test(stats): click backdrop, not disabled button, in cleanup close test

- Assert the Close button is disabled while apply is in flight
- Click the always-enabled backdrop instead, since that's the path that actually reaches the close guard

changes/duplicate-line-stats-cleanup.md

@@ -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: when the active subtitle file has been parsed, stats record exactly the cues the sidebar shows, and for sources with no parsed cue list a run of identical, contiguous, sub-0.1s lines stops counting after a few frames. Ordinary repeated dialogue and rewatches are unaffected.
+- Added a cleanup for stats already affected. The Vocabulary tab has a **Duplicates** button that scans a chosen window (7 days through all time), shows the bursts it found and the word and kanji counts they added, and collapses each run to one line once confirmed. `subminer stats cleanup --duplicate-lines` does the same from the terminal, with `--dry-run` and `--lookback-days <n>`. Only subtitle lines and the vocabulary counts they feed are touched; watch time and lines-seen totals are left as recorded.

changes/styled-subtitle-cue-parsing.md

@@ -0,0 +1,5 @@
+type: fixed
+area: subtitles
+
+- Heavily typeset ASS scripts (karaoke OP/ED, sign work) no longer fill the subtitle sidebar with garbage. Vector drawing runs (`\p1` … `\p0`) are no longer shown as subtitle text (e.g. `m 20 0 b 10 0 0 10 0 20 …`), and duplicate events in a parsed subtitle file collapse into one cue: identical text over an identical span (layered "shadow" copies), and per-frame animation bursts. An ASS burst has to prove itself with authoring evidence — a temporal tag (`\t`, `\move`, karaoke timing), an animated `Effect` column, or override values that change from event to event — plus one shared style and actor, so three rapid `えっ` reactions from three characters, or a sign repeated with the same static `\clip`, stay separate. SRT and VTT carry no such metadata, so there the run has to be at least five contiguous events all shorter than 0.1s, which is where ASS-to-SRT conversion leaves karaoke frames.
+- Subtitle text is now decoded from ASS exactly once, where it enters the app, and matches how mpv renders the same line (including `\N`, `\n`, `\h`, unclosed `{`, and the fact that `\{` is not an escape). Renderer, timing tracker, tokenizer and the tokenization cache take that decoded text as-is instead of each re-deriving it, so one authored line can no longer produce two different cache keys, and a cue that normalizes to nothing is no longer stored as subtitle text or cached under an empty key.

changes/x11-backend-keeps-mpv-renderer.md

@@ -0,0 +1,4 @@
+type: fixed
+area: playback
+
+- XWayland/X11 mode (`--backend=x11`, or the automatic fallback on non-Hyprland/Sway Wayland sessions) no longer forces mpv onto `--vo=gpu --gpu-api=opengl`. It now only pins the window context (`--gpu-context=x11vk,x11egl,x11`), so a `vo=gpu-next` config keeps its renderer, API, and user shaders. Forcing the legacy OpenGL renderer crashed mpv on the first fullscreen toggle for anyone using a gpu-next user shader that emits a 4-component LUMA hook (ArtCNN and friends), which asserts in mpv's old renderer (`copy_image: *offset + count < sizeof(dst)`) as soon as the shader's upscale-only condition turns on.

docs-site/immersion-tracking.md

@@ -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 +125,32 @@ 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.
+- Otherwise only timing is available, so the strict metadata-free rule applies: a run of identical, contiguous lines each shorter than 0.1s stops being recorded after a few frames. 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 `--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.
+
+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:

docs-site/launcher-script.md

@@ -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                                                                 |

docs-site/troubleshooting.md

@@ -406,7 +406,7 @@ 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.
+- 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.
 - 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 +420,7 @@ 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).
 

docs-site/usage.md

@@ -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

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

@@ -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)

launcher/commands/stats-command.ts

@@ -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);
     }

launcher/config/args-normalizer.test.ts

@@ -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: [],

launcher/config/args-normalizer.ts

@@ -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 (

launcher/config/cli-parser-builder.ts

@@ -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; anything unusable means no limit. */
+function parseStatsLookbackDays(value: unknown): number | null {
+  if (typeof value !== 'string' && typeof value !== 'number') return null;
+  const days = Number(value);
+  if (!Number.isFinite(days) || days <= 0) {
+    throw new Error('Stats --lookback-days must be a positive number of days.');
+  }
+  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 +308,29 @@ 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)) {
+        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;
+        if (statsCleanupLifetime && statsCleanupDuplicateLines) {
+          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 +518,9 @@ export function parseCliPrograms(
       statsCleanup,
       statsCleanupVocab,
       statsCleanupLifetime,
+      statsCleanupDuplicateLines,
+      statsCleanupDryRun,
+      statsCleanupLookbackDays,
       statsLogLevel,
       syncTriggered,
       syncCliTokens,

launcher/mpv.test.ts

@@ -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',
       ],
     );

launcher/parse-args.test.ts

@@ -232,6 +232,53 @@ 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);
+});
+
+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 combining lifetime and duplicate-line cleanup modes', () => {
+  const error = withProcessExitIntercept(() => {
+    parseArgs(['stats', 'cleanup', '--lifetime', '--duplicate-lines'], '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', '-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 a positive number of days/);
+  }
+});
+
 test('parseArgs rejects cleanup-only stats flags without cleanup action', () => {
   const error = withProcessExitIntercept(() => {
     parseArgs(['stats', '--vocab'], 'subminer', {});
@@ -239,7 +286,10 @@ 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', () => {

launcher/types.ts

@@ -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;

src/cli/args.ts

@@ -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;
@@ -167,6 +170,8 @@ export function parseArgs(argv: string[]): CliArgs {
     statsCleanup: false,
     statsCleanupVocab: false,
     statsCleanupLifetime: false,
+    statsCleanupDuplicateLines: false,
+    statsCleanupDryRun: false,
     jellyfin: false,
     jellyfinLogin: false,
     jellyfinLogout: false,
@@ -368,7 +373,15 @@ 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=')) {
+      const value = Number(arg.split('=', 2)[1]);
+      if (Number.isFinite(value) && value > 0) args.statsCleanupLookbackDays = Math.floor(value);
+    } else if (arg === '--stats-cleanup-lookback-days') {
+      const value = Number(readValue(argv[i + 1]));
+      if (Number.isFinite(value) && value > 0) args.statsCleanupLookbackDays = Math.floor(value);
+    } else if (arg.startsWith('--stats-response-path=')) {
       const value = arg.split('=', 2)[1];
       if (value) args.statsResponsePath = value;
     } else if (arg === '--stats-response-path') {

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

@@ -1032,6 +1032,95 @@ 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 ignores a window shorter than a day', async () => {
+    let seenOptions: unknown = null;
+    const app = createStatsApp(
+      createMockTracker({
+        cleanupDuplicateSubtitleLines: async (options: unknown) => {
+          seenOptions = options;
+          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, 200);
+    // Half a day must not floor to a zero-day window; it means no limit.
+    assert.deepEqual(seenOptions, { dryRun: true, lookbackDays: null });
+  });
+
+  it('POST /api/stats/maintenance/duplicate-lines treats a missing body as an apply over 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' });
+
+    assert.equal(res.status, 200);
+    assert.deepEqual(seenOptions, { dryRun: false, lookbackDays: null });
+  });
+
   it('PUT /api/stats/excluded-words rejects malformed rows', async () => {
     const app = createStatsApp(createMockTracker());
 

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

@@ -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);
+});

src/core/services/ass-text.ts

@@ -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);
+}

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

@@ -91,6 +91,11 @@ import {
   markVideoWatched,
   upsertCoverArt,
 } from './immersion-tracker/query-maintenance';
+import {
+  cleanupDuplicateSubtitleLines,
+  type DuplicateSubtitleLineCleanupOptions,
+  type DuplicateSubtitleLineCleanupSummary,
+} from './immersion-tracker/duplicate-line-cleanup';
 import { repairJellyfinStreamVideoLinks } from './immersion-tracker/jellyfin-link-repair';
 import {
   repairLegacySeasonlessAnimeRows,
@@ -595,6 +600,18 @@ 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();
@@ -1799,6 +1816,24 @@ 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);
   }

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

@@ -0,0 +1,312 @@
+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 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('a short run below the threshold survives', () => {
+  const { db, dbPath } = createDb(karaokeFrames(1, '飛び上がる', 1_000, 4, 40));
+
+  try {
+    const summary = cleanupDuplicateSubtitleLines(db);
+
+    assert.equal(summary.burstGroups, 0);
+    assert.equal(countLines(db), 4);
+  } 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);
+  }
+});

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

@@ -0,0 +1,373 @@
+/*
+ * 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
+ * each line may be as long as the animation-frame bound rather than the much tighter
+ * timing-only one. 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_TIMING_ONLY_FRAMES,
+} 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 LINE_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;
+  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' && 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),
+    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 {
+  if (run.length < bounds.minRunLength) {
+    return false;
+  }
+  return run.every((row) => row.endMs - row.startMs <= bounds.maxFrameMs);
+}
+
+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.
+ *
+ * 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);
+  const bursts: DuplicateSubtitleLineBurst[] = [];
+  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 rows) {
+    const previous = run[run.length - 1];
+    const continuesRun =
+      previous !== undefined &&
+      previous.sessionId === row.sessionId &&
+      previous.videoId === row.videoId &&
+      previous.text === row.text &&
+      row.startMs <= chainEndMs + bounds.gapToleranceMs;
+
+    if (continuesRun) {
+      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, LINE_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, LINE_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, LINE_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) {
+    db.exec('ROLLBACK');
+    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;
+}

src/core/services/immersion-tracker/query-shared.ts

@@ -268,6 +268,19 @@ 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,

src/core/services/stats-server/library-routes.ts

@@ -5,6 +5,7 @@ import {
   buildSentenceSearchOptions,
   enrichSessionsWithKnownWordMetrics,
   parseBooleanQuery,
+  parseDuplicateLineCleanupBody,
   parseExcludedWordsBody,
   parseIntQuery,
 } from './route-support.js';
@@ -40,6 +41,15 @@ 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 body = await c.req.json().catch(() => null);
+    const { dryRun, lookbackDays } = parseDuplicateLineCleanupBody(body);
+    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();

src/core/services/stats-server/route-support.ts

@@ -88,6 +88,25 @@ export function parseExcludedWordsBody(body: unknown): StatsExcludedWord[] | nul
   return words;
 }
 
+/**
+ * Read a duplicate-line cleanup request. An absent or unusable `lookbackDays` scans all
+ * history, which is what the CLI does; only a positive number narrows the window.
+ */
+export function parseDuplicateLineCleanupBody(body: unknown): {
+  dryRun: boolean;
+  lookbackDays: number | null;
+} {
+  const source = body && typeof body === 'object' ? (body as Record<string, unknown>) : {};
+  const rawLookback = source.lookbackDays;
+  // Floor before the bounds check, or a fraction of a day arrives as a zero-day window.
+  const wholeDays =
+    typeof rawLookback === 'number' && Number.isFinite(rawLookback)
+      ? Math.floor(rawLookback)
+      : null;
+  const lookbackDays = wholeDays !== null && wholeDays >= 1 ? wholeDays : null;
+  return { dryRun: source.dryRun === true, lookbackDays };
+}
+
 export function loadKnownWordsSet(cachePath: string | undefined): Set<string> | null {
   if (!cachePath || !existsSync(cachePath)) return null;
   try {

src/core/services/subtitle-burst-constants.ts

@@ -0,0 +1,40 @@
+/*
+ * 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;

src/core/services/subtitle-cue-dedup.ts

@@ -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);
+}

src/core/services/subtitle-cue-parser.test.ts

@@ -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]',

src/core/services/subtitle-cue-parser.ts

@@ -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));
 }

src/core/services/subtitle-line-dedup-gate.test.ts

@@ -0,0 +1,116 @@
+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, 0.1).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('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('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 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);
+});

src/core/services/subtitle-line-dedup-gate.ts

@@ -0,0 +1,188 @@
+/*
+ * 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 the streaming run state, e.g. when playback moves to another file. */
+  reset: () => void;
+}
+
+interface CueSpan {
+  startTime: number;
+  endTime: number;
+}
+
+interface StreamingRunState {
+  text: string;
+  startMs: number;
+  chainEndSec: number;
+  /** Contiguous identical short frames seen so far, including the recorded first one. */
+  frames: number;
+}
+
+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 {
+  const startsOwnCue = spans.some(
+    (span) => Math.abs(startSec - span.startTime) <= DUPLICATE_CUE_GAP_TOLERANCE_SECONDS,
+  );
+  if (startsOwnCue) {
+    return false;
+  }
+  return spans.some(
+    (span) =>
+      startSec > span.startTime + DUPLICATE_CUE_GAP_TOLERANCE_SECONDS &&
+      startSec <= span.endTime + DUPLICATE_CUE_GAP_TOLERANCE_SECONDS,
+  );
+}
+
+export function createSubtitleLineDedupGate(
+  deps: SubtitleLineDedupGateDeps,
+): SubtitleLineDedupGate {
+  let indexedCues: readonly SubtitleCue[] | null = null;
+  let spansByText: Map<string, CueSpan[]> = new Map();
+  let run: StreamingRunState | null = null;
+
+  const lookupSpans = (text: string): CueSpan[] | null => {
+    const cues = deps.getParsedCues();
+    if (!cues?.length) {
+      indexedCues = null;
+      spansByText = new Map();
+      return null;
+    }
+    if (cues !== indexedCues) {
+      indexedCues = cues;
+      spansByText = buildSpansByText(cues);
+    }
+    return spansByText.get(text) ?? null;
+  };
+
+  /**
+   * 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 => {
+    const startMs = Math.round(sample.startSec * 1000);
+    // 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.text === text && 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 !== null &&
+      run.text === text &&
+      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;
+    } else {
+      run = {
+        text,
+        startMs,
+        chainEndSec: sample.endSec,
+        frames: isShortFrame ? 1 : 0,
+      };
+    }
+
+    return run.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) {
+        run = null;
+        return !isMergedAwayFrame(spans, sample.startSec);
+      }
+
+      return advanceStreamingRun(text, sample);
+    },
+    reset: () => {
+      run = null;
+    },
+  };
+}

src/core/services/subtitle-processing-controller.test.ts

@@ -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;

src/core/services/subtitle-processing-controller.ts

@@ -1,4 +1,5 @@
 import type { SubtitleData } from '../../types';
+import { normalizePlainSubtitleText } from './ass-text';
 
 export interface SubtitleProcessingControllerDeps {
   tokenizeSubtitle: (text: string) => Promise<SubtitleData | null>;
@@ -47,8 +48,15 @@ export interface SubtitleProcessingController {
   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(
@@ -72,6 +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;
@@ -83,7 +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) {
@@ -254,7 +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);
     },
   };
 }

src/core/services/tokenizer.test.ts

@@ -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 () => {

src/core/services/tokenizer.ts

@@ -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');
 
@@ -886,14 +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

src/main/runtime/mpv-main-event-main-deps.ts

@@ -1,4 +1,5 @@
-import type { MergedToken, SubtitleData } from '../../types';
+import { createSubtitleLineDedupGate } from '../../core/services/subtitle-line-dedup-gate';
+import type { MergedToken, SubtitleCue, SubtitleData } from '../../types';
 
 type AnilistPostWatchRunOptions = {
   watchedSeconds?: number;
@@ -34,6 +35,7 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
     subtitleTimingTracker: {
       recordSubtitle?: (text: string, start: number, end: number, secondaryText?: string) => void;
     } | null;
+    activeParsedSubtitleCues?: SubtitleCue[] | null;
     currentMediaPath?: string | null;
     currentSubText: string;
     currentSubAssText: string;
@@ -86,6 +88,11 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
     deps.ensureImmersionTrackerInitialized();
     deps.appState.immersionTracker?.recordPlaybackPosition?.(normalizedTimeSec);
   };
+  // mpv reports every animation frame of a typeset line as its own subtitle event, so
+  // stats have to collapse bursts the same way the parsed cue list already does.
+  const immersionLineDedupGate = createSubtitleLineDedupGate({
+    getParsedCues: () => deps.appState.activeParsedSubtitleCues,
+  });
   const hasInitialPlaybackQuitOnDisconnectArg = (): boolean =>
     Boolean(
       deps.appState.initialArgs?.managedPlayback ||
@@ -110,6 +117,9 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
       if (!tracker?.recordSubtitleLine) {
         return;
       }
+      if (!immersionLineDedupGate.shouldRecord({ text, startSec: start, endSec: end })) {
+        return;
+      }
       const secondaryText = deps.appState.mpvClient?.currentSecondarySubText || null;
       const cachedTokens =
         deps.appState.currentSubtitleData?.text === text

src/main/runtime/stats-cli-command.test.ts

@@ -200,6 +200,59 @@ test('stats cli command fails when immersion tracking is disabled', async () =>
   ]);
 });
 
+test('stats cli command runs a duplicate-line cleanup preview without touching the dashboard', async () => {
+  const { handler, calls, responses } = makeHandler({
+    getImmersionTracker: () => ({
+      cleanupDuplicateSubtitleLines: async (options: {
+        dryRun?: boolean;
+        lookbackDays?: number | null;
+      }) => ({
+        dryRun: options.dryRun === true,
+        lookbackDays: options.lookbackDays ?? null,
+        scannedLines: 900,
+        burstGroups: 2,
+        removedLines: 180,
+        removedWordOccurrences: 540,
+        removedKanjiOccurrences: 120,
+        samples: [
+          {
+            videoId: 7,
+            videoTitle: 'Ep 1',
+            text: '飛び上がる',
+            frames: 90,
+            removedLines: 89,
+            startMs: 1000,
+            endMs: 5000,
+          },
+        ],
+      }),
+    }),
+  });
+
+  await handler(
+    {
+      statsResponsePath: '/tmp/subminer-stats-response.json',
+      statsCleanup: true,
+      statsCleanupDuplicateLines: true,
+      statsCleanupDryRun: true,
+      statsCleanupLookbackDays: 30,
+    },
+    'initial',
+  );
+
+  assert.deepEqual(calls, [
+    'ensureImmersionTrackerStarted',
+    'info:Stats duplicate-line cleanup preview (last 30d): scanned=900 bursts=2 removedLines=180 removedWordCounts=540 removedKanjiCounts=120',
+    'info:  Ep 1: "飛び上がる" x90',
+  ]);
+  assert.deepEqual(responses, [
+    {
+      responsePath: '/tmp/subminer-stats-response.json',
+      payload: { ok: true },
+    },
+  ]);
+});
+
 test('stats cli command runs vocab cleanup instead of opening dashboard when cleanup mode is requested', async () => {
   const { handler, calls, responses } = makeHandler({
     getImmersionTracker: () => ({

src/main/runtime/stats-cli-command.ts

@@ -1,6 +1,7 @@
 import fs from 'node:fs';
 import path from 'node:path';
 import type { CliArgs, CliCommandSource } from '../../cli/args';
+import type { DuplicateSubtitleLineCleanupSummary } from '../../core/services/immersion-tracker/duplicate-line-cleanup';
 import type {
   LifetimeRebuildSummary,
   VocabularyCleanupSummary,
@@ -50,6 +51,10 @@ export function createRunStatsCliCommandHandler(deps: {
   ensureVocabularyCleanupTokenizerReady?: () => Promise<void> | void;
   getImmersionTracker: () => {
     cleanupVocabularyStats?: () => Promise<VocabularyCleanupSummary>;
+    cleanupDuplicateSubtitleLines?: (options: {
+      dryRun?: boolean;
+      lookbackDays?: number | null;
+    }) => Promise<DuplicateSubtitleLineCleanupSummary>;
     rebuildLifetimeSummaries?: () => Promise<LifetimeRebuildSummary>;
   } | null;
   ensureStatsServerStarted: () => string;
@@ -83,6 +88,9 @@ export function createRunStatsCliCommandHandler(deps: {
       | 'statsCleanup'
       | 'statsCleanupVocab'
       | 'statsCleanupLifetime'
+      | 'statsCleanupDuplicateLines'
+      | 'statsCleanupDryRun'
+      | 'statsCleanupLookbackDays'
     >,
     source: CliCommandSource,
   ): Promise<void> => {
@@ -126,6 +134,7 @@ export function createRunStatsCliCommandHandler(deps: {
         const cleanupModes = [
           args.statsCleanupVocab ? 'vocab' : null,
           args.statsCleanupLifetime ? 'lifetime' : null,
+          args.statsCleanupDuplicateLines ? 'duplicate-lines' : null,
         ].filter(Boolean);
         if (cleanupModes.length !== 1) {
           throw new Error('Choose exactly one stats cleanup mode.');
@@ -142,6 +151,27 @@ export function createRunStatsCliCommandHandler(deps: {
           writeResponseSafe(args.statsResponsePath, { ok: true });
           return;
         }
+        if (args.statsCleanupDuplicateLines && tracker.cleanupDuplicateSubtitleLines) {
+          const result = await tracker.cleanupDuplicateSubtitleLines({
+            dryRun: args.statsCleanupDryRun === true,
+            lookbackDays: args.statsCleanupLookbackDays ?? null,
+          });
+          const window =
+            result.lookbackDays === null ? 'all history' : `last ${result.lookbackDays}d`;
+          deps.logInfo(
+            `Stats duplicate-line cleanup ${result.dryRun ? 'preview' : 'complete'} (${window}): ` +
+              `scanned=${result.scannedLines} bursts=${result.burstGroups} ` +
+              `removedLines=${result.removedLines} removedWordCounts=${result.removedWordOccurrences} ` +
+              `removedKanjiCounts=${result.removedKanjiOccurrences}`,
+          );
+          for (const sample of result.samples.slice(0, 5)) {
+            deps.logInfo(
+              `  ${sample.videoTitle ?? `video ${sample.videoId}`}: "${sample.text}" x${sample.frames}`,
+            );
+          }
+          writeResponseSafe(args.statsResponsePath, { ok: true });
+          return;
+        }
         if (!args.statsCleanupLifetime || !tracker.rebuildLifetimeSummaries) {
           throw new Error('Stats cleanup mode is not available.');
         }

src/renderer/subtitle-render.test.ts

@@ -1004,6 +1004,24 @@ test('normalizeSubtitle collapses explicit line breaks when collapseLineBreaks i
   );
 });
 
+test('normalizeSubtitle leaves already-decoded text alone', () => {
+  // Primary subtitle text is decoded from ASS once, upstream: by mpv for live lines and
+  // by the cue parser for prefetched ones. A brace that survives that is literal text.
+  assert.equal(normalizeSubtitle('本文{\\pos(1,2)'), '本文{\\pos(1,2)');
+  assert.equal(normalizeSubtitle('  余白  ', false), '  余白  ');
+});
+
+test('prepareSecondarySubtitleLines drops ASS vector drawing runs', () => {
+  assert.deepEqual(
+    prepareSecondarySubtitleLines(
+      '{\\an5\\pos(730,1042)\\p1\\blur1}m 20 0 b 10 0 0 10 0 20 b 0 31 10 40 20 40 {\\p0}',
+    ),
+    [],
+  );
+  assert.deepEqual(prepareSecondarySubtitleLines('{\\p1}m 0 0 l 10 10{\\p0}本文'), ['本文']);
+  assert.deepEqual(prepareSecondarySubtitleLines('{\\pos(960,1068)\\bord3}位置指定'), ['位置指定']);
+});
+
 test('shouldRenderTokenizedSubtitle enables token rendering when tokens exist', () => {
   assert.equal(shouldRenderTokenizedSubtitle(5), true);
   assert.equal(shouldRenderTokenizedSubtitle(0), false);

src/renderer/subtitle-render.ts

@@ -5,6 +5,7 @@ import type {
   SubtitleData,
   SubtitleRendererStyleConfig,
 } from '../types';
+import { assToPlainText, normalizePlainSubtitleText } from '../core/services/ass-text.js';
 import type { RendererContext } from './context';
 import { PRIMARY_SUB_VISIBLE_ON_YOMITAN_POPUP_CLASS } from './yomitan-popup.js';
 
@@ -42,17 +43,10 @@ function isWhitespaceOnly(value: string): boolean {
   return value.trim().length === 0;
 }
 
+// Text reaching the overlay has already been decoded from ASS -- by mpv for live lines,
+// by the cue parser for prefetched ones -- so this only settles line breaks.
 export function normalizeSubtitle(text: string, trim = true, collapseLineBreaks = false): string {
-  if (!text) return '';
-
-  let normalized = text.replace(/\\N/g, '\n').replace(/\\n/g, '\n');
-  normalized = normalized.replace(/\{[^}]*\}/g, '');
-  if (collapseLineBreaks) {
-    normalized = normalized.replace(/\n/g, ' ');
-    normalized = normalized.replace(/\s+/g, ' ');
-  }
-
-  return trim ? normalized.trim() : normalized;
+  return normalizePlainSubtitleText(text, { trim, collapseLineBreaks });
 }
 
 const HEX_COLOR_PATTERN = /^#(?:[0-9a-fA-F]{3}|[0-9a-fA-F]{4}|[0-9a-fA-F]{6}|[0-9a-fA-F]{8})$/;
@@ -672,7 +666,10 @@ function isKaraokeLikeLineSet(lines: string[]): boolean {
 }
 
 export function prepareSecondarySubtitleLines(text: string): string[] {
-  const normalized = normalizeSubtitle(text, true, false);
+  // The one display-side ASS decode: secondary text also reaches the overlay from
+  // websocket clients that forward their source line untouched, so unlike the primary
+  // path it cannot assume mpv already decoded it.
+  const normalized = assToPlainText(text).trim();
 
   if (!normalized) return [];
 

src/shared/mpv-x11-backend.test.ts

@@ -99,9 +99,10 @@ test('applyX11EnvOverrides strips Wayland hints and pins session type to x11', (
   assert.equal(result.XDG_SESSION_TYPE, 'x11');
 });
 
-test('MPV_X11_BACKEND_ARGS pins the GPU stack to X11', () => {
-  assert.deepEqual(
-    [...MPV_X11_BACKEND_ARGS],
-    ['--vo=gpu', '--gpu-api=opengl', '--gpu-context=x11egl,x11'],
+test('MPV_X11_BACKEND_ARGS pins the window context to X11 without overriding the renderer', () => {
+  assert.deepEqual([...MPV_X11_BACKEND_ARGS], ['--gpu-context=x11vk,x11egl,x11']);
+  assert.equal(
+    MPV_X11_BACKEND_ARGS.some((arg) => arg.startsWith('--vo=') || arg.startsWith('--gpu-api=')),
+    false,
   );
 });

src/shared/mpv-x11-backend.ts

@@ -12,12 +12,18 @@
   so the gate and the mpv backend args stay in one place.
 */
 
-/** mpv args that pin the GPU/windowing stack to X11/XWayland (libGL via EGL on X11). */
-export const MPV_X11_BACKEND_ARGS = [
-  '--vo=gpu',
-  '--gpu-api=opengl',
-  '--gpu-context=x11egl,x11',
-] as const;
+/**
+ * mpv args that pin the *windowing* stack to X11/XWayland, in Vulkan-then-OpenGL order.
+ * mpv walks the list and skips contexts that do not match the configured `--gpu-api`,
+ * so this works for both a Vulkan and an OpenGL config.
+ *
+ * Deliberately does NOT set `--vo`/`--gpu-api`: forcing `--vo=gpu --gpu-api=opengl` here
+ * used to drop configs off `vo=gpu-next` onto the legacy renderer, where user shaders
+ * written for gpu-next (e.g. ArtCNN, `//!COMPONENTS 4` LUMA hooks) abort mpv with
+ * `copy_image: Assertion '*offset + count < sizeof(dst)' failed` as soon as their
+ * upscale-only `//!WHEN` condition turns on, i.e. on the first fullscreen toggle.
+ */
+export const MPV_X11_BACKEND_ARGS = ['--gpu-context=x11vk,x11egl,x11'] as const;
 
 export type LinuxDesktopEnv = {
   xdgCurrentDesktop: string;

src/subtitle-timing-tracker.ts

@@ -16,6 +16,8 @@
  * along with this program.  If not, see <https://www.gnu.org/licenses/>.
  */
 
+import { normalizePlainSubtitleText } from './core/services/ass-text';
+
 interface TimingEntry {
   startTime: number;
   endTime: number;
@@ -191,23 +193,14 @@ export class SubtitleTimingTracker {
     return costs[shorter.length] || 0;
   }
 
+  // Both sides take text mpv has already decoded from ASS; only whitespace differs
+  // between the lookup key (single line) and the display form (line breaks kept).
   private normalizeText(text: string): string {
-    return text
-      .replace(/\\N/g, ' ')
-      .replace(/\\n/g, ' ')
-      .replace(/\n/g, ' ')
-      .replace(/{[^}]*}/g, '')
-      .replace(/\s+/g, ' ')
-      .trim();
+    return normalizePlainSubtitleText(text, { collapseLineBreaks: true });
   }
 
   private prepareDisplayText(text: string): string {
-    // Convert ASS/SSA newlines to real newlines, strip tags
-    return text
-      .replace(/\\N/g, '\n')
-      .replace(/\\n/g, '\n')
-      .replace(/{[^}]*}/g, '')
-      .trim();
+    return normalizePlainSubtitleText(text);
   }
 
   private startCleanup(): void {

src/types/stats-http-contract.ts

@@ -18,6 +18,7 @@ import type {
   SessionTimelinePoint,
   StatsAnkiNoteInfo,
   StatsCoverImagesData,
+  StatsDuplicateLineCleanupResult,
   StatsExcludedWord,
   StreakCalendarDay,
   TrendsDashboardData,
@@ -31,6 +32,13 @@ export type StatsTrendRange = '7d' | '30d' | '90d' | '365d' | 'all';
 export type StatsTrendGroupBy = 'day' | 'month';
 export type StatsMineMode = 'word' | 'sentence' | 'audio';
 
+/** Body of `POST /api/stats/maintenance/duplicate-lines`. */
+export interface StatsDuplicateLineCleanupRequest {
+  dryRun?: boolean;
+  /** Null means every recorded line, whatever its age. */
+  lookbackDays?: number | null;
+}
+
 export interface StatsSessionKnownWordsTimelinePoint {
   linesSeen: number;
   knownWordsSeen: number;
@@ -125,6 +133,7 @@ export interface StatsJsonResponseMap {
   vocabulary: VocabularyEntry[];
   excludedWords: StatsExcludedWord[];
   setExcludedWords: StatsOkResponse;
+  duplicateLineCleanup: StatsDuplicateLineCleanupResult;
   wordOccurrences: VocabularyOccurrenceEntry[];
   sentenceSearch: SentenceSearchResult[];
   kanji: KanjiEntry[];
@@ -178,6 +187,9 @@ export interface StatsHttpClient {
   getVocabulary: (limit?: number) => Promise<VocabularyEntry[]>;
   getExcludedWords: () => Promise<StatsExcludedWord[]>;
   setExcludedWords: (words: StatsExcludedWord[]) => Promise<void>;
+  cleanupDuplicateLines: (
+    options?: StatsDuplicateLineCleanupRequest,
+  ) => Promise<StatsDuplicateLineCleanupResult>;
   getWordOccurrences: (
     headword: string,
     word: string,

src/types/stats-wire.ts

@@ -82,6 +82,29 @@ export interface StatsExcludedWord {
   reading: string;
 }
 
+/** One animation burst the duplicate-line cleanup found in the stats database. */
+export interface StatsDuplicateLineSample {
+  videoId: number;
+  videoTitle: string | null;
+  text: string;
+  /** Events recorded for this run, including the one that is kept. */
+  frames: number;
+  removedLines: number;
+  startMs: number;
+  endMs: number;
+}
+
+export interface StatsDuplicateLineCleanupResult {
+  dryRun: boolean;
+  lookbackDays: number | null;
+  scannedLines: number;
+  burstGroups: number;
+  removedLines: number;
+  removedWordOccurrences: number;
+  removedKanjiOccurrences: number;
+  samples: StatsDuplicateLineSample[];
+}
+
 export interface StatsCoverImage {
   contentType: string;
   dataUrl: string;

stats/src/components/vocabulary/DuplicateLineCleanup.test.tsx

@@ -0,0 +1,254 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { Window } from 'happy-dom';
+import { act } from 'react';
+import { createRoot } from 'react-dom/client';
+import { apiClient } from '../../lib/api-client';
+import type { StatsDuplicateLineCleanupResult } from '../../types/stats';
+import { DuplicateLineCleanup } from './DuplicateLineCleanup';
+
+interface TestWindow extends Window {
+  IS_REACT_ACT_ENVIRONMENT?: boolean;
+}
+
+function installDom(): () => void {
+  const previousWindow = globalThis.window;
+  const previousDocument = globalThis.document;
+  const previousHTMLElement = globalThis.HTMLElement;
+  const previousISReactActEnvironment = (
+    globalThis as typeof globalThis & { IS_REACT_ACT_ENVIRONMENT?: boolean }
+  ).IS_REACT_ACT_ENVIRONMENT;
+  const window = new Window() as TestWindow;
+
+  Object.defineProperty(globalThis, 'window', { value: window, configurable: true });
+  Object.defineProperty(globalThis, 'document', { value: window.document, configurable: true });
+  Object.defineProperty(globalThis, 'HTMLElement', {
+    value: window.HTMLElement,
+    configurable: true,
+  });
+  (
+    globalThis as typeof globalThis & { IS_REACT_ACT_ENVIRONMENT?: boolean }
+  ).IS_REACT_ACT_ENVIRONMENT = true;
+
+  return () => {
+    Object.defineProperty(globalThis, 'window', { value: previousWindow, configurable: true });
+    Object.defineProperty(globalThis, 'document', { value: previousDocument, configurable: true });
+    Object.defineProperty(globalThis, 'HTMLElement', {
+      value: previousHTMLElement,
+      configurable: true,
+    });
+    (
+      globalThis as typeof globalThis & { IS_REACT_ACT_ENVIRONMENT?: boolean }
+    ).IS_REACT_ACT_ENVIRONMENT = previousISReactActEnvironment;
+  };
+}
+
+function findButton(container: Element, label: string): HTMLButtonElement {
+  const match = [...container.querySelectorAll('button')].find(
+    (button) => (button.textContent ?? '').trim() === label,
+  );
+  assert.ok(match, `expected a "${label}" button`);
+  return match as unknown as HTMLButtonElement;
+}
+
+/** The backdrop stays clickable during an apply, so it reaches the guard in `close`. */
+function findBackdrop(container: Element): HTMLButtonElement {
+  const match = container.querySelector('button[aria-label="Close duplicate line cleanup"]');
+  assert.ok(match, 'expected the backdrop close button');
+  return match as unknown as HTMLButtonElement;
+}
+
+function deferred<T>(): { promise: Promise<T>; resolve: (value: T) => void } {
+  let resolve!: (value: T) => void;
+  const promise = new Promise<T>((done) => {
+    resolve = done;
+  });
+  return { promise, resolve };
+}
+
+function summary(
+  overrides: Partial<StatsDuplicateLineCleanupResult> = {},
+): StatsDuplicateLineCleanupResult {
+  return {
+    dryRun: false,
+    lookbackDays: 30,
+    scannedLines: 900,
+    burstGroups: 2,
+    removedLines: 180,
+    removedWordOccurrences: 540,
+    removedKanjiOccurrences: 120,
+    samples: [],
+    ...overrides,
+  };
+}
+
+interface Harness {
+  container: Element;
+  cleanedCalls: () => number;
+  closedCalls: () => number;
+  teardown: () => void;
+}
+
+async function mount(cleanup: (typeof apiClient)['cleanupDuplicateLines']): Promise<Harness> {
+  const uninstallDom = installDom();
+  const originalCleanup = apiClient.cleanupDuplicateLines;
+  apiClient.cleanupDuplicateLines = cleanup;
+
+  let cleaned = 0;
+  let closed = 0;
+  const container = document.createElement('div');
+  document.body.append(container);
+  const root = createRoot(container);
+
+  await act(async () => {
+    root.render(
+      <DuplicateLineCleanup
+        onClose={() => {
+          closed += 1;
+        }}
+        onCleaned={() => {
+          cleaned += 1;
+        }}
+      />,
+    );
+  });
+
+  return {
+    container,
+    cleanedCalls: () => cleaned,
+    closedCalls: () => closed,
+    teardown: () => {
+      apiClient.cleanupDuplicateLines = originalCleanup;
+      uninstallDom();
+    },
+  };
+}
+
+test('a reload is still owed after a later scan replaces the applied result', async () => {
+  const harness = await mount(async ({ dryRun } = {}) => summary({ dryRun: dryRun === true }));
+
+  try {
+    await act(async () => {
+      findButton(harness.container, 'Scan').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Clean Up').click();
+    });
+    assert.equal(harness.cleanedCalls(), 0, 'reload must wait for the result to be read');
+
+    // The follow-up scan clears the applied summary, but the rows are already gone.
+    await act(async () => {
+      findButton(harness.container, 'Scan').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Close').click();
+    });
+
+    assert.equal(harness.cleanedCalls(), 1);
+    assert.equal(harness.closedCalls(), 1);
+  } finally {
+    harness.teardown();
+  }
+});
+
+test('a reload is still owed after the lookback window changes', async () => {
+  const harness = await mount(async ({ dryRun } = {}) => summary({ dryRun: dryRun === true }));
+
+  try {
+    await act(async () => {
+      findButton(harness.container, 'Scan').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Clean Up').click();
+    });
+    await act(async () => {
+      findButton(harness.container, '7 days').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Close').click();
+    });
+
+    assert.equal(harness.cleanedCalls(), 1);
+  } finally {
+    harness.teardown();
+  }
+});
+
+test('closing is refused while an apply is in flight', async () => {
+  const pending = deferred<StatsDuplicateLineCleanupResult>();
+  const harness = await mount(async ({ dryRun } = {}) =>
+    dryRun === true ? summary({ dryRun: true }) : pending.promise,
+  );
+
+  try {
+    await act(async () => {
+      findButton(harness.container, 'Scan').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Clean Up').click();
+    });
+
+    assert.equal(findButton(harness.container, 'Close').disabled, true);
+    // The backdrop is never disabled, so this is the path that has to be refused.
+    await act(async () => {
+      findBackdrop(harness.container).click();
+    });
+    assert.equal(harness.closedCalls(), 0, 'the modal must stay open mid-apply');
+    assert.equal(harness.cleanedCalls(), 0);
+
+    await act(async () => {
+      pending.resolve(summary({ removedLines: 12 }));
+      await pending.promise;
+    });
+    await act(async () => {
+      findButton(harness.container, 'Close').click();
+    });
+
+    assert.equal(harness.closedCalls(), 1);
+    assert.equal(harness.cleanedCalls(), 1);
+  } finally {
+    harness.teardown();
+  }
+});
+
+test('a scan on its own owes no reload', async () => {
+  const harness = await mount(async ({ dryRun } = {}) => summary({ dryRun: dryRun === true }));
+
+  try {
+    await act(async () => {
+      findButton(harness.container, 'Scan').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Close').click();
+    });
+
+    assert.equal(harness.cleanedCalls(), 0);
+    assert.equal(harness.closedCalls(), 1);
+  } finally {
+    harness.teardown();
+  }
+});
+
+test('an apply that removes nothing owes no reload', async () => {
+  // The scan saw work to do, but by the time it ran another cleanup had taken it.
+  const harness = await mount(async ({ dryRun } = {}) =>
+    dryRun === true ? summary({ dryRun: true }) : summary({ burstGroups: 0, removedLines: 0 }),
+  );
+
+  try {
+    await act(async () => {
+      findButton(harness.container, 'Scan').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Clean Up').click();
+    });
+    await act(async () => {
+      findButton(harness.container, 'Close').click();
+    });
+
+    assert.equal(harness.cleanedCalls(), 0);
+    assert.equal(harness.closedCalls(), 1);
+  } finally {
+    harness.teardown();
+  }
+});

stats/src/components/vocabulary/DuplicateLineCleanup.tsx

@@ -0,0 +1,202 @@
+import { useCallback, useState } from 'react';
+import { getStatsClient } from '../../hooks/useStatsApi';
+import { formatNumber } from '../../lib/formatters';
+import type { StatsDuplicateLineCleanupResult } from '../../types/stats';
+
+interface DuplicateLineCleanupProps {
+  onClose: () => void;
+  /** Called after rows are actually removed, so the charts can reload. */
+  onCleaned: () => void;
+}
+
+const LOOKBACK_CHOICES: Array<{ label: string; days: number | null }> = [
+  { label: '7 days', days: 7 },
+  { label: '30 days', days: 30 },
+  { label: '90 days', days: 90 },
+  { label: '1 year', days: 365 },
+  { label: 'All time', days: null },
+];
+
+function formatTimecode(ms: number): string {
+  const totalSeconds = Math.max(0, Math.floor(ms / 1000));
+  const minutes = Math.floor(totalSeconds / 60);
+  const seconds = totalSeconds % 60;
+  return `${minutes}:${String(seconds).padStart(2, '0')}`;
+}
+
+export function DuplicateLineCleanup({ onClose, onCleaned }: DuplicateLineCleanupProps) {
+  const [lookbackDays, setLookbackDays] = useState<number | null>(30);
+  const [preview, setPreview] = useState<StatsDuplicateLineCleanupResult | null>(null);
+  const [applied, setApplied] = useState<StatsDuplicateLineCleanupResult | null>(null);
+  const [busy, setBusy] = useState<'scan' | 'apply' | null>(null);
+  const [error, setError] = useState<string | null>(null);
+  // Survives everything the displayed result does not: another scan, a different window.
+  // Rows are gone from the moment an apply succeeds, so the reload is owed until it runs.
+  const [needsReload, setNeedsReload] = useState(false);
+
+  const run = useCallback(
+    async (dryRun: boolean) => {
+      setBusy(dryRun ? 'scan' : 'apply');
+      setError(null);
+      try {
+        const result = await getStatsClient().cleanupDuplicateLines({ dryRun, lookbackDays });
+        if (dryRun) {
+          setPreview(result);
+          setApplied(null);
+        } else {
+          setApplied(result);
+          setPreview(null);
+          if (result.removedLines > 0) {
+            setNeedsReload(true);
+          }
+        }
+      } catch (cause) {
+        setError(cause instanceof Error ? cause.message : String(cause));
+      } finally {
+        setBusy(null);
+      }
+    },
+    [lookbackDays],
+  );
+
+  // Reloading the vocabulary tables unmounts this modal along with the rest of the tab,
+  // so it waits for the user to close: they get to read what was removed first. Closing
+  // is refused mid-apply, which would drop the reload on the floor along with the report.
+  const close = useCallback(() => {
+    if (busy === 'apply') {
+      return;
+    }
+    if (needsReload) {
+      onCleaned();
+    }
+    onClose();
+  }, [busy, needsReload, onCleaned, onClose]);
+
+  const result = applied ?? preview;
+  const nothingToDo = preview !== null && preview.removedLines === 0;
+
+  return (
+    <div className="fixed inset-0 z-50">
+      <button
+        type="button"
+        aria-label="Close duplicate line cleanup"
+        className="absolute inset-0 bg-ctp-crust/70 backdrop-blur-[2px]"
+        onClick={close}
+      />
+      <div className="absolute inset-x-0 top-1/2 mx-auto max-w-xl -translate-y-1/2 rounded-xl border border-ctp-surface1 bg-ctp-mantle shadow-2xl">
+        <div className="flex items-center justify-between border-b border-ctp-surface1 px-5 py-4">
+          <h2 className="text-sm font-semibold text-ctp-text">Duplicate Lines</h2>
+          <button
+            type="button"
+            disabled={busy === 'apply'}
+            className="rounded-md border border-ctp-surface2 px-3 py-1.5 text-xs font-medium text-ctp-subtext0 transition hover:border-ctp-blue hover:text-ctp-blue disabled:opacity-50"
+            onClick={close}
+          >
+            Close
+          </button>
+        </div>
+
+        <div className="space-y-4 px-5 py-4">
+          <p className="text-xs leading-relaxed text-ctp-subtext0">
+            Typeset subtitles — karaoke openings, animated signs — are authored as one event per
+            animation frame, and older versions counted every frame as its own line. This finds
+            those runs and collapses each one back to a single line, giving back the word and kanji
+            counts they inflated. Ordinary repeated dialogue is left alone.
+          </p>
+
+          <div>
+            <div className="mb-2 text-xs font-medium text-ctp-subtext1">Look back over</div>
+            <div className="flex flex-wrap gap-2">
+              {LOOKBACK_CHOICES.map((choice) => (
+                <button
+                  key={choice.label}
+                  type="button"
+                  disabled={busy !== null}
+                  onClick={() => {
+                    setLookbackDays(choice.days);
+                    setPreview(null);
+                    setApplied(null);
+                  }}
+                  className={`rounded-lg border px-3 py-1.5 text-xs transition disabled:opacity-50 ${
+                    lookbackDays === choice.days
+                      ? 'border-ctp-blue/50 bg-ctp-surface2 text-ctp-text'
+                      : 'border-ctp-surface1 bg-ctp-surface0 text-ctp-overlay2 hover:text-ctp-subtext0'
+                  }`}
+                >
+                  {choice.label}
+                </button>
+              ))}
+            </div>
+          </div>
+
+          {error && (
+            <div className="rounded-lg border border-ctp-red/30 bg-ctp-red/10 px-3 py-2 text-xs text-ctp-red">
+              {error}
+            </div>
+          )}
+
+          {result && (
+            <div className="rounded-lg bg-ctp-surface0 px-4 py-3">
+              <div className="text-sm text-ctp-text">
+                {applied
+                  ? `Removed ${formatNumber(applied.removedLines)} repeated lines`
+                  : nothingToDo
+                    ? 'No animation bursts found in this window'
+                    : `Found ${formatNumber(preview!.burstGroups)} bursts covering ${formatNumber(preview!.removedLines)} extra lines`}
+              </div>
+              <div className="mt-1 text-xs text-ctp-overlay2">
+                {formatNumber(result.scannedLines)} lines scanned ·{' '}
+                {formatNumber(result.removedWordOccurrences)} word counts ·{' '}
+                {formatNumber(result.removedKanjiOccurrences)} kanji counts
+                {applied ? ' removed' : ' would be removed'}
+              </div>
+
+              {result.samples.length > 0 && (
+                <div className="mt-3 max-h-52 space-y-1.5 overflow-y-auto">
+                  {result.samples.map((sample) => (
+                    <div
+                      key={`${sample.videoId}:${sample.startMs}:${sample.text}`}
+                      className="flex items-center justify-between gap-3 rounded-md bg-ctp-mantle px-3 py-1.5"
+                    >
+                      <div className="min-w-0">
+                        <div className="truncate text-xs text-ctp-text">{sample.text}</div>
+                        <div className="truncate text-[11px] text-ctp-overlay1">
+                          {sample.videoTitle ?? `Video ${sample.videoId}`} ·{' '}
+                          {formatTimecode(sample.startMs)}
+                        </div>
+                      </div>
+                      <span className="shrink-0 text-xs text-ctp-peach">×{sample.frames}</span>
+                    </div>
+                  ))}
+                </div>
+              )}
+            </div>
+          )}
+
+          <div className="flex items-center justify-end gap-2">
+            <button
+              type="button"
+              disabled={busy !== null}
+              onClick={() => void run(true)}
+              className="rounded-md border border-ctp-surface2 px-3 py-1.5 text-xs font-medium text-ctp-subtext0 transition hover:border-ctp-blue hover:text-ctp-blue disabled:opacity-50"
+            >
+              {busy === 'scan' ? 'Scanning…' : 'Scan'}
+            </button>
+            <button
+              type="button"
+              disabled={busy !== null || preview === null || nothingToDo}
+              onClick={() => void run(false)}
+              className="rounded-md border border-ctp-red/30 px-3 py-1.5 text-xs font-medium text-ctp-red transition hover:bg-ctp-red/10 disabled:opacity-40"
+            >
+              {busy === 'apply' ? 'Cleaning…' : 'Clean Up'}
+            </button>
+          </div>
+          <p className="text-[11px] text-ctp-overlay1">
+            Scan first: cleanup removes rows and cannot be undone. Session watch time and lines-seen
+            totals are left untouched.
+          </p>
+        </div>
+      </div>
+    </div>
+  );
+}

stats/src/components/vocabulary/VocabularyTab.tsx

@@ -5,6 +5,7 @@ import { WordList } from './WordList';
 import { KanjiBreakdown } from './KanjiBreakdown';
 import { KanjiDetailPanel } from './KanjiDetailPanel';
 import { ExclusionManager } from './ExclusionManager';
+import { DuplicateLineCleanup } from './DuplicateLineCleanup';
 import { formatNumber } from '../../lib/formatters';
 import { TrendChart } from '../trends/TrendChart';
 import { FrequencyRankTable } from './FrequencyRankTable';
@@ -34,10 +35,11 @@ export function VocabularyTab({
   onRemoveExclusion,
   onClearExclusions,
 }: VocabularyTabProps) {
-  const { words, kanji, knownWords, loading, error } = useVocabulary();
+  const { words, kanji, knownWords, loading, error, reload } = useVocabulary();
   const [selectedKanjiId, setSelectedKanjiId] = useState<number | null>(null);
   const [hideNames, setHideNames] = useState(false);
   const [showExclusionManager, setShowExclusionManager] = useState(false);
+  const [showDuplicateLineCleanup, setShowDuplicateLineCleanup] = useState(false);
 
   const hasNames = useMemo(() => words.some(isProperNoun), [words]);
   const filteredWords = useMemo(() => {
@@ -129,6 +131,13 @@ export function VocabularyTab({
             Hide Names
           </button>
         )}
+        <button
+          type="button"
+          onClick={() => setShowDuplicateLineCleanup(true)}
+          className="shrink-0 rounded-lg border border-ctp-surface1 bg-ctp-surface0 px-3 py-2 text-xs text-ctp-overlay2 transition-colors hover:text-ctp-subtext0"
+        >
+          Duplicates
+        </button>
         <button
           type="button"
           onClick={() => setShowExclusionManager(true)}
@@ -193,6 +202,13 @@ export function VocabularyTab({
           onClose={() => setShowExclusionManager(false)}
         />
       )}
+
+      {showDuplicateLineCleanup && (
+        <DuplicateLineCleanup
+          onClose={() => setShowDuplicateLineCleanup(false)}
+          onCleaned={reload}
+        />
+      )}
     </div>
   );
 }

stats/src/hooks/useVocabulary.ts

@@ -1,4 +1,4 @@
-import { useState, useEffect } from 'react';
+import { useState, useEffect, useCallback } from 'react';
 import { getStatsClient } from './useStatsApi';
 import type { VocabularyEntry, KanjiEntry } from '../types/stats';
 
@@ -8,6 +8,9 @@ export function useVocabulary() {
   const [knownWords, setKnownWords] = useState<Set<string>>(new Set());
   const [loading, setLoading] = useState(true);
   const [error, setError] = useState<string | null>(null);
+  // Bumped by `reload` after maintenance rewrites the vocabulary tables.
+  const [reloadToken, setReloadToken] = useState(0);
+  const reload = useCallback(() => setReloadToken((token) => token + 1), []);
 
   useEffect(() => {
     let cancelled = false;
@@ -46,7 +49,7 @@ export function useVocabulary() {
     return () => {
       cancelled = true;
     };
-  }, []);
+  }, [reloadToken]);
 
-  return { words, kanji, knownWords, loading, error };
+  return { words, kanji, knownWords, loading, error, reload };
 }

stats/src/lib/api-client.ts

@@ -6,6 +6,8 @@ import type {
   StatsAnkiNotesInfoRequest,
   StatsCoverImagesRequest,
   StatsDeleteSessionsRequest,
+  StatsDuplicateLineCleanupRequest,
+  StatsDuplicateLineCleanupResult,
   StatsExcludedWordsRequest,
   StatsHttpClient,
   StatsJsonResponseMap,
@@ -102,6 +104,19 @@ export const apiClient = {
       body: JSON.stringify({ words } satisfies StatsExcludedWordsRequest),
     });
   },
+  cleanupDuplicateLines: async (
+    options: StatsDuplicateLineCleanupRequest = {},
+  ): Promise<StatsDuplicateLineCleanupResult> => {
+    const res = await fetchResponse('/api/stats/maintenance/duplicate-lines', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        dryRun: options.dryRun === true,
+        lookbackDays: options.lookbackDays ?? null,
+      } satisfies StatsDuplicateLineCleanupRequest),
+    });
+    return res.json() as Promise<StatsDuplicateLineCleanupResult>;
+  },
   getWordOccurrences: (headword: string, word: string, reading: string, limit = 50, offset = 0) =>
     fetchJson(
       'wordOccurrences',