-          if (src.matchType !== 'exact') { continue; }
-          return true;
-        }
-        return false;
-      }
-      function collectExactHeadwordMatches(dictionaryEntries, token, requirePrimary) {
-        const matches = [];
-        for (const dictionaryEntry of dictionaryEntries || []) {
-          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
-          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
-            const headword = headwords[headwordIndex];
-            if (!hasExactSource(headword, token, requirePrimary)) { continue; }
-            matches.push({ dictionaryEntry, headword, headwordIndex });
-          }
-        }
-        return matches;
-      }
-      function sameHeadword(match, preferredMatch) {
-        if (!match || !preferredMatch) {
-          return false;
-        }
-        if (match.headword?.term !== preferredMatch.headword?.term) {
-          return false;
-        }
-        const matchReading = typeof match.headword?.reading === 'string' ? match.headword.reading : '';
-        const preferredReading =
-          typeof preferredMatch.headword?.reading === 'string' ? preferredMatch.headword.reading : '';
-        if (!matchReading || !preferredReading) {
-          return true;
-        }
-        return matchReading === preferredReading;
-      }
-      function getBestFrequencyRankForMatches(matches, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
-        let best = null;
-        for (const match of matches) {
-          const rank = getBestFrequencyRank(
-            match.dictionaryEntry,
-            match.headwordIndex,
-            dictionaryPriorityByName,
-            dictionaryFrequencyModeByName
-          );
-          if (rank === null) { continue; }
-          if (best === null || rank < best) {
-            best = rank;
-          }
-        }
-        return best;
-      }
-      function normalizeWordClasses(headword) {
-          if (!Array.isArray(headword?.wordClasses)) { return undefined; }
-          const classes = headword.wordClasses.filter((wordClass) => typeof wordClass === "string" && wordClass.trim().length > 0);
-          return classes.length > 0 ? classes : undefined;
-        }
-        function appendDictionaryNames(target, value) {
-          if (!value || typeof value !== 'object') {
-            return;
-          }
-          const candidates = [
-            value.dictionary,
-            value.dictionaryName,
-            value.name,
-            value.title,
-            value.dictionaryTitle,
-            value.dictionaryAlias
-          ];
-          for (const candidate of candidates) {
-            if (typeof candidate === 'string' && candidate.trim().length > 0) {
-              target.push(candidate.trim());
-            }
-          }
-        }
-        function getDictionaryEntryNames(entry) {
-          const names = [];
-          appendDictionaryNames(names, entry);
-          for (const definition of entry?.definitions || []) {
-            appendDictionaryNames(names, definition);
-          }
-          for (const frequency of entry?.frequencies || []) {
-            appendDictionaryNames(names, frequency);
-          }
-          for (const pronunciation of entry?.pronunciations || []) {
-            appendDictionaryNames(names, pronunciation);
-          }
-          return names;
-        }
-        function isNameDictionaryEntry(entry) {
-          if (!includeNameMatchMetadata || !entry || typeof entry !== 'object') {
-            return false;
-          }
-          return getDictionaryEntryNames(entry).some((name) => name.startsWith(${JSON.stringify(CHARACTER_DICTIONARY_TITLE_PREFIX)}));
-        }
-        function parseSubMinerMediaIdFromString(value) {
-          const imageMatch = value.match(/\bimg\/m(\d+)-/i);
-          if (imageMatch) {
-            const parsed = Number.parseInt(imageMatch[1], 10);
-            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
-          }
-          const titleMatch = value.match(/${CHARACTER_DICTIONARY_TITLE_PREFIX}[^\d]*(?:AniList\s*)?(\d+)/i);
-          if (titleMatch) {
-            const parsed = Number.parseInt(titleMatch[1], 10);
-            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
-          }
-          return null;
-        }
-        function parseSubMinerMediaIdCandidate(value) {
-          if (typeof value === 'number' && Number.isSafeInteger(value) && value > 0) {
-            return value;
-          }
-          if (typeof value === 'string' && /^\d+$/.test(value.trim())) {
-            const parsed = Number.parseInt(value.trim(), 10);
-            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
-          }
-          return null;
-        }
-        function collectSubMinerMediaIds(value, target) {
-          if (typeof value === 'string') {
-            const parsed = parseSubMinerMediaIdFromString(value);
-            if (parsed !== null) { target.add(parsed); }
-            return;
-          }
-          if (!value || typeof value !== 'object') {
-            return;
-          }
-          if (Array.isArray(value)) {
-            for (const item of value) { collectSubMinerMediaIds(item, target); }
-            return;
-          }
-          const mediaIdCandidates = [
-            value.subminerMediaId,
-            value.subMinerMediaId,
-            value.characterDictionaryMediaId,
-            value.data?.subminerMediaId,
-            value.data?.subMinerMediaId,
-            value.data?.characterDictionaryMediaId
-          ];
-          for (const candidate of mediaIdCandidates) {
-            const parsed = parseSubMinerMediaIdCandidate(candidate);
-            if (parsed !== null) { target.add(parsed); }
-          }
-          for (const child of Object.values(value)) {
-            collectSubMinerMediaIds(child, target);
-          }
-        }
-        function getSubMinerMediaIds(entry) {
-          const mediaIds = new Set();
-          collectSubMinerMediaIds(entry, mediaIds);
-          return mediaIds;
-        }
-        function isCurrentMediaNameDictionaryEntry(entry) {
-          if (!isNameDictionaryEntry(entry)) {
-            return false;
-          }
-          if (currentCharacterDictionaryMediaId === null) {
-            return true;
-          }
-          const mediaIds = getSubMinerMediaIds(entry);
-          return mediaIds.size === 0 || mediaIds.has(currentCharacterDictionaryMediaId);
-      }
-      function findLongestNameMatch(dictionaryEntries, textWindow) {
-        let best = null;
-        for (const dictionaryEntry of dictionaryEntries || []) {
-          if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
-          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
-          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
-            const headword = headwords[headwordIndex];
-            for (const src of headword?.sources || []) {
-              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
-              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
-              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
-              if (best === null || originalText.length > best.sourceLength) {
-                best = { dictionaryEntry, headword, headwordIndex, sourceLength: originalText.length };
-              }
-            }
-          }
-        }
-        return best;
-      }
-      function findLongestGenericMatchLength(dictionaryEntries, textWindow) {
-        let best = 0;
-        for (const dictionaryEntry of dictionaryEntries || []) {
-          if (isNameDictionaryEntry(dictionaryEntry)) { continue; }
-          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
-          for (const headword of headwords) {
-            for (const src of headword?.sources || []) {
-              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
-              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
-              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
-              if (originalText.length > best) { best = originalText.length; }
-            }
-          }
-        }
-        return best;
-      }
-      function getPreferredHeadword(dictionaryEntries, token, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
-        const currentMediaDictionaryEntries =
-          currentCharacterDictionaryMediaId === null
-            ? (dictionaryEntries || [])
-            : (dictionaryEntries || []).filter((entry) => {
-                if (!isNameDictionaryEntry(entry)) { return true; }
-                return isCurrentMediaNameDictionaryEntry(entry);
-              });
-        const exactPrimaryMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, true);
-        let matchedNameDictionary = false;
-        if (includeNameMatchMetadata) {
-          for (const dictionaryEntry of currentMediaDictionaryEntries || []) {
-            if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
-            for (const match of exactPrimaryMatches) {
-              if (match.dictionaryEntry !== dictionaryEntry) { continue; }
-              matchedNameDictionary = true;
-              break;
-            }
-            if (matchedNameDictionary) { break; }
-          }
-        }
-        const preferredMatch = exactPrimaryMatches[0];
-        if (preferredMatch) {
-          const exactFrequencyMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, false)
-            .filter((match) => sameHeadword(match, preferredMatch));
-          return {
-            term: preferredMatch.headword.term,
-            reading: preferredMatch.headword.reading,
-            wordClasses: normalizeWordClasses(preferredMatch.headword),
-            isNameMatch:
-              matchedNameDictionary || isCurrentMediaNameDictionaryEntry(preferredMatch.dictionaryEntry),
-            frequencyRank: getBestFrequencyRankForMatches(
-              exactFrequencyMatches.length > 0 ? exactFrequencyMatches : exactPrimaryMatches,
-              dictionaryPriorityByName,
-              dictionaryFrequencyModeByName
-            )
-          };
-        }
-        return null;
-      }
-`;
+async function installYomitanScanRuntime(parserWindow: BrowserWindow): Promise<void> {
+  await parserWindow.webContents.executeJavaScript(YOMITAN_SCAN_RUNTIME_INSTALL_SCRIPT, true);
+  // A fresh runtime has no candidate list; force the next scan to reinstall it.
+  yomitanScanNameCandidateKeyByWindow.delete(parserWindow);
+}

-function buildYomitanScanningScript(
-  text: string,
-  profileIndex: number,
-  scanLength: number,
-  includeNameMatchMetadata: boolean,
-  greedyNameScanEnabled: boolean,
-  currentCharacterDictionaryMediaId: number | null,
-  dictionaryPriorityByName: Record<string, number>,
-  dictionaryFrequencyModeByName: Partial<Record<string, YomitanFrequencyMode>>,
-): string {
-  return `
-    (async () => {
-      const invoke = (action, params) =>
-        new Promise((resolve, reject) => {
-          chrome.runtime.sendMessage({ action, params }, (response) => {
-            if (chrome.runtime.lastError) {
-              reject(new Error(chrome.runtime.lastError.message));
-              return;
-            }
-            if (!response || typeof response !== "object") {
-              reject(new Error("Invalid response from Yomitan backend"));
-              return;
-            }
-            if (response.error) {
-              reject(new Error(response.error.message || "Yomitan backend error"));
-              return;
-            }
-            resolve(response.result);
-          });
-        });
-${YOMITAN_SCANNING_HELPERS}
-      const includeNameMatchMetadata = ${includeNameMatchMetadata ? 'true' : 'false'};
-      const greedyNameScanEnabled = ${greedyNameScanEnabled ? 'true' : 'false'};
-      const currentCharacterDictionaryMediaId = ${
-        currentCharacterDictionaryMediaId !== null
-          ? String(currentCharacterDictionaryMediaId)
-          : 'null'
-      };
-      const dictionaryPriorityByName = ${JSON.stringify(dictionaryPriorityByName)};
-      const dictionaryFrequencyModeByName = ${JSON.stringify(dictionaryFrequencyModeByName)};
-      const text = ${JSON.stringify(text)};
-      const details = {matchType: "exact", deinflect: true};
-      const tokens = [];
-      const termsFindCache = new Map();
-      async function termsFindAt(position, windowLength) {
-        const cacheKey = position + ":" + windowLength;
-        const cached = termsFindCache.get(cacheKey);
-        if (cached) { return cached; }
-        const substring = text.substring(position, position + windowLength);
-        const result = await invoke("termsFind", { text: substring, details, optionsContext: { index: ${profileIndex} } });
-        termsFindCache.set(cacheKey, result);
-        return result;
-      }
-      function buildScanToken(position, source, preferredHeadword) {
-        const reading = typeof preferredHeadword.reading === "string" ? preferredHeadword.reading : "";
-        const segments = distributeFuriganaInflected(preferredHeadword.term, reading, source);
-        const tokenPayload = {
-          surface: segments.map((segment) => segment.text).join("") || source,
-          reading: segments.map(getSegmentReadingContribution).join(""),
-          headword: preferredHeadword.term,
-          headwordReading: reading || undefined,
-          startPos: position,
-          endPos: position + source.length,
-          isNameMatch: includeNameMatchMetadata && preferredHeadword.isNameMatch === true,
-          frequencyRank:
-            typeof preferredHeadword.frequencyRank === "number" && Number.isFinite(preferredHeadword.frequencyRank)
-              ? Math.max(1, Math.floor(preferredHeadword.frequencyRank))
-              : undefined,
-        };
-        if (Array.isArray(preferredHeadword.wordClasses) && preferredHeadword.wordClasses.length > 0) {
-          tokenPayload.wordClasses = preferredHeadword.wordClasses;
-        }
-        return tokenPayload;
-      }
-      async function findTokenAt(position, windowLength) {
-        const codePoint = text.codePointAt(position);
-        const character = String.fromCodePoint(codePoint);
-        const result = await termsFindAt(position, windowLength);
-        const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
-        const originalTextLength = typeof result?.originalTextLength === "number" ? result.originalTextLength : 0;
-        if (dictionaryEntries.length === 0 || originalTextLength <= 0 || (originalTextLength === character.length && !isCodePointJapanese(codePoint))) {
-          return { token: null, matchedLength: 0 };
-        }
-        const source = text.substring(position, position + originalTextLength);
-        const preferredHeadword = getPreferredHeadword(
-          dictionaryEntries,
-          source,
-          dictionaryPriorityByName,
-          dictionaryFrequencyModeByName
-        );
-        if (!preferredHeadword || typeof preferredHeadword.term !== "string") {
-          return { token: null, matchedLength: originalTextLength };
-        }
-        return { token: buildScanToken(position, source, preferredHeadword), matchedLength: originalTextLength };
-      }
-      // Greedy name pre-pass: character-name matches claim their spans before
-      // the left-to-right walk, so a longer generic match starting earlier
-      // (e.g. とヨー → 渡洋) cannot swallow the start of a name (ヨータ).
-      const nameTokens = [];
-      if (greedyNameScanEnabled) {
-        let namePos = 0;
-        while (namePos < text.length) {
-          const codePoint = text.codePointAt(namePos);
-          if (!isCodePointJapanese(codePoint)) {
-            namePos += String.fromCodePoint(codePoint).length;
-            continue;
-          }
-          const result = await termsFindAt(namePos, ${scanLength});
-          const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
-          const textWindow = text.substring(namePos, namePos + ${scanLength});
-          const nameMatch = findLongestNameMatch(dictionaryEntries, textWindow);
-          // A name only claims its span when no strictly longer generic word
-          // starts at the same position (a character named 空 must not split
-          // 空気). Ties go to the name. Generic matches that start earlier and
-          // overlap the name are still blocked by the reservation.
-          if (
-            !nameMatch ||
-            findLongestGenericMatchLength(dictionaryEntries, textWindow) > nameMatch.sourceLength
-          ) {
-            namePos += String.fromCodePoint(codePoint).length;
-            continue;
-          }
-          const source = text.substring(namePos, namePos + nameMatch.sourceLength);
-          nameTokens.push(buildScanToken(namePos, source, {
-            term: nameMatch.headword.term,
-            reading: nameMatch.headword.reading,
-            wordClasses: normalizeWordClasses(nameMatch.headword),
-            isNameMatch: true,
-            frequencyRank: getBestFrequencyRank(
-              nameMatch.dictionaryEntry,
-              nameMatch.headwordIndex,
-              dictionaryPriorityByName,
-              dictionaryFrequencyModeByName
-            )
-          }));
-          namePos += nameMatch.sourceLength;
-        }
-      }
-      let i = 0;
-      let nameIndex = 0;
-      while (i < text.length) {
-        while (nameIndex < nameTokens.length && nameTokens[nameIndex].startPos < i) { nameIndex += 1; }
-        const nextNameToken = nameIndex < nameTokens.length ? nameTokens[nameIndex] : null;
-        if (nextNameToken && nextNameToken.startPos === i) {
-          tokens.push(nextNameToken);
-          i = nextNameToken.endPos;
-          nameIndex += 1;
-          continue;
-        }
-        // Cap the window at the next reserved name span so a generic match
-        // cannot consume into it.
-        const windowLength = nextNameToken ? Math.min(${scanLength}, nextNameToken.startPos - i) : ${scanLength};
-        let attempt = await findTokenAt(i, windowLength);
-        // Yomitan text normalization can consume characters (whitespace,
-        // punctuation) beyond the matched term, leaving no headword whose
-        // source equals the consumed text. Retry with shorter windows so a
-        // valid prefix term (e.g. a character name before a paren) still
-        // tokenizes instead of the position being skipped.
-        let retryLength = Math.min(attempt.matchedLength, windowLength) - 1;
-        while (!attempt.token && retryLength >= 1) {
-          const retry = await findTokenAt(i, retryLength);
-          if (retry.token) {
-            attempt = retry;
-            break;
-          }
-          retryLength = Math.min(retryLength - 1, retry.matchedLength - 1);
-        }
-        if (attempt.token) {
-          tokens.push(attempt.token);
-          i += attempt.matchedLength;
-          continue;
-        }
-        i += String.fromCodePoint(text.codePointAt(i)).length;
-      }
-      return tokens;
-    })();
-  `;
+// Key of the character-name candidate list currently installed in each parser
+// window, so an unchanged list costs nothing per line.
+const yomitanScanNameCandidateKeyByWindow = new WeakMap<BrowserWindow, string>();
+
+async function ensureYomitanScanNameCandidates(
+  parserWindow: BrowserWindow,
+  nameCandidates: { key: string; forms: string[] } | null,
+  logger: LoggerLike,
+): Promise<void> {
+  const installedKey = yomitanScanNameCandidateKeyByWindow.get(parserWindow);
+  const nextKey = nameCandidates?.key ?? '';
+  if (installedKey === nextKey) {
+    return;
+  }
+
+  try {
+    await parserWindow.webContents.executeJavaScript(
+      buildYomitanScanNameCandidatesScript(nameCandidates),
+      true,
+    );
+    yomitanScanNameCandidateKeyByWindow.set(parserWindow, nextKey);
+  } catch (err) {
+    // The scan falls back to checking every position when the list is absent,
+    // so a failed install costs speed, never a missed name.
+    logger.warn?.(
+      'Failed to install Yomitan character-name scan candidates:',
+      (err as Error).message,
+    );
+    yomitanScanNameCandidateKeyByWindow.delete(parserWindow);
+  }
 }

 export async function requestYomitanParseResults(
@@ -1635,6 +963,20 @@ export async function requestYomitanParseResults(
  }
 }

+// parseText fallback for when the in-window scanner cannot run (script eval
+// failure, unexpected payload). The scanner walk is the primary tokenizer and
+// emits its own filler runs, so this extra full parse only happens on errors.
+async function requestYomitanParseFallbackTokens(
+  text: string,
+  deps: YomitanParserRuntimeDeps,
+  logger: LoggerLike,
+): Promise<YomitanScanToken[] | null> {
+  const parseResults = await requestYomitanParseResults(text, deps, logger);
+  const selectedTokens = selectYomitanParseTokens(parseResults, () => false, 'headword');
+  const parseScanTokens = selectedTokens?.map(toYomitanScanToken) ?? null;
+  return parseScanTokens && parseScanTokens.length > 0 ? parseScanTokens : null;
+}
+
 export async function requestYomitanScanTokens(
  text: string,
  deps: YomitanParserRuntimeDeps,
@@ -1642,6 +984,7 @@ export async function requestYomitanScanTokens(
  options?: {
    includeNameMatchMetadata?: boolean;
    currentCharacterDictionaryMediaId?: number | null;
+    nameCandidates?: { key: string; forms: string[] } | null;
  },
 ): Promise<YomitanScanToken[] | null> {
  const yomitanExt = deps.getYomitanExt();
@@ -1655,10 +998,6 @@ export async function requestYomitanScanTokens(
    return null;
  }

-  const parseResults = await requestYomitanParseResults(text, deps, logger);
-  const selectedParseTokens = selectYomitanParseTokens(parseResults, () => false, 'headword');
-  const parseScanTokens = selectedParseTokens?.map(toYomitanScanToken) ?? null;
-
  const metadata = await requestYomitanProfileMetadata(parserWindow, logger);
  const profileIndex = metadata?.profileIndex ?? 0;
  const scanLength = metadata?.scanLength ?? DEFAULT_YOMITAN_SCAN_LENGTH;
@@ -1669,44 +1008,63 @@ export async function requestYomitanScanTokens(
      name.startsWith(CHARACTER_DICTIONARY_TITLE_PREFIX),
    );

+  // Candidate name forms let the in-page pre-pass skip positions where no
+  // character name can start. Installed only when it changes (per media), so
+  // the per-line call stays a single tiny script.
+  const nameCandidates = greedyNameScanEnabled ? (options?.nameCandidates ?? null) : null;
+  await ensureYomitanScanNameCandidates(parserWindow, nameCandidates, logger);
+
+  const callScript = buildYomitanScanCallScript({
+    text,
+    profileIndex,
+    scanLength,
+    includeNameMatchMetadata,
+    greedyNameScanEnabled,
+    currentCharacterDictionaryMediaId:
+      typeof options?.currentCharacterDictionaryMediaId === 'number' &&
+      Number.isFinite(options.currentCharacterDictionaryMediaId) &&
+      options.currentCharacterDictionaryMediaId > 0
+        ? Math.floor(options.currentCharacterDictionaryMediaId)
+        : null,
+    dictionaryPriorityByName: metadata?.dictionaryPriorityByName ?? {},
+    dictionaryFrequencyModeByName: metadata?.dictionaryFrequencyModeByName ?? {},
+    cacheEpoch: getYomitanScanCacheEpoch(parserWindow),
+    nameCandidateKey: nameCandidates?.key ?? null,
+  });
+
  try {
-    const rawResult = await parserWindow.webContents.executeJavaScript(
-      buildYomitanScanningScript(
-        text,
-        profileIndex,
-        scanLength,
-        includeNameMatchMetadata,
-        greedyNameScanEnabled,
-        typeof options?.currentCharacterDictionaryMediaId === 'number' &&
-          Number.isFinite(options.currentCharacterDictionaryMediaId) &&
-          options.currentCharacterDictionaryMediaId > 0
-          ? Math.floor(options.currentCharacterDictionaryMediaId)
-          : null,
-        metadata?.dictionaryPriorityByName ?? {},
-        metadata?.dictionaryFrequencyModeByName ?? {},
-      ),
-      true,
-    );
-    if (isScanTokenArray(rawResult)) {
-      if (parseScanTokens && parseScanTokens.length > 0) {
-        return mergeScannerTokensIntoParseTokens(parseScanTokens, rawResult);
+    let rawResult = await parserWindow.webContents.executeJavaScript(callScript, true);
+    if (rawResult === YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL) {
+      // First request for this window, or the page reloaded and dropped the
+      // installed runtime: install and retry once. The candidate list lives in
+      // the same page state, so it has to be reinstalled alongside it.
+      await installYomitanScanRuntime(parserWindow);
+      await ensureYomitanScanNameCandidates(parserWindow, nameCandidates, logger);
+      rawResult = await parserWindow.webContents.executeJavaScript(callScript, true);
+    }
+    // The scanner reports a line where a position ran out of shrinking-window
+    // retries: it stopped short of windows an uncapped ladder would have tried,
+    // so a real term may be sitting in an unparsed run. One parseText for the
+    // line is the bounded way to get the exhaustive answer back (this is the
+    // parse the scanner replaced, and it only runs for these rare lines).
+    if (isObject(rawResult) && rawResult.retryBudgetExhausted === true) {
+      logger.info?.('Yomitan scanner exhausted its retry budget; parsing the line as a fallback.');
+      const fallbackTokens = await requestYomitanParseFallbackTokens(text, deps, logger);
+      if (fallbackTokens) {
+        return fallbackTokens;
      }
-      return rawResult;
+      rawResult = rawResult.tokens;
    }
-    if (Array.isArray(rawResult)) {
-      const selectedTokens = selectYomitanParseTokens(rawResult, () => false, 'headword');
-      return selectedTokens?.map(toYomitanScanToken) ?? null;
+    if (isScanTokenArray(rawResult)) {
+      // Filler-only results carry no dictionary match; keep the historical
+      // contract of returning null so callers fall back to raw text.
+      return rawResult.some((token) => token.isUnparsedRun !== true) ? rawResult : null;
    }
-    if (parseScanTokens && parseScanTokens.length > 0) {
-      return parseScanTokens;
-    }
-    return null;
+    logger.error('Yomitan scanner returned an unexpected payload; using parseText fallback.');
+    return await requestYomitanParseFallbackTokens(text, deps, logger);
  } catch (err) {
-    if (parseScanTokens && parseScanTokens.length > 0) {
-      return parseScanTokens;
-    }
    logger.error('Yomitan scanner request failed:', (err as Error).message);
-    return null;
+    return await requestYomitanParseFallbackTokens(text, deps, logger);
  }
 }

@@ -0,0 +1,563 @@
+// In-page Yomitan scan runtime: the scan walk that gets installed once per
+// parser window as globalThis.__subminerYomitanScan, plus the tiny per-line
+// call script. Kept separate from the host runtime module so the injected
+// script text (which is data, not executed here) does not dominate that file;
+// the helper bundle it embeds is composed in yomitan-scanning-helpers-script.ts
+// from the yomitan-*-script.ts fragments.
+import { YOMITAN_SCANNING_HELPERS } from './yomitan-scanning-helpers-script';
+
+export { CHARACTER_DICTIONARY_TITLE_PREFIX } from './yomitan-scanning-helpers-script';
+
+export type YomitanFrequencyMode = 'occurrence-based' | 'rank-based';
+
+// Bump whenever the install script below changes so already-loaded parser
+// windows re-install the new scan runtime instead of running the stale one.
+export const YOMITAN_SCAN_RUNTIME_VERSION = 12;
+export const YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL = '__subminer-yomitan-scan-runtime-missing__';
+
+export interface YomitanScanRequestParams {
+  text: string;
+  profileIndex: number;
+  scanLength: number;
+  includeNameMatchMetadata: boolean;
+  greedyNameScanEnabled: boolean;
+  currentCharacterDictionaryMediaId: number | null;
+  dictionaryPriorityByName: Record<string, number>;
+  dictionaryFrequencyModeByName: Partial<Record<string, YomitanFrequencyMode>>;
+  cacheEpoch: number;
+  /**
+   * Key of the character-name candidate list installed for the current media,
+   * or null to scan every Japanese position (see the pre-pass prefilter).
+   */
+  nameCandidateKey: string | null;
+}
+
+// Installed once per parser window (and re-installed after in-page reloads):
+// keeps V8 from re-parsing the helper bundle on every subtitle line, and hosts
+// the cross-line termsFind cache. Each subtitle line then only evaluates a tiny
+// call into globalThis.__subminerYomitanScan.
+export const YOMITAN_SCAN_RUNTIME_INSTALL_SCRIPT = String.raw`
+  (() => {
+    if (globalThis.__subminerYomitanScanVersion === ${YOMITAN_SCAN_RUNTIME_VERSION}) {
+      return true;
+    }
+    const invoke = (action, params) =>
+      new Promise((resolve, reject) => {
+        chrome.runtime.sendMessage({ action, params }, (response) => {
+          if (chrome.runtime.lastError) {
+            reject(new Error(chrome.runtime.lastError.message));
+            return;
+          }
+          if (!response || typeof response !== "object") {
+            reject(new Error("Invalid response from Yomitan backend"));
+            return;
+          }
+          if (response.error) {
+            reject(new Error(response.error.message || "Yomitan backend error"));
+            return;
+          }
+          resolve(response.result);
+        });
+      });
+    // Cross-line termsFind LRU keyed by profile + substring: subtitle lines
+    // repeat particles and inflections constantly, so most lookups hit here.
+    // Entries hold in-flight promises so concurrent identical lookups dedupe.
+    const termsFindCache = new Map();
+    // Two bounds. The key count keeps the map itself small; the accumulated
+    // dictionary-entry count stands in for retained bytes, because a single
+    // lookup over a common prefix can hold hundreds of entries with their full
+    // glossaries and a key-count cap alone would not bound that.
+    const TERMS_FIND_CACHE_LIMIT = 2000;
+    const TERMS_FIND_CACHE_DICTIONARY_ENTRY_LIMIT = 20000;
+    let termsFindCacheDictionaryEntries = 0;
+    let termsFindCacheEpoch = -1;
+    function dropCachedTermsFind(cacheKey, entry) {
+      if (termsFindCache.get(cacheKey) !== entry) { return; }
+      termsFindCache.delete(cacheKey);
+      termsFindCacheDictionaryEntries -= entry.dictionaryEntryCount;
+    }
+    // Runs on insert and again once a lookup resolves: an entry is only worth
+    // its estimated weight of 1 until then, so a single oversized response
+    // would otherwise sit in the cache forever, over the limit and reused.
+    function evictOverflowingTermsFindEntries() {
+      while (
+        termsFindCache.size > TERMS_FIND_CACHE_LIMIT ||
+        termsFindCacheDictionaryEntries > TERMS_FIND_CACHE_DICTIONARY_ENTRY_LIMIT
+      ) {
+        const oldest = termsFindCache.entries().next().value;
+        if (oldest === undefined) { break; }
+        dropCachedTermsFind(oldest[0], oldest[1]);
+      }
+    }
+    // Classification of a dictionary entry (which dictionaries it came from,
+    // which media ids it mentions) depends only on the entry object, so it is
+    // memoized for as long as that object lives. Entries are shared with the
+    // termsFind cache above, which is what makes this worth keeping: the same
+    // objects come back for every repeated lookup, on every line.
+    const dictionaryEntryNamesCache = new WeakMap();
+    const subMinerMediaIdsCache = new WeakMap();
+    const EMPTY_MEDIA_ID_SET = new Set();
+    // Only blind ladder steps are capped (see the retry loop): those are the
+    // ones that would otherwise degrade into O(scanLength) lookups at a single
+    // position. Steps the backend guides by reporting a shorter consumed length
+    // stay uncapped, so a valid prefix term is still found on lines where
+    // normalization eats a long tail.
+    const MAX_BLIND_SHRINKING_WINDOW_RETRIES = 4;
+    // Character-name candidate forms for the current media, installed
+    // separately from the per-line scan call so the per-line script stays tiny.
+    // Stored raw here; the normalized lookup index is built inside the scan,
+    // where the kana-normalization helper is in scope, and reused by key.
+    let rawNameCandidates = null;
+    let nameCandidateIndex = null;
+    globalThis.__subminerYomitanScanSetNameCandidates = (key, forms) => {
+      if (!key || !Array.isArray(forms) || forms.length === 0) {
+        rawNameCandidates = null;
+        nameCandidateIndex = null;
+        return false;
+      }
+      rawNameCandidates = { key, forms };
+      nameCandidateIndex = null;
+      return true;
+    };
+    globalThis.__subminerYomitanScanVersion = ${YOMITAN_SCAN_RUNTIME_VERSION};
+    globalThis.__subminerYomitanScan = async (scanParams) => {
+      const {
+        text,
+        profileIndex,
+        scanLength,
+        includeNameMatchMetadata,
+        greedyNameScanEnabled,
+        currentCharacterDictionaryMediaId,
+        dictionaryPriorityByName,
+        dictionaryFrequencyModeByName,
+        cacheEpoch,
+        nameCandidateKey
+      } = scanParams;
+      if (cacheEpoch !== termsFindCacheEpoch) {
+        termsFindCache.clear();
+        termsFindCacheDictionaryEntries = 0;
+        termsFindCacheEpoch = cacheEpoch;
+      }
+${YOMITAN_SCANNING_HELPERS}
+      const CAPTION_OPENING_BRACKETS = new Set(["(", "（", "[", "［", "{", "｛", "「", "『", "【", "〈", "《", "≪", "＜", "<"]);
+      function shouldEmitUnparsedRunAsToken(runText) {
+        if (!/[\p{L}\p{N}]/u.test(runText)) { return false; }
+        const firstChar = Array.from(runText.trim())[0];
+        return firstChar !== undefined && !CAPTION_OPENING_BRACKETS.has(firstChar);
+      }
+      function isLookupWorthyCodePoint(codePoint) {
+        if (isCodePointJapanese(codePoint)) { return true; }
+        return /[\p{L}\p{N}]/u.test(String.fromCodePoint(codePoint));
+      }
+      function isKanaOnlyRunText(runText) {
+        const chars = Array.from(runText);
+        return chars.length > 0 && chars.every((char) => isCodePointKana(char.codePointAt(0)));
+      }
+      const details = {matchType: "exact", deinflect: true};
+      const tokens = [];
+      async function termsFindAt(position, windowLength) {
+        const substring = text.substring(position, position + windowLength);
+        const cacheKey = profileIndex + "\u0000" + substring;
+        const cached = termsFindCache.get(cacheKey);
+        if (cached !== undefined) {
+          termsFindCache.delete(cacheKey);
+          termsFindCache.set(cacheKey, cached);
+          return await cached.promise;
+        }
+        // An in-flight lookup counts as one entry until it resolves; the real
+        // weight replaces that estimate once the result is known.
+        const entry = { promise: null, dictionaryEntryCount: 1 };
+        entry.promise = invoke("termsFind", { text: substring, details, optionsContext: { index: profileIndex } })
+          .then((result) => {
+            const resolvedCount =
+              1 + (Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries.length : 0);
+            const isCached = termsFindCache.get(cacheKey) === entry;
+            if (isCached) {
+              termsFindCacheDictionaryEntries += resolvedCount - entry.dictionaryEntryCount;
+            }
+            entry.dictionaryEntryCount = resolvedCount;
+            // The real weight can push the cache over its budget, and a single
+            // response can exceed it on its own, so re-check here.
+            if (isCached) { evictOverflowingTermsFindEntries(); }
+            return result;
+          });
+        termsFindCache.set(cacheKey, entry);
+        termsFindCacheDictionaryEntries += entry.dictionaryEntryCount;
+        evictOverflowingTermsFindEntries();
+        try {
+          return await entry.promise;
+        } catch (error) {
+          dropCachedTermsFind(cacheKey, entry);
+          throw error;
+        }
+      }
+      // Text the walk skips accumulates into unparsed runs, mirroring the
+      // filler chunks the parseText segmentation used to provide: runs stay
+      // hoverable (flagged isUnparsedRun) unless they are punctuation-only or
+      // caption-style asides, and kana continuations of a longer headword
+      // extend the previous token instead.
+      function flushUnparsedRun(runStart, runEnd) {
+        if (runStart === null || runEnd <= runStart) { return; }
+        const runText = text.substring(runStart, runEnd);
+        const previousToken = tokens[tokens.length - 1];
+        if (
+          previousToken &&
+          previousToken.endPos === runStart &&
+          isKanaOnlyRunText(runText) &&
+          typeof previousToken.headword === "string" &&
+          previousToken.headword.length > previousToken.surface.length &&
+          previousToken.headword.startsWith(previousToken.surface + runText)
+        ) {
+          previousToken.surface += runText;
+          // The run is kana-only, so its reading is itself: append it or the
+          // reading stops covering the surface, which disables the known-word
+          // reading fallback (isCompleteReadingForSurface) downstream.
+          previousToken.reading += runText;
+          // The run is kana-only, so its reading is itself: append it or the
+          // reading stops covering the surface, which disables the known-word
+          // reading fallback (isCompleteReadingForSurface) downstream.
+          previousToken.endPos = runEnd;
+          return;
+        }
+        if (!shouldEmitUnparsedRunAsToken(runText)) { return; }
+        tokens.push({
+          surface: runText,
+          reading: "",
+          headword: runText,
+          startPos: runStart,
+          endPos: runEnd,
+          isUnparsedRun: true
+        });
+      }
+      function buildScanToken(position, source, preferredHeadword) {
+        const reading = typeof preferredHeadword.reading === "string" ? preferredHeadword.reading : "";
+        const segments = distributeFuriganaInflected(preferredHeadword.term, reading, source);
+        const tokenPayload = {
+          surface: segments.map((segment) => segment.text).join("") || source,
+          reading: segments.map(getSegmentReadingContribution).join(""),
+          headword: preferredHeadword.term,
+          headwordReading: reading || undefined,
+          startPos: position,
+          endPos: position + source.length,
+          isNameMatch: includeNameMatchMetadata && preferredHeadword.isNameMatch === true,
+          frequencyRank:
+            typeof preferredHeadword.frequencyRank === "number" && Number.isFinite(preferredHeadword.frequencyRank)
+              ? Math.max(1, Math.floor(preferredHeadword.frequencyRank))
+              : undefined,
+        };
+        if (Array.isArray(preferredHeadword.wordClasses) && preferredHeadword.wordClasses.length > 0) {
+          tokenPayload.wordClasses = preferredHeadword.wordClasses;
+        }
+        return tokenPayload;
+      }
+      // findTokenAt plus the shrinking-window ladder below it: Yomitan text
+      // normalization can consume characters (whitespace, punctuation) beyond
+      // the matched term, leaving no headword whose source equals the consumed
+      // text. Retry with shorter windows so a valid prefix term (e.g. a
+      // character name before a paren) still tokenizes instead of the position
+      // being skipped.
+      // Every window at or above the consumed length repeats the same result,
+      // so the next informative window sits just below it. A lookup that
+      // consumed its whole window reports nothing to aim at, and the step down
+      // from it is a blind guess: only those are budgeted.
+      // The window can run past the end of the line, so blindness is judged
+      // against the text the lookup actually saw.
+      // Set when a position stopped short of windows an uncapped ladder would
+      // still have tried; the line then escalates to parseText at the end.
+      let blindRetryBudgetExhausted = false;
+      async function resolveTokenAt(position, windowLength) {
+        let attempt = await findTokenAt(position, windowLength);
+        const scannedLength = Math.min(windowLength, text.length - position);
+        let retryLength = Math.min(attempt.matchedLength, scannedLength) - 1;
+        let stepIsBlind = attempt.matchedLength >= scannedLength;
+        let blindRetriesRemaining = MAX_BLIND_SHRINKING_WINDOW_RETRIES;
+        while (!attempt.token && retryLength >= 1) {
+          if (stepIsBlind) {
+            if (blindRetriesRemaining <= 0) {
+              blindRetryBudgetExhausted = true;
+              break;
+            }
+            blindRetriesRemaining -= 1;
+          }
+          const retry = await findTokenAt(position, retryLength);
+          if (retry.token) { return retry; }
+          const guidedLength = retry.matchedLength - 1;
+          stepIsBlind = guidedLength >= retryLength - 1;
+          retryLength = Math.min(retryLength - 1, guidedLength);
+        }
+        return attempt;
+      }
+      async function findTokenAt(position, windowLength) {
+        const codePoint = text.codePointAt(position);
+        const character = String.fromCodePoint(codePoint);
+        const result = await termsFindAt(position, windowLength);
+        const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
+        const originalTextLength = typeof result?.originalTextLength === "number" ? result.originalTextLength : 0;
+        if (dictionaryEntries.length === 0 || originalTextLength <= 0 || (originalTextLength === character.length && !isCodePointJapanese(codePoint))) {
+          return { token: null, matchedLength: 0 };
+        }
+        const source = text.substring(position, position + originalTextLength);
+        const preferredHeadword = getPreferredHeadword(
+          dictionaryEntries,
+          source,
+          dictionaryPriorityByName,
+          dictionaryFrequencyModeByName
+        );
+        if (!preferredHeadword || typeof preferredHeadword.term !== "string") {
+          return { token: null, matchedLength: originalTextLength };
+        }
+        return { token: buildScanToken(position, source, preferredHeadword), matchedLength: originalTextLength };
+      }
+      // Kana normalization folds halfwidth katakana one code point to one, so an
+      // unvoiced halfwidth spelling prefix-matches a candidate form like any
+      // other. What it cannot fold is a voiced pair: ｶ + ﾞ stays two characters
+      // where the candidate form carries the single が, so the comparison fails
+      // at that character. That break can sit anywhere inside the name, not
+      // just at its first character (山ｶﾞｸ starts on a kanji), so the bypass is
+      // keyed on the region a candidate could cover, not on how it starts.
+      function isHalfwidthKanaVoicedMarkCodePoint(codePoint) {
+        return codePoint === 0xff9e || codePoint === 0xff9f;
+      }
+      // Build (once per candidate list) a first-character bucket index of the
+      // normalized name forms, so the pre-pass can reject a position with a
+      // single map hit instead of a backend round trip.
+      if (rawNameCandidates && nameCandidateIndex?.key !== rawNameCandidates.key) {
+        const byFirstChar = new Map();
+        for (const form of rawNameCandidates.forms) {
+          const normalized = typeof form === "string" ? convertKatakanaToHiragana(form.trim()) : "";
+          if (!normalized) { continue; }
+          const bucket = byFirstChar.get(normalized[0]);
+          if (bucket) { bucket.push(normalized); } else { byFirstChar.set(normalized[0], [normalized]); }
+        }
+        nameCandidateIndex = byFirstChar.size > 0 ? { key: rawNameCandidates.key, byFirstChar } : null;
+      } else if (!rawNameCandidates) {
+        nameCandidateIndex = null;
+      }
+      // Only meaningful when the installed list matches the media this scan is
+      // for; otherwise fall back to scanning every position.
+      const activeNameCandidateIndex =
+        nameCandidateKey !== null && nameCandidateIndex?.key === nameCandidateKey
+          ? nameCandidateIndex
+          : null;
+      const normalizedText = activeNameCandidateIndex ? convertKatakanaToHiragana(text) : "";
+      // Yomitan collapses emphatic sequences before matching (すっっごーーい →
+      // すごい), so a stretched name still resolves to its entry. Skipping these
+      // characters keeps such spellings candidates; the filter only ever grows
+      // the probe set, so a false positive costs one lookup, never a name.
+      const EMPHATIC_SKIP_CHARS = new Set(["ぁ", "ぃ", "ぅ", "ぇ", "ぉ", "っ", "ゃ", "ゅ", "ょ", "ー"]);
+      function matchesCandidateFormAt(form, position) {
+        let textIndex = position;
+        for (let formIndex = 0; formIndex < form.length; formIndex += 1) {
+          while (
+            textIndex < normalizedText.length &&
+            normalizedText[textIndex] !== form[formIndex] &&
+            EMPHATIC_SKIP_CHARS.has(normalizedText[textIndex])
+          ) {
+            textIndex += 1;
+          }
+          if (normalizedText[textIndex] !== form[formIndex]) { return false; }
+          textIndex += 1;
+        }
+        return true;
+      }
+      // Where the folding gives up, listed once per line. Matching may skip any
+      // number of emphatic characters on its way through a form (山ーーーーーーｶﾞｸ),
+      // so there is no shorter honest bound than the window a name lookup
+      // covers: scanLength. The list is almost always empty, which is what
+      // keeps the check below free on ordinary lines.
+      const halfwidthVoicedMarkPositions = [];
+      if (activeNameCandidateIndex) {
+        for (let index = 0; index < text.length; index += 1) {
+          if (isHalfwidthKanaVoicedMarkCodePoint(text.charCodeAt(index))) {
+            halfwidthVoicedMarkPositions.push(index);
+          }
+        }
+      }
+      function hasHalfwidthVoicedMarkInScanWindow(position) {
+        const end = position + scanLength;
+        for (const markPosition of halfwidthVoicedMarkPositions) {
+          if (markPosition >= position && markPosition < end) { return true; }
+        }
+        return false;
+      }
+      // A name written ｶﾞ... folds to か + ﾞ, so its first character never leads
+      // to the が bucket the candidate form is filed under. Nothing else can
+      // find it, so such a position is always worth a probe.
+      function startsHalfwidthVoicedPair(position, codePoint) {
+        if (codePoint < 0xff66 || codePoint > 0xff9d) { return false; }
+        return isHalfwidthKanaVoicedMarkCodePoint(text.charCodeAt(position + 1));
+      }
+      function couldNameStartAt(position, codePoint) {
+        // Nothing starts with a combining voiced mark, whether or not the
+        // prefilter is active.
+        if (isHalfwidthKanaVoicedMarkCodePoint(codePoint)) { return false; }
+        if (!activeNameCandidateIndex) { return true; }
+        const bucket = activeNameCandidateIndex.byFirstChar.get(normalizedText[position]);
+        if (!bucket) {
+          // No candidate begins with this character, and the window search
+          // below would only ever say yes to positions like this one, so an
+          // unrelated ｶﾞ elsewhere in the line must not drag them in.
+          return startsHalfwidthVoicedPair(position, codePoint);
+        }
+        for (const form of bucket) {
+          if (matchesCandidateFormAt(form, position)) { return true; }
+        }
+        // A candidate does start here but did not match: an unfoldable voiced
+        // pair anywhere in the window is a reason the comparison could not see
+        // it (山ｶﾞｸ, 山ーーーーーーｶﾞｸ), so probe rather than drop the name.
+        return hasHalfwidthVoicedMarkInScanWindow(position);
+      }
+      // Greedy name pre-pass: character-name matches claim their spans before
+      // the left-to-right walk, so a longer generic match starting earlier
+      // (e.g. とヨー → 渡洋) cannot swallow the start of a name (ヨータ).
+      const nameTokens = [];
+      if (greedyNameScanEnabled) {
+        let namePos = 0;
+        while (namePos < text.length) {
+          const codePoint = text.codePointAt(namePos);
+          if (!isCodePointJapanese(codePoint) || !couldNameStartAt(namePos, codePoint)) {
+            namePos += String.fromCodePoint(codePoint).length;
+            continue;
+          }
+          const result = await termsFindAt(namePos, scanLength);
+          const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
+          const textWindow = text.substring(namePos, namePos + scanLength);
+          const nameMatch = findLongestNameMatch(dictionaryEntries, textWindow);
+          // A name only claims its span when no strictly longer generic word
+          // starts at the same position (a character named 空 must not split
+          // 空気). Ties go to the name. Generic matches that start earlier and
+          // overlap the name are still blocked by the reservation.
+          if (
+            !nameMatch ||
+            findLongestGenericMatchLength(dictionaryEntries, textWindow) > nameMatch.sourceLength
+          ) {
+            namePos += String.fromCodePoint(codePoint).length;
+            continue;
+          }
+          const source = text.substring(namePos, namePos + nameMatch.sourceLength);
+          nameTokens.push(buildScanToken(namePos, source, {
+            term: nameMatch.headword.term,
+            reading: nameMatch.headword.reading,
+            wordClasses: normalizeWordClasses(nameMatch.headword),
+            isNameMatch: true,
+            frequencyRank: getBestFrequencyRank(
+              nameMatch.dictionaryEntry,
+              nameMatch.headwordIndex,
+              dictionaryPriorityByName,
+              dictionaryFrequencyModeByName
+            )
+          }));
+          namePos += nameMatch.sourceLength;
+        }
+      }
+      // First reserved name span that a match ending at endPos would leave
+      // half-consumed. Spans the match covers entirely are not returned: those
+      // lose to the longer word instead of splitting it.
+      function findSplitNameToken(startIndex, endPos) {
+        for (let index = startIndex; index < nameTokens.length; index += 1) {
+          const nameToken = nameTokens[index];
+          if (nameToken.startPos >= endPos) { return null; }
+          if (nameToken.endPos > endPos) { return nameToken; }
+        }
+        return null;
+      }
+      let i = 0;
+      let nameIndex = 0;
+      let unparsedRunStart = null;
+      while (i < text.length) {
+        while (nameIndex < nameTokens.length && nameTokens[nameIndex].startPos < i) { nameIndex += 1; }
+        const nextNameToken = nameIndex < nameTokens.length ? nameTokens[nameIndex] : null;
+        if (nextNameToken && nextNameToken.startPos === i) {
+          flushUnparsedRun(unparsedRunStart, i);
+          unparsedRunStart = null;
+          tokens.push(nextNameToken);
+          i = nextNameToken.endPos;
+          nameIndex += 1;
+          continue;
+        }
+        const codePoint = text.codePointAt(i);
+        // Punctuation and whitespace can never start a token: skip the backend
+        // round trip entirely. Latin letters and digits stay lookup-worthy
+        // (terms like Tシャツ start on an ASCII letter).
+        if (!isLookupWorthyCodePoint(codePoint)) {
+          if (unparsedRunStart === null) { unparsedRunStart = i; }
+          i += String.fromCodePoint(codePoint).length;
+          continue;
+        }
+        // A reservation only outranks generic matches that would cut into it.
+        // Look the position up unrestricted first: a generic word that starts
+        // earlier and covers the whole name span (写真 over a character named
+        // 真) is the better reading, so the reservation yields rather than
+        // splitting the word. Only a match that ends inside a name span gets
+        // re-run against a window capped at that span.
+        let attempt = await resolveTokenAt(i, scanLength);
+        if (attempt.token) {
+          const splitNameToken = findSplitNameToken(nameIndex, attempt.token.endPos);
+          if (splitNameToken) {
+            attempt = await resolveTokenAt(i, splitNameToken.startPos - i);
+          }
+        }
+        if (attempt.token) {
+          flushUnparsedRun(unparsedRunStart, i);
+          unparsedRunStart = null;
+          tokens.push(attempt.token);
+          i += attempt.matchedLength;
+          continue;
+        }
+        if (unparsedRunStart === null) { unparsedRunStart = i; }
+        i += String.fromCodePoint(text.codePointAt(i)).length;
+      }
+      flushUnparsedRun(unparsedRunStart, text.length);
+      if (blindRetryBudgetExhausted) {
+        // A position gave up with shorter windows still worth trying. The walk
+        // is the only tokenizer now, so stopping there would leave a real term
+        // as an unparsed run; report it so the host can spend one parseText on
+        // the line instead of letting the ladder run to O(scanLength) lookups.
+        return { tokens, retryBudgetExhausted: true };
+      }
+      return tokens;
+    };
+    return true;
+  })();
+`;
+
+// Installs (or clears) the character-name candidate forms for the current
+// media. Runs only when the list changes, not per line. Passing null restores
+// the exhaustive every-position pre-pass.
+export function buildYomitanScanNameCandidatesScript(
+  nameCandidates: { key: string; forms: string[] } | null,
+): string {
+  if (!nameCandidates) {
+    return `
+    (() => {
+      if (typeof globalThis.__subminerYomitanScanSetNameCandidates !== "function") {
+        return false;
+      }
+      return globalThis.__subminerYomitanScanSetNameCandidates(null, null);
+    })();
+  `;
+  }
+
+  return `
+    (() => {
+      if (typeof globalThis.__subminerYomitanScanSetNameCandidates !== "function") {
+        return false;
+      }
+      return globalThis.__subminerYomitanScanSetNameCandidates(
+        ${JSON.stringify(nameCandidates.key)},
+        ${JSON.stringify(nameCandidates.forms)}
+      );
+    })();
+  `;
+}
+
+export function buildYomitanScanCallScript(params: YomitanScanRequestParams): string {
+  return `
+    (async () => {
+      if (typeof globalThis.__subminerYomitanScan !== "function") {
+        return ${JSON.stringify(YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL)};
+      }
+      return await globalThis.__subminerYomitanScan(${JSON.stringify(params)});
+    })();
+  `;
+}
@@ -0,0 +1,304 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { requestYomitanScanTokens } from './yomitan-parser-runtime';
+import {
+  countTermsFindLookups,
+  createNameScanDeps,
+  NAME_SCAN_WORDS,
+} from './yomitan-scan-test-harness';
+
+// Behaviour of the in-page scan runtime around character names and kana:
+// which positions the greedy pre-pass probes, and what the walk makes of
+// halfwidth spellings. Driven end to end through requestYomitanScanTokens
+// because the runtime only exists inside the parser window.
+
+const NAME_SCAN_LINE = 'ミナトはまだ学校にいない';
+
+test('requestYomitanScanTokens skips name pre-pass lookups where no candidate name can start', async () => {
+  const exhaustiveLookups: string[] = [];
+  const exhaustive = await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    createNameScanDeps(exhaustiveLookups),
+    { error: () => undefined },
+    { includeNameMatchMetadata: true },
+  );
+
+  const prefilteredLookups: string[] = [];
+  const prefiltered = await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    createNameScanDeps(prefilteredLookups),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      // Terms and readings the generated dictionary exposes for this media.
+      nameCandidates: { key: 'media-1', forms: ['ミナト', 'みなと'] },
+    },
+  );
+
+  // Same tokenization, including the name match, with fewer round trips.
+  assert.deepEqual(prefiltered, exhaustive);
+  assert.equal(prefiltered?.[0]?.surface, 'ミナト');
+  assert.equal(prefiltered?.[0]?.isNameMatch, true);
+  assert.ok(
+    prefilteredLookups.length < exhaustiveLookups.length,
+    `expected fewer lookups with candidates (${prefilteredLookups.length} vs ${exhaustiveLookups.length})`,
+  );
+  // Mid-token positions are exactly what the pre-pass used to probe (a name can
+  // start mid-token); with candidates they cost nothing, while the main walk's
+  // own token-start lookups are unaffected.
+  assert.ok(countTermsFindLookups(exhaustiveLookups, '校に') > 0);
+  assert.equal(countTermsFindLookups(prefilteredLookups, '校に'), 0);
+});
+
+test('requestYomitanScanTokens matches a katakana name from its kana-normalized candidate form', async () => {
+  const lookups: string[] = [];
+  const result = await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    createNameScanDeps(lookups),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      // Only the hiragana reading is listed; the katakana surface in the line
+      // must still be found through kana normalization.
+      nameCandidates: { key: 'media-1', forms: ['みなと'] },
+    },
+  );
+
+  assert.equal(result?.[0]?.surface, 'ミナト');
+  assert.equal(result?.[0]?.isNameMatch, true);
+});
+
+// Kana normalization folds halfwidth katakana, so a name written that way does
+// prefix-match a candidate form — but only if the position counts as Japanese
+// in the first place. The generic word here reaches into the name, so only a
+// pre-pass reservation can keep the name whole.
+const HALFWIDTH_NAME_SCAN_WORDS: Array<[string, string, string, boolean]> = [
+  ['ﾈｺ', 'ネコ', 'ねこ', false],
+  ['まだﾐ', 'まだミ', 'まだみ', false],
+  ['まだ', 'まだ', 'まだ', false],
+  ['ﾐﾅﾄ', 'ミナト', 'みなと', true],
+];
+
+test('requestYomitanScanTokens probes halfwidth katakana positions during the name pre-pass', async () => {
+  const lookups: string[] = [];
+  const result = await requestYomitanScanTokens(
+    'ﾈｺまだﾐﾅﾄ',
+    createNameScanDeps(lookups, HALFWIDTH_NAME_SCAN_WORDS),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      // Fullwidth forms only, as the generated dictionary stores them.
+      nameCandidates: { key: 'media-1', forms: ['ミナト', 'みなと'] },
+    },
+  );
+
+  assert.equal(countTermsFindLookups(lookups, 'ﾐﾅﾄ'), 1);
+  // ｺ is mid-token, so only the pre-pass would ever look it up, and it matches
+  // no candidate: folding halfwidth made those positions indexable, so they no
+  // longer cost a round trip apiece.
+  assert.equal(countTermsFindLookups(lookups, 'ｺ'), 0);
+  assert.deepEqual(
+    result?.map((token) => token.surface),
+    ['ﾈｺ', 'まだ', 'ﾐﾅﾄ'],
+  );
+  assert.equal(result?.[2]?.isNameMatch, true);
+  // The reading is written the way the fullwidth katakana path writes it
+  // (surface spelling, fullwidth): halfwidth kana is not kana to the known-word
+  // and frequency code downstream, and an empty reading there disables the
+  // reading fallback entirely.
+  assert.equal(result?.[2]?.reading, 'ミナト');
+  assert.equal(result?.[2]?.headwordReading, 'みなと');
+});
+
+test('a voiced halfwidth name still bypasses the candidate prefilter', async () => {
+  const lookups: string[] = [];
+  const result = await requestYomitanScanTokens(
+    'まだｶﾞｸ',
+    createNameScanDeps(lookups, [
+      ['まだｶ', 'まだカ', 'まだか', false],
+      ['まだ', 'まだ', 'まだ', false],
+      ['ｶﾞｸ', 'ガク', 'がく', true],
+    ]),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      nameCandidates: { key: 'media-1', forms: ['ガク', 'がく'] },
+    },
+  );
+
+  // ｶ + ﾞ folds to か + ﾞ, which cannot prefix-match が, so the prefilter would
+  // drop this position; the voiced-mark bypass is what keeps the name.
+  assert.deepEqual(
+    result?.map((token) => token.surface),
+    ['まだ', 'ｶﾞｸ'],
+  );
+  assert.equal(result?.[1]?.isNameMatch, true);
+});
+
+test('an unrelated halfwidth voiced word does not restore the exhaustive pre-pass', async () => {
+  const baseline: string[] = [];
+  await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    createNameScanDeps(baseline),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      nameCandidates: { key: 'media-1', forms: ['ミナト', 'みなと'] },
+    },
+  );
+
+  const withVoicedTail: string[] = [];
+  await requestYomitanScanTokens(
+    `${NAME_SCAN_LINE}ｶﾞ`,
+    createNameScanDeps(withVoicedTail),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      nameCandidates: { key: 'media-1', forms: ['ミナト', 'みなと'] },
+    },
+  );
+
+  // Mid-token positions are the ones only the pre-pass would ever probe. A ｶﾞ
+  // anywhere in the line used to drag every position within scanLength of it
+  // back in; now only the voiced pair itself, which the fold cannot index, is
+  // added to what the line already looked up.
+  for (const midTokenPrefix of ['ナト', 'だ学', '校に', 'ない']) {
+    assert.equal(countTermsFindLookups(baseline, midTokenPrefix), 0, midTokenPrefix);
+    assert.equal(countTermsFindLookups(withVoicedTail, midTokenPrefix), 0, midTokenPrefix);
+  }
+  assert.ok(
+    withVoicedTail.length - baseline.length <= 3,
+    `expected the ｶﾞ tail to add only its own lookups, saw ${JSON.stringify(withVoicedTail)}`,
+  );
+});
+
+test('a mixed-width voiced name survives the candidate prefilter', async () => {
+  const lookups: string[] = [];
+  const result = await requestYomitanScanTokens(
+    'まだ山ｶﾞｸ',
+    createNameScanDeps(lookups, [
+      ['まだ山', 'まだ山', 'まだやま', false],
+      ['まだ', 'まだ', 'まだ', false],
+      ['山ｶﾞｸ', '山ガク', 'やまがく', true],
+    ]),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      nameCandidates: { key: 'media-1', forms: ['山ガク', 'やまがく'] },
+    },
+  );
+
+  // The name starts on a kanji, so the fold only breaks mid-name: 山ｶﾞｸ
+  // normalizes to 山かﾞく, which still cannot match the candidate 山がく. The
+  // bypass is keyed on the scan window rather than the first character, so the
+  // position is still probed and the generic まだ山 cannot swallow the 山.
+  assert.deepEqual(
+    result?.map((token) => token.surface),
+    ['まだ', '山ｶﾞｸ'],
+  );
+  assert.equal(result?.[1]?.isNameMatch, true);
+});
+
+test('a stretched mixed-width voiced name survives the candidate prefilter', async () => {
+  const lookups: string[] = [];
+  const result = await requestYomitanScanTokens(
+    'まだ山ーーーーーーｶﾞｸ',
+    createNameScanDeps(lookups, [
+      ['まだ山', 'まだ山', 'まだやま', false],
+      ['まだ', 'まだ', 'まだ', false],
+      ['山ーーーーーーｶﾞｸ', '山ガク', 'やまがく', true],
+    ]),
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      nameCandidates: { key: 'media-1', forms: ['山ガク', 'やまがく'] },
+    },
+  );
+
+  // Matching skips any number of emphatic characters, so the voiced mark that
+  // defeats the fold can sit arbitrarily far into the name: the search for it
+  // has to cover the whole lookup window, not a multiple of the form length.
+  assert.deepEqual(
+    result?.map((token) => token.surface),
+    ['まだ', '山ーーーーーーｶﾞｸ'],
+  );
+  assert.equal(result?.[1]?.isNameMatch, true);
+});
+
+test('halfwidth voiced kana compose into the reading instead of leaving a stray mark', async () => {
+  const lookups: string[] = [];
+  const result = await requestYomitanScanTokens(
+    'ｶﾞｸ ﾊﾟﾝ',
+    createNameScanDeps(lookups, [
+      ['ｶﾞｸ', 'ガク', 'がく', false],
+      ['ﾊﾟﾝ', 'パン', 'ぱん', false],
+    ]),
+    { error: () => undefined },
+    { includeNameMatchMetadata: true },
+  );
+
+  // The name pre-pass runs over every position here (no candidate list), but a
+  // standalone voiced mark can never start a name, so it costs no lookup.
+  assert.equal(countTermsFindLookups(lookups, 'ﾞ'), 0);
+  assert.equal(countTermsFindLookups(lookups, 'ﾟ'), 0);
+  const readings = (result ?? [])
+    .filter((token) => token.isUnparsedRun !== true)
+    .map((token) => [token.surface, token.reading]);
+  assert.deepEqual(readings, [
+    ['ｶﾞｸ', 'ガク'],
+    ['ﾊﾟﾝ', 'パン'],
+  ]);
+});
+
+test('requestYomitanScanTokens falls back to the exhaustive name scan without candidates', async () => {
+  const withoutLookups: string[] = [];
+  const withoutCandidates = await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    createNameScanDeps(withoutLookups),
+    { error: () => undefined },
+    { includeNameMatchMetadata: true, currentCharacterDictionaryMediaId: 1, nameCandidates: null },
+  );
+
+  assert.equal(withoutCandidates?.[0]?.isNameMatch, true);
+  // No candidate list means every Japanese position is probed, as before.
+  assert.ok(countTermsFindLookups(withoutLookups, '校に') > 0);
+});
+
+test('requestYomitanScanTokens reinstalls name candidates when the media changes', async () => {
+  const lookups: string[] = [];
+  const deps = createNameScanDeps(lookups);
+
+  // First media's candidates cannot match this line's name.
+  const otherMedia = await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    deps,
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 2,
+      nameCandidates: { key: 'media-2', forms: ['カズマ'] },
+    },
+  );
+  assert.equal(otherMedia?.[0]?.isNameMatch, undefined);
+
+  const correctMedia = await requestYomitanScanTokens(
+    NAME_SCAN_LINE,
+    deps,
+    { error: () => undefined },
+    {
+      includeNameMatchMetadata: true,
+      currentCharacterDictionaryMediaId: 1,
+      nameCandidates: { key: 'media-1', forms: ['ミナト'] },
+    },
+  );
+  assert.equal(correctMedia?.[0]?.surface, 'ミナト');
+  assert.equal(correctMedia?.[0]?.isNameMatch, true);
+});
@@ -0,0 +1,166 @@
+// Shared harness for the Yomitan parser-runtime and scan-runtime tests: fake
+// parser-window deps whose injected scripts run in a vm context, plus the
+// backend stubs the scanner tests drive them with. Kept out of the test files
+// so the runtime tests and the in-page scanner tests can share one setup.
+import * as vm from 'node:vm';
+
+export function createDeps(
+  executeJavaScript: (script: string) => Promise<unknown>,
+  options?: {
+    createYomitanExtensionWindow?: (pageName: string) => Promise<unknown>;
+  },
+) {
+  const parserWindow = {
+    isDestroyed: () => false,
+    webContents: {
+      executeJavaScript: async (script: string) => await executeJavaScript(script),
+    },
+  };
+
+  return {
+    getYomitanExt: () => ({ id: 'ext-id' }) as never,
+    getYomitanParserWindow: () => parserWindow as never,
+    setYomitanParserWindow: () => undefined,
+    getYomitanParserReadyPromise: () => null,
+    setYomitanParserReadyPromise: () => undefined,
+    getYomitanParserInitPromise: () => null,
+    setYomitanParserInitPromise: () => undefined,
+    createYomitanExtensionWindow: options?.createYomitanExtensionWindow as never,
+  };
+}
+
+function createYomitanScriptSandbox(handler: (action: string, params: unknown) => unknown) {
+  return {
+    chrome: {
+      runtime: {
+        lastError: null,
+        sendMessage: (
+          payload: { action?: string; params?: unknown },
+          callback: (response: { result?: unknown; error?: { message?: string } }) => void,
+        ) => {
+          try {
+            callback({ result: handler(payload.action ?? '', payload.params) });
+          } catch (error) {
+            callback({ error: { message: (error as Error).message } });
+          }
+        },
+      },
+    },
+    Array,
+    Error,
+    JSON,
+    Map,
+    Math,
+    Number,
+    Object,
+    Promise,
+    RegExp,
+    Set,
+    String,
+  };
+}
+
+export async function runInjectedYomitanScript(
+  script: string,
+  handler: (action: string, params: unknown) => unknown,
+): Promise<unknown> {
+  return await vm.runInNewContext(script, createYomitanScriptSandbox(handler));
+}
+
+// Persistent page context shared across executeJavaScript calls, matching the
+// real parser window: the scan runtime is installed once via
+// globalThis.__subminerYomitanScan and per-line calls reuse it (and its
+// cross-line termsFind cache).
+function createPersistentYomitanScriptRunner(
+  handler: (action: string, params: unknown) => unknown,
+): (script: string) => Promise<unknown> {
+  const context = vm.createContext(createYomitanScriptSandbox(handler));
+  return async (script: string) => await vm.runInContext(script, context);
+}
+
+// Deps whose parser window executes every injected script (profile metadata,
+// scan runtime install, per-line scan calls, parseText fallback) inside one
+// persistent vm context, dispatching backend actions to `handler`.
+export function createScanDeps(
+  handler: (action: string, params: unknown) => unknown,
+  options?: { onScript?: (script: string) => void },
+) {
+  const runScript = createPersistentYomitanScriptRunner(handler);
+  return createDeps(async (script) => {
+    options?.onScript?.(script);
+    return await runScript(script);
+  });
+}
+
+export function countTermsFindLookups(lookups: string[], prefix: string): number {
+  return lookups.filter((lookupText) => lookupText.startsWith(prefix)).length;
+}
+
+// Backend stub for the greedy name pre-pass: one character name (ミナト) in a
+// line of ordinary words, with the SubMiner character dictionary enabled.
+export const NAME_SCAN_WORDS: Array<[string, string, string, boolean]> = [
+  ['ミナト', 'ミナト', 'みなと', true],
+  ['は', 'は', 'は', false],
+  ['まだ', 'まだ', 'まだ', false],
+  ['学校', '学校', 'がっこう', false],
+  ['に', 'に', 'に', false],
+  ['いない', 'いる', 'いる', false],
+];
+
+export function createNameScanDeps(
+  lookups: string[],
+  words: Array<[string, string, string, boolean]> = NAME_SCAN_WORDS,
+) {
+  return createScanDeps((action, params) => {
+    if (action === 'optionsGetFull') {
+      return {
+        profileCurrent: 0,
+        profiles: [
+          {
+            options: {
+              scanning: { length: 40 },
+              dictionaries: [
+                { name: 'JMdict', enabled: true, id: 0 },
+                {
+                  name: 'SubMiner Character Dictionary (AniList 1)',
+                  enabled: true,
+                  id: 1,
+                },
+              ],
+            },
+          },
+        ],
+      };
+    }
+    if (action === 'getDictionaryInfo') {
+      return [];
+    }
+    if (action !== 'termsFind') {
+      throw new Error(`unexpected action: ${action}`);
+    }
+    const text = (params as { text?: string } | undefined)?.text ?? '';
+    lookups.push(text);
+    for (const [surface, term, reading, isName] of words) {
+      if (text.startsWith(surface)) {
+        return {
+          originalTextLength: surface.length,
+          dictionaryEntries: [
+            {
+              headwords: [
+                {
+                  term,
+                  reading,
+                  sources: [{ originalText: surface, isPrimary: true, matchType: 'exact' }],
+                },
+              ],
+              definitions: [
+                { dictionary: isName ? 'SubMiner Character Dictionary (AniList 1)' : 'JMdict' },
+              ],
+            },
+          ],
+        };
+      }
+    }
+    return { originalTextLength: 0, dictionaryEntries: [] };
+  });
+}
@@ -0,0 +1,21 @@
+// Helper bundle for the in-page Yomitan scan runtime, composed from the
+// fragments below. Injected as text into the parser window by
+// yomitan-scan-runtime-script.ts, so it is data here, not code this process
+// runs. The fragments are concatenated into a single function body and share
+// one lexical scope: every function in them is hoisted, but the constants are
+// not, so kana stays first — the later fragments read its ranges as they run.
+import { YOMITAN_DICTIONARY_CLASSIFICATION_HELPERS } from './yomitan-dictionary-classification-script';
+import { YOMITAN_FREQUENCY_HELPERS } from './yomitan-frequency-script';
+import { YOMITAN_FURIGANA_HELPERS } from './yomitan-furigana-script';
+import { YOMITAN_KANA_HELPERS } from './yomitan-kana-script';
+import { YOMITAN_MATCH_SELECTION_HELPERS } from './yomitan-match-selection-script';
+
+export { CHARACTER_DICTIONARY_TITLE_PREFIX } from './character-dictionary-title';
+
+export const YOMITAN_SCANNING_HELPERS = [
+  YOMITAN_KANA_HELPERS,
+  YOMITAN_FURIGANA_HELPERS,
+  YOMITAN_FREQUENCY_HELPERS,
+  YOMITAN_DICTIONARY_CLASSIFICATION_HELPERS,
+  YOMITAN_MATCH_SELECTION_HELPERS,
+].join('\n');
@@ -0,0 +1,54 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { HAN_CODE_POINT_RANGES, HAN_REGEXP_CLASS_BODY, isHanCodePoint } from './han-code-points';
+
+test('every range boundary is inside the table', () => {
+  for (const [start, end] of HAN_CODE_POINT_RANGES) {
+    for (const codePoint of [start, end]) {
+      assert.ok(isHanCodePoint(codePoint), `expected U+${codePoint.toString(16)} to be Han`);
+    }
+  }
+
+  // Extension J (Unicode 17) and the Compatibility blocks are the ones a
+  // BMP-only table used to miss.
+  assert.ok(isHanCodePoint(0x323b0));
+  assert.ok(isHanCodePoint(0x33479));
+  assert.ok(isHanCodePoint(0xf900));
+  assert.ok(isHanCodePoint(0x2f800));
+});
+
+test('no unified ideograph the runtime knows about falls outside the table', () => {
+  // One direction only: a runtime with older Unicode data simply checks fewer
+  // code points, where asserting the reverse would fail on Extension J.
+  const unifiedIdeograph = /\p{Unified_Ideograph}/u;
+
+  for (let codePoint = 0x3000; codePoint <= 0x40000; codePoint += 1) {
+    if (unifiedIdeograph.test(String.fromCodePoint(codePoint))) {
+      assert.ok(
+        isHanCodePoint(codePoint),
+        `expected unified ideograph U+${codePoint.toString(16)} to be in the table`,
+      );
+    }
+  }
+});
+
+test('code points just outside the table are rejected', () => {
+  for (const codePoint of [0x33ff, 0x4dc0, 0xa000, 0x1f000, 0x3347a]) {
+    assert.equal(
+      isHanCodePoint(codePoint),
+      false,
+      `expected U+${codePoint.toString(16)} not to be Han`,
+    );
+  }
+});
+
+test('the regexp class body matches the same code points as the predicate', () => {
+  const classRegExp = new RegExp(`^[${HAN_REGEXP_CLASS_BODY}]$`, 'u');
+
+  for (const codePoint of [0x3400, 0x4e00, 0x9fff, 0xf900, 0x20000, 0x323b0, 0x33479]) {
+    assert.match(String.fromCodePoint(codePoint), classRegExp);
+  }
+  for (const codePoint of [0x3040, 0x30ff, 0x33fa, 0x3347a]) {
+    assert.doesNotMatch(String.fromCodePoint(codePoint), classRegExp);
+  }
+});
@@ -0,0 +1,31 @@
+// Single source of truth for "this code point is a Han character", shared by
+// the main-process character dictionary and the in-page Yomitan scan runtime.
+// The two used to carry separate range lists, and they drifted: a name written
+// with a supplementary-plane kanji could enter the generated dictionary while
+// the scanner's greedy name pre-pass refused to probe the position.
+//
+// Ranges rather than \p{Script=Han}: the scan walk tests one code point per
+// character of every subtitle line, where an integer compare beats building a
+// string for a regex, and the script is injected as text into a page where a
+// shared helper cannot be imported.
+export const HAN_CODE_POINT_RANGES: ReadonlyArray<readonly [number, number]> = [
+  [0x3400, 0x4dbf], // Extension A
+  [0x4e00, 0x9fff], // CJK Unified Ideographs
+  [0xf900, 0xfaff], // Compatibility Ideographs
+  [0x20000, 0x2a6df], // Extension B
+  [0x2a700, 0x2ebef], // Extensions C-F
+  [0x2ebf0, 0x2ee5f], // Extension I
+  [0x2f800, 0x2fa1f], // Compatibility Ideographs Supplement
+  [0x30000, 0x3134f], // Extension G
+  [0x31350, 0x323af], // Extension H
+  [0x323b0, 0x33479], // Extension J (Unicode 17)
+];
+
+export function isHanCodePoint(codePoint: number): boolean {
+  return HAN_CODE_POINT_RANGES.some(([start, end]) => codePoint >= start && codePoint <= end);
+}
+
+/** The same ranges as a regular expression character class body (needs the `u` flag). */
+export const HAN_REGEXP_CLASS_BODY = HAN_CODE_POINT_RANGES.map(
+  ([start, end]) => `\\u{${start.toString(16)}}-\\u{${end.toString(16)}}`,
+).join('');
@@ -0,0 +1,203 @@
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import path from 'node:path';
+import test from 'node:test';
+
+import { parseChangelog, resolveChangelogGroupKey } from './changelog-parse';
+
+const SAMPLE = `# Changelog
+
+## v0.19.2 (2026-08-04)
+
+### Changed
+- Subsync: picks both tracks now.
+
+### Fixed
+- Overlay: shows the plain line immediately.
+
+<details>
+<summary>Internal changes</summary>
+
+### Internal
+- Patched \`undici\`.
+
+</details>
+
+## v0.19.1 (2026-08-01)
+
+### Added
+- **Word Card Type:**
+  - Adds a setting.
+  - Flags clear each other.
+
+## v0.18.0 (2026-07-01)
+
+### Fixed
+- Something older.
+`;
+
+test('changelog parser reads versions, dates, and sections in file order', () => {
+  const entries = parseChangelog(SAMPLE);
+
+  assert.deepEqual(
+    entries.map((entry) => `${entry.version}@${entry.date}`),
+    ['0.19.2@2026-08-04', '0.19.1@2026-08-01', '0.18.0@2026-07-01'],
+  );
+  assert.deepEqual(
+    entries[0]?.sections.map((section) => section.heading),
+    ['Changed', 'Fixed', 'Internal'],
+  );
+  assert.deepEqual(entries[0]?.sections[1]?.items, [
+    { text: 'Overlay: shows the plain line immediately.', children: [] },
+  ]);
+});
+
+test('changelog parser flags sections inside the details block as internal', () => {
+  const entries = parseChangelog(SAMPLE);
+  const sections = entries[0]?.sections ?? [];
+
+  assert.deepEqual(
+    sections.map((section) => section.internal),
+    [false, false, true],
+  );
+  assert.deepEqual(sections[2]?.items, [{ text: 'Patched `undici`.', children: [] }]);
+});
+
+test('changelog parser groups entries by major.minor', () => {
+  const entries = parseChangelog(SAMPLE);
+
+  assert.deepEqual(
+    entries.map((entry) => entry.groupKey),
+    ['0.19', '0.19', '0.18'],
+  );
+  assert.equal(resolveChangelogGroupKey('1.2.3'), '1.2');
+});
+
+test('changelog parser keeps bullets that precede any section heading', () => {
+  const entries = parseChangelog('## v0.1.0 (2025-01-01)\n\n- Initial release.\n');
+
+  assert.deepEqual(entries[0]?.sections, [
+    {
+      heading: 'Changes',
+      items: [{ text: 'Initial release.', children: [] }],
+      internal: false,
+    },
+  ]);
+});
+
+test('changelog parser drops empty sections and tolerates missing dates', () => {
+  const entries = parseChangelog('## v0.2.0\n\n### Added\n\n### Fixed\n- One fix.\n');
+
+  assert.equal(entries[0]?.date, '');
+  assert.deepEqual(
+    entries[0]?.sections.map((section) => section.heading),
+    ['Fixed'],
+  );
+});
+
+test('changelog parser keeps indented sub-bullets nested under their lead bullet', () => {
+  const entries = parseChangelog(SAMPLE);
+  const added = entries[1]?.sections.find((section) => section.heading === 'Added');
+
+  assert.deepEqual(added?.items, [
+    {
+      text: '**Word Card Type:**',
+      children: [
+        { text: 'Adds a setting.', children: [] },
+        { text: 'Flags clear each other.', children: [] },
+      ],
+    },
+  ]);
+});
+
+test('changelog parser nests three bullet levels and rejoins wrapped lines', () => {
+  const entries = parseChangelog(
+    [
+      '## v0.9.0 (2025-05-05)',
+      '',
+      '### Added',
+      '- Top level',
+      '  - Second level',
+      '    - Third level',
+      '      continued on the next line',
+      '  - Back to second level',
+      '- Another top level',
+      '',
+    ].join('\n'),
+  );
+
+  assert.deepEqual(entries[0]?.sections[0]?.items, [
+    {
+      text: 'Top level',
+      children: [
+        {
+          text: 'Second level',
+          children: [{ text: 'Third level continued on the next line', children: [] }],
+        },
+        { text: 'Back to second level', children: [] },
+      ],
+    },
+    { text: 'Another top level', children: [] },
+  ]);
+});
+
+test('changelog parser reads prerelease and build metadata version headings', () => {
+  const entries = parseChangelog(
+    [
+      '## v0.16.0 (2026-06-01)',
+      '',
+      '### Added',
+      '- New in 0.16.',
+      '',
+      '## v0.15.0-rc.1+build.2 (2026-05-29)',
+      '',
+      '### Added',
+      '- Release candidate note.',
+      '',
+    ].join('\n'),
+  );
+
+  // The prerelease heading has to become its own entry. Asserting the exact
+  // version list is what catches the failure mode: a heading the regex misses
+  // is not skipped, its notes silently fold into the release above it.
+  assert.deepEqual(
+    entries.map((entry) => entry.version),
+    ['0.16.0', '0.15.0-rc.1+build.2'],
+  );
+  assert.equal(entries[1]?.date, '2026-05-29');
+  assert.equal(entries[1]?.groupKey, '0.15');
+  assert.equal(entries[0]?.sections.length, 1);
+  // The prerelease body has to land on its own entry, not fold into 0.16.0.
+  assert.deepEqual(entries[1]?.sections, [
+    {
+      heading: 'Added',
+      items: [{ text: 'Release candidate note.', children: [] }],
+      internal: false,
+    },
+  ]);
+});
+
+test('changelog parser handles the repo CHANGELOG.md', () => {
+  const markdown = fs.readFileSync(path.join(process.cwd(), 'CHANGELOG.md'), 'utf8');
+  const entries = parseChangelog(markdown);
+
+  assert.ok(entries.length > 3);
+  for (const entry of entries) {
+    assert.match(entry.version, /^\d+\.\d+\.\d+/);
+    assert.ok(entry.sections.length > 0, `expected sections for v${entry.version}`);
+    for (const section of entry.sections) {
+      for (const item of section.items) {
+        assert.ok(item.text.length > 0, `empty bullet in v${entry.version}`);
+      }
+    }
+  }
+
+  // Older entries group notes under a bold lead bullet; nesting must survive.
+  const breaking = entries
+    .find((entry) => entry.version === '0.15.0')
+    ?.sections.find((section) => section.heading === 'Breaking Changes');
+  assert.deepEqual(
+    breaking?.items.map((item) => `${item.text}:${item.children.length}`),
+    ['**Subsync:**:2', '**N+1 Highlighting:**:2'],
+  );
+});
@@ -0,0 +1,128 @@
+import type { ChangelogEntry, ChangelogItem, ChangelogSection } from '../../types/changelog';
+
+// Prerelease and build metadata are matched separately: a single `[-+]`-led
+// group cannot span `-rc.1+build.2`, and an unmatched heading silently folds
+// that release's notes into the previous entry.
+const VERSION_HEADING =
+  /^##\s+v(\d+\.\d+\.\d+(?:-[0-9A-Za-z.-]+)?(?:\+[0-9A-Za-z.-]+)?)\s*(?:\(([^)]*)\))?\s*$/;
+const SECTION_HEADING = /^###\s+(.+?)\s*$/;
+const BULLET = /^(\s*)[-*]\s+(.*)$/;
+
+/**
+ * Entries are grouped by `major.minor` so the whole current minor line renders
+ * expanded, matching how docs-site/changelog.md splits current vs previous.
+ */
+export function resolveChangelogGroupKey(version: string): string {
+  const match = version.match(/^(\d+)\.(\d+)/);
+  if (!match) return version;
+  return `${match[1]}.${match[2]}`;
+}
+
+/**
+ * Parses the repo CHANGELOG.md into version entries. Bullets keep their inline
+ * markdown and their nesting: older entries group related notes under a bold
+ * lead bullet with indented children, and flattening them loses that structure.
+ */
+export function parseChangelog(markdown: string): ChangelogEntry[] {
+  const entries: ChangelogEntry[] = [];
+  let entry: ChangelogEntry | null = null;
+  let section: ChangelogSection | null = null;
+  let internal = false;
+  // Open bullets from outermost to innermost, used to place the next bullet.
+  let openItems: Array<{ indent: number; item: ChangelogItem }> = [];
+
+  function startSection(heading: string): void {
+    section = { heading, items: [], internal };
+    openItems = [];
+    entry?.sections.push(section);
+  }
+
+  function addBullet(indent: number, text: string): void {
+    if (!section) {
+      // Bullets before any "###" heading (older entries) land in a generic group.
+      startSection('Changes');
+    }
+    const item: ChangelogItem = { text, children: [] };
+
+    while (openItems.length > 0 && (openItems[openItems.length - 1]?.indent ?? 0) >= indent) {
+      openItems.pop();
+    }
+    const parent = openItems[openItems.length - 1];
+    if (parent) {
+      parent.item.children.push(item);
+    } else {
+      section?.items.push(item);
+    }
+    openItems.push({ indent, item });
+  }
+
+  function appendContinuation(text: string): void {
+    const current = openItems[openItems.length - 1];
+    if (!current) return;
+    current.item.text = `${current.item.text} ${text}`;
+  }
+
+  for (const rawLine of markdown.split(/\r?\n/)) {
+    const line = rawLine.trimEnd();
+    const trimmed = line.trim();
+
+    const versionMatch = trimmed.match(VERSION_HEADING);
+    if (versionMatch) {
+      const version = versionMatch[1] ?? '';
+      entry = {
+        version,
+        date: versionMatch[2]?.trim() ?? '',
+        groupKey: resolveChangelogGroupKey(version),
+        sections: [],
+      };
+      entries.push(entry);
+      section = null;
+      internal = false;
+      openItems = [];
+      continue;
+    }
+
+    if (!entry) continue;
+
+    if (trimmed.startsWith('<details')) {
+      internal = true;
+      section = null;
+      openItems = [];
+      continue;
+    }
+    if (trimmed.startsWith('</details')) {
+      internal = false;
+      section = null;
+      openItems = [];
+      continue;
+    }
+    if (trimmed.startsWith('<summary')) continue;
+
+    const sectionMatch = trimmed.match(SECTION_HEADING);
+    if (sectionMatch) {
+      startSection(sectionMatch[1] ?? '');
+      continue;
+    }
+
+    const bulletMatch = line.match(BULLET);
+    if (bulletMatch) {
+      addBullet((bulletMatch[1] ?? '').length, bulletMatch[2] ?? '');
+      continue;
+    }
+
+    // An indented non-bullet line continues the bullet above it, including
+    // across a blank line: that is CommonMark's continuation paragraph, and
+    // dropping the open bullets here would silently discard the text.
+    if (!trimmed) {
+      continue;
+    }
+    if (/^\s/.test(line)) {
+      appendContinuation(trimmed);
+    }
+  }
+
+  return entries.map((item) => ({
+    ...item,
+    sections: item.sections.filter((entrySection) => entrySection.items.length > 0),
+  }));
+}
@@ -1,6 +1,6 @@
 import assert from 'node:assert/strict';
 import test from 'node:test';
-import { shouldForceX11ElectronBackend } from './electron-backend';
+import { resolveX11ElectronRelaunchArgs, shouldForceX11ElectronBackend } from './electron-backend';

 function withPlatform(platform: NodeJS.Platform, run: () => void): void {
  const original = Object.getOwnPropertyDescriptor(process, 'platform');
@@ -32,3 +32,70 @@ test('shouldForceX11ElectronBackend is false off Linux', () => {
    assert.equal(shouldForceX11ElectronBackend({}), false);
  });
 });
+
+test('resolveX11ElectronRelaunchArgs adds the raw X11 Ozone argument on unsupported Linux', () => {
+  assert.deepEqual(
+    resolveX11ElectronRelaunchArgs(
+      ['--start'],
+      {
+        DISPLAY: ':1',
+        WAYLAND_DISPLAY: 'wayland-0',
+        XDG_CURRENT_DESKTOP: 'KDE',
+      },
+      'linux',
+    ),
+    ['--start', '--ozone-platform=x11'],
+  );
+});
+
+test('resolveX11ElectronRelaunchArgs avoids loops and preserves native Wayland backends', () => {
+  const kdeWayland = {
+    DISPLAY: ':1',
+    WAYLAND_DISPLAY: 'wayland-0',
+    XDG_CURRENT_DESKTOP: 'KDE',
+  };
+  assert.equal(
+    resolveX11ElectronRelaunchArgs(['--start', '--ozone-platform=x11'], kdeWayland, 'linux'),
+    null,
+  );
+  assert.equal(
+    resolveX11ElectronRelaunchArgs(
+      ['--start'],
+      { ...kdeWayland, HYPRLAND_INSTANCE_SIGNATURE: 'hypr' },
+      'linux',
+    ),
+    null,
+  );
+  assert.equal(resolveX11ElectronRelaunchArgs(['--start'], kdeWayland, 'darwin'), null);
+  assert.equal(
+    resolveX11ElectronRelaunchArgs(
+      [],
+      {
+        ...kdeWayland,
+        SUBMINER_APP_ARGC: '1',
+        SUBMINER_APP_ARG_0: '--start',
+      },
+      'linux',
+    )?.at(-1),
+    '--ozone-platform=x11',
+  );
+  assert.equal(
+    resolveX11ElectronRelaunchArgs([], { ...kdeWayland, SUBMINER_X11_BOOTSTRAPPED: '1' }, 'linux'),
+    null,
+  );
+});
+
+test('resolveX11ElectronRelaunchArgs replaces an explicit unsupported Wayland argument', () => {
+  assert.deepEqual(
+    resolveX11ElectronRelaunchArgs(
+      ['--start', '--ozone-platform', 'wayland'],
+      {
+        DISPLAY: ':1',
+        WAYLAND_DISPLAY: 'wayland-0',
+        XDG_CURRENT_DESKTOP: 'KDE',
+      },
+      'linux',
+    ),
+    ['--start', '--ozone-platform=x11'],
+  );
+});
@@ -4,6 +4,9 @@ import { isSupportedWaylandCompositor } from '../../shared/mpv-x11-backend';

 const logger = createLogger('core:electron-backend');

+export const X11_ELECTRON_BOOTSTRAP_ENV = 'SUBMINER_X11_BOOTSTRAPPED';
+const X11_ELECTRON_OZONE_ARG = '--ozone-platform=x11';
+
 function getElectronOzonePlatformHint(env: NodeJS.ProcessEnv = process.env): string | null {
  const hint = env.ELECTRON_OZONE_PLATFORM_HINT?.trim().toLowerCase();
  if (hint) return hint;
@@ -24,11 +27,42 @@ function getElectronOzonePlatformHint(env: NodeJS.ProcessEnv = process.env): str
 * Electron Wayland backend is unsupported); the Hyprland/Sway case is left untouched so
 * {@link enforceUnsupportedWaylandMode} can report it.
 */
-export function shouldForceX11ElectronBackend(env: NodeJS.ProcessEnv = process.env): boolean {
-  if (process.platform !== 'linux') return false;
+export function shouldForceX11ElectronBackend(
+  env: NodeJS.ProcessEnv = process.env,
+  platform: NodeJS.Platform = process.platform,
+): boolean {
+  if (platform !== 'linux') return false;
  return !isSupportedWaylandCompositor(env);
 }

+export function resolveX11ElectronRelaunchArgs(
+  args: string[],
+  env: NodeJS.ProcessEnv = process.env,
+  platform: NodeJS.Platform = process.platform,
+): string[] | null {
+  if (!shouldForceX11ElectronBackend(env, platform)) return null;
+  if (env[X11_ELECTRON_BOOTSTRAP_ENV] === '1') return null;
+
+  const retainedArgs: string[] = [];
+  let alreadyForced = false;
+  for (let index = 0; index < args.length; index += 1) {
+    const arg = args[index];
+    if (arg === '--ozone-platform') {
+      const value = args[index + 1];
+      alreadyForced = value?.trim().toLowerCase() === 'x11';
+      if (value && !value.startsWith('--')) index += 1;
+      continue;
+    }
+    if (arg?.startsWith('--ozone-platform=')) {
+      alreadyForced = arg.slice('--ozone-platform='.length).trim().toLowerCase() === 'x11';
+      continue;
+    }
+    if (arg) retainedArgs.push(arg);
+  }
+
+  return alreadyForced ? null : [...retainedArgs, X11_ELECTRON_OZONE_ARG];
+}
+
 export function forceX11Backend(args: CliArgs): void {
  if (!shouldStartApp(args)) return;
  if (!shouldForceX11ElectronBackend()) return;
@@ -52,9 +52,15 @@ function resolveRuntimeDefaultNotificationIconPath(): string | null {
  });
 }

+/**
+ * Live notifications keyed by `replaceId`. Electron exposes no native "replace this notification"
+ * flag, so a repeated status closes its predecessor instead of stacking a fresh toast per update.
+ */
+const notificationsByReplaceId = new Map<string, Electron.Notification>();
+
 export function showDesktopNotification(
  title: string,
-  options: { body?: string; icon?: string },
+  options: { body?: string; icon?: string; replaceId?: string },
 ): void {
  const notificationOptions: {
    title: string;
@@ -98,5 +104,15 @@ export function showDesktopNotification(
  }

  const notification = new Notification(notificationOptions);
+  const replaceId = options.replaceId?.trim();
+  if (replaceId) {
+    notificationsByReplaceId.get(replaceId)?.close();
+    notificationsByReplaceId.set(replaceId, notification);
+    notification.once('close', () => {
+      if (notificationsByReplaceId.get(replaceId) === notification) {
+        notificationsByReplaceId.delete(replaceId);
+      }
+    });
+  }
  notification.show();
 }
@@ -0,0 +1,59 @@
+/**
+ * Loose semver ordering shared by the updater and the changelog UI.
+ * Returns >0 when `a` is newer, <0 when older, 0 when equal.
+ */
+export function compareSemverLike(a: string, b: string): number {
+  const parse = (
+    value: string,
+  ): {
+    core: number[];
+    prerelease: Array<number | string>;
+  } => {
+    // Build metadata ("+build.2") is not part of precedence per semver, and
+    // leaving it attached makes it leak into the prerelease comparison.
+    const normalized = value.replace(/^v/i, '').split('+', 1)[0] ?? '';
+    const [coreText = '', ...prereleaseParts] = normalized.split('-');
+    const core = coreText
+      .split('.')
+      .slice(0, 3)
+      .map((part) => Number.parseInt(part, 10) || 0);
+    while (core.length < 3) core.push(0);
+    const prereleaseText = prereleaseParts.join('-');
+    return {
+      core,
+      prerelease: prereleaseText
+        ? prereleaseText.split('.').map((part) => {
+            const numeric = Number.parseInt(part, 10);
+            return /^\d+$/.test(part) ? numeric : part;
+          })
+        : [],
+    };
+  };
+  const left = parse(a);
+  const right = parse(b);
+  for (let i = 0; i < 3; i += 1) {
+    const diff = (left.core[i] ?? 0) - (right.core[i] ?? 0);
+    if (diff !== 0) return diff;
+  }
+
+  if (left.prerelease.length === 0 && right.prerelease.length === 0) return 0;
+  if (left.prerelease.length === 0) return 1;
+  if (right.prerelease.length === 0) return -1;
+
+  const length = Math.max(left.prerelease.length, right.prerelease.length);
+  for (let i = 0; i < length; i += 1) {
+    const leftPart = left.prerelease[i];
+    const rightPart = right.prerelease[i];
+    if (leftPart === undefined && rightPart === undefined) return 0;
+    if (leftPart === undefined) return -1;
+    if (rightPart === undefined) return 1;
+    if (leftPart === rightPart) continue;
+    if (typeof leftPart === 'number' && typeof rightPart === 'number') {
+      return leftPart - rightPart;
+    }
+    if (typeof leftPart === 'number') return -1;
+    if (typeof rightPart === 'number') return 1;
+    return leftPart > rightPart ? 1 : -1;
+  }
+  return 0;
+}
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
sudacode	b98d4d65c7	fix(stats): stop counting duplicate typeset subtitle lines (#191 )	2026-08-14 00:12:56 -07:00
sudacode	046e74ea91	chore(workflow): simplify agent tooling and docs	2026-08-13 23:34:04 -07:00
sudacode	8bf847503d	fix(media): tolerate slow MKV audio extraction (#195 )	2026-08-13 23:19:22 -07:00
sudacode	47b5903392	chore(release): prepare v0.19.3	2026-08-13 23:02:57 -07:00
sudacode	bf85554d1e	fix(stats): batch deletes off the main thread (#194 )	2026-08-13 22:31:52 -07:00
sudacode	d74c7e1235	chore: add Claude instructions symlink and update js-yaml	2026-08-11 22:31:26 -07:00
sudacode	57ddd19953	fix(overlay): handle X11 display scaling across monitors (#193 )	2026-08-11 22:24:01 -07:00
sudacode	ee25536d90	fix(dictionary): stop large character dictionaries from timing out (#189 )	2026-08-11 18:40:18 -07:00
sudacode	7b0fbdf254	fix(subtitles): collapse duplicate ASS events and decode text once (#186 )	2026-08-10 22:21:44 -07:00
sudacode	2fefc83e3f	fix(playback): stop forcing legacy OpenGL renderer on X11 mpv backend (#188 )	2026-08-06 23:52:35 -07:00
sudacode	dbdf578c68	perf(tokenizer): single-pass Yomitan scan with cross-line caching and prefetch fixes (#185 )	2026-08-06 21:44:09 -07:00
sudacode	441ecf3c04	feat(overlay): add in-app changelog modal (#187 )	2026-08-05 22:19:13 -07:00