fix(subtitles): recover canonical lines from ASS animation (#207)

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

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

src/core/services/mining.test.ts

@@ -125,6 +125,49 @@ test('mineSentenceCard creates sentence card from mpv subtitle state', async ()
   ]);
 });
 
+test('mineSentenceCard prefers a canonical primary subtitle snapshot', async () => {
+  const created: Array<{
+    sentence: string;
+    startTime: number;
+    endTime: number;
+    secondarySub?: string;
+  }> = [];
+
+  await mineSentenceCard({
+    ankiIntegration: {
+      updateLastAddedFromClipboard: async () => {},
+      triggerFieldGroupingForLastAddedCard: async () => {},
+      markLastCardAsAudioCard: async () => {},
+      createSentenceCard: async (sentence, startTime, endTime, secondarySub) => {
+        created.push({ sentence, startTime, endTime, secondarySub });
+        return true;
+      },
+    },
+    mpvClient: {
+      connected: true,
+      currentSubText: '今今今手手手',
+      currentSubStart: 11.4,
+      currentSubEnd: 11.8,
+      currentSecondarySubText: 'English subtitle',
+    },
+    primarySubtitle: {
+      text: '今　手にある物差しでは',
+      startTime: 11.13,
+      endTime: 13.83,
+    },
+    showMpvOsd: () => {},
+  });
+
+  assert.deepEqual(created, [
+    {
+      sentence: '今　手にある物差しでは',
+      startTime: 11.13,
+      endTime: 13.83,
+      secondarySub: 'English subtitle',
+    },
+  ]);
+});
+
 test('mineSentenceCard refreshes secondary subtitle text before creating card', async () => {
   const created: Array<{ sentence: string; secondarySub?: string }> = [];
   const requestedProperties: string[] = [];

src/core/services/mining.ts

@@ -131,8 +131,8 @@ function normalizeSecondarySubText(text: unknown, primaryText: string): string |
 
 async function getCurrentSecondarySubTextForSentenceCard(
   mpvClient: MpvClientLike,
+  primaryText: string,
 ): Promise<string | undefined> {
-  const primaryText = mpvClient.currentSubText;
   if (mpvClient.requestProperty) {
     try {
       const latestSecondaryText = await mpvClient.requestProperty('secondary-sub-text');
@@ -175,6 +175,7 @@ export async function markLastCardAsAudioCard(deps: {
 export async function mineSentenceCard(deps: {
   ankiIntegration: AnkiIntegrationLike | null;
   mpvClient: MpvClientLike | null;
+  primarySubtitle?: Pick<SubtitleMiningContext, 'text' | 'startTime' | 'endTime'>;
   showMpvOsd: (text: string) => void;
 }): Promise<boolean> {
   const anki = requireAnkiIntegration(deps.ankiIntegration, deps.showMpvOsd);
@@ -185,16 +186,17 @@ export async function mineSentenceCard(deps: {
     deps.showMpvOsd('MPV not connected');
     return false;
   }
-  if (!mpvClient.currentSubText) {
+  const primaryText = deps.primarySubtitle?.text ?? mpvClient.currentSubText;
+  if (!primaryText) {
     deps.showMpvOsd('No current subtitle');
     return false;
   }
 
-  const secondarySubText = await getCurrentSecondarySubTextForSentenceCard(mpvClient);
+  const secondarySubText = await getCurrentSecondarySubTextForSentenceCard(mpvClient, primaryText);
   return await anki.createSentenceCard(
-    mpvClient.currentSubText,
-    mpvClient.currentSubStart,
-    mpvClient.currentSubEnd,
+    primaryText,
+    deps.primarySubtitle?.startTime ?? mpvClient.currentSubStart,
+    deps.primarySubtitle?.endTime ?? mpvClient.currentSubEnd,
     secondarySubText,
   );
 }

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

@@ -27,17 +27,188 @@ function cueKey(cue: SubtitleCue): string {
 
 /**
  * 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.
+ * often a layered ASS event stacking a shadow copy under the visible one. When one of
+ * the duplicates is a recovered canonical cue, that copy survives: dropping it would
+ * strip the `source` marker and animation envelope the live overlay substitutes on.
  */
 function collapseExactDuplicates(cues: AnnotatedSubtitleCue[]): AnnotatedSubtitleCue[] {
-  const seen = new Set<string>();
-  return cues.filter((cue) => {
+  const survivorByKey = new Map<string, AnnotatedSubtitleCue>();
+  const keysInOrder: string[] = [];
+  for (const cue of cues) {
     const key = cueKey(cue);
-    if (seen.has(key)) {
-      return false;
+    const existing = survivorByKey.get(key);
+    if (!existing) {
+      survivorByKey.set(key, cue);
+      keysInOrder.push(key);
+    } else if (!existing.source && cue.source) {
+      survivorByKey.set(key, cue);
     }
-    seen.add(key);
+  }
+  return keysInOrder.map((key) => survivorByKey.get(key)!);
+}
+
+const SPATIAL_ASS_OVERRIDE_COMMANDS = new Set([
+  'a',
+  'an',
+  'clip',
+  'iclip',
+  'move',
+  'org',
+  'pbo',
+  'pos',
+  'q',
+]);
+
+interface RepeatedPhaseRun {
+  cues: AnnotatedSubtitleCue[];
+  indices: number[];
+}
+
+// A changing override signature alone is weak: two ordinary repeats restyled with
+// different colors look identical to a phase pair. Real phase redraws carry a styling
+// stack over a full lyric line, and they exist to move a color/highlight boundary
+// *within* the line -- so every event also has an override block after visible text
+// began. An ordinary restyled repeat carries only a leading block and stays separate.
+const MIN_PHASE_EVIDENCE_OVERRIDES = 2;
+const MIN_PHASE_TEXT_LENGTH = 4;
+
+function hasMidLineOverrideBlock(rawText: string): boolean {
+  let sawVisibleText = false;
+  for (let i = 0; i < rawText.length; i += 1) {
+    if (rawText[i] === '{') {
+      const close = rawText.indexOf('}', i);
+      if (close === -1) {
+        // Unclosed brace renders as literal text; nothing after it is markup.
+        return false;
+      }
+      if (sawVisibleText) {
+        return true;
+      }
+      i = close;
+    } else if (!/\s/.test(rawText[i]!)) {
+      sawVisibleText = true;
+    }
+  }
+  return false;
+}
+
+function assStyleKey(cue: AnnotatedSubtitleCue): string {
+  return `${cue.style}\0${cue.name}\0${cue.layer}`;
+}
+
+function spatialOverrideSignature(cue: AnnotatedSubtitleCue): string {
+  return cue.overrides
+    .filter((command) => SPATIAL_ASS_OVERRIDE_COMMANDS.has(command.name.toLowerCase()))
+    .map((command) => `${command.name.toLowerCase()}(${command.args})`)
+    .join('|');
+}
+
+function hasStableSpatialOverrides(run: readonly AnnotatedSubtitleCue[]): boolean {
+  const firstSignature = spatialOverrideSignature(run[0]!);
+  return run.every((cue) => spatialOverrideSignature(cue) === firstSignature);
+}
+
+function hasDirectPhaseEvidence(run: readonly AnnotatedSubtitleCue[]): boolean {
+  // Phases redraw one authored line in place. Whatever the animation evidence, a run
+  // whose spatial placement changes is separate authored occurrences -- two flush
+  // same-text `\move` signs at different coordinates must never merge.
+  if (!hasStableSpatialOverrides(run)) {
+    return false;
+  }
+  if (run.every((cue) => hasAssTemporalOverride(cue.overrides))) {
     return true;
+  }
+  if (run.every((cue) => isAnimatedAssEffectKind(cue.effectKind))) {
+    return true;
+  }
+
+  const [first] = run;
+  return (
+    first!.text.replace(/\s+/gu, '').length >= MIN_PHASE_TEXT_LENGTH &&
+    run.every(
+      (cue) =>
+        cue.overrides.length >= MIN_PHASE_EVIDENCE_OVERRIDES &&
+        hasMidLineOverrideBlock(cue.rawText),
+    ) &&
+    run.some((cue) => cue.overrideSignature !== first!.overrideSignature)
+  );
+}
+
+function collectRepeatedPhaseRuns(cues: AnnotatedSubtitleCue[]): RepeatedPhaseRun[] {
+  const runs: RepeatedPhaseRun[] = [];
+  let start = 0;
+
+  while (start < cues.length) {
+    const first = cues[start]!;
+    const styleKey = assStyleKey(first);
+    let end = start;
+
+    while (end + 1 < cues.length) {
+      const current = cues[end]!;
+      const next = cues[end + 1]!;
+      const isFlush =
+        Math.abs(next.startTime - current.endTime) <= DUPLICATE_CUE_GAP_TOLERANCE_SECONDS;
+      if (
+        first.source === 'canonical-ass' ||
+        next.source === 'canonical-ass' ||
+        next.text !== first.text ||
+        assStyleKey(next) !== styleKey ||
+        !isFlush
+      ) {
+        break;
+      }
+      end += 1;
+    }
+
+    if (end > start) {
+      const indices = Array.from({ length: end - start + 1 }, (_, offset) => start + offset);
+      runs.push({
+        cues: indices.map((index) => cues[index]!),
+        indices,
+      });
+    }
+    start = end + 1;
+  }
+
+  return runs;
+}
+
+/**
+ * Some karaoke scripts redraw one complete lyric for each color/highlight phase. These
+ * events last far longer than animation frames, but are still one sidebar/history line.
+ * The events must prove themselves through direct animation metadata or changing
+ * non-spatial overrides. Plain repeated dialogue and separately positioned signs stay
+ * intact.
+ */
+function collapseAnimatedStylePhases(cues: AnnotatedSubtitleCue[]): AnnotatedSubtitleCue[] {
+  const runs = collectRepeatedPhaseRuns(cues);
+  if (runs.length === 0) {
+    return cues;
+  }
+
+  const dropped = new Set<number>();
+  const extendedEnd = new Map<number, number>();
+  for (const run of runs) {
+    if (!hasDirectPhaseEvidence(run.cues)) {
+      continue;
+    }
+
+    const [firstIndex, ...remainingIndices] = run.indices;
+    for (const index of remainingIndices) {
+      dropped.add(index);
+    }
+    extendedEnd.set(firstIndex!, Math.max(...run.cues.map((cue) => cue.endTime)));
+  }
+
+  if (dropped.size === 0) {
+    return cues;
+  }
+  return cues.flatMap((cue, index) => {
+    if (dropped.has(index)) {
+      return [];
+    }
+    const endTime = extendedEnd.get(index);
+    return endTime !== undefined ? [{ ...cue, endTime }] : [cue];
   });
 }
 
@@ -176,5 +347,8 @@ export function mergeDuplicateCues(
   cues: AnnotatedSubtitleCue[],
   format: SubtitleSourceFormat,
 ): AnnotatedSubtitleCue[] {
-  return collapseAnimationBursts(collapseExactDuplicates(cues), format);
+  const exactDeduplicated = collapseExactDuplicates(cues);
+  const phaseDeduplicated =
+    format === 'ass' ? collapseAnimatedStylePhases(exactDeduplicated) : exactDeduplicated;
+  return collapseAnimationBursts(phaseDeduplicated, format);
 }

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

@@ -327,6 +327,122 @@ test('parseSubtitleCues collapses per-frame karaoke duplicates into one cue', ()
   assert.equal(cues[0]!.text, '過ぎ去ってしまう瞬間を');
 });
 
+test('parseSubtitleCues collapses long full-line color phases', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:03:49.75,0:03:51.21,OPJP,,0,0,0,,{\\blur0.6\\c&H312D38&\\4c&HFFFFFF&}ちゃんと目を{\\4c&HD590FF&}合わせてよ',
+    'Dialogue: 1,0:03:51.21,0:03:52.25,OPJP,,0,0,0,,{\\blur0.6\\4c&H312D38&\\c&HFFFFFF&}ちゃんと目を{\\4c&HD590FF&}合わせてよ',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    {
+      startTime: 229.75,
+      endTime: 232.25,
+      text: 'ちゃんと目を合わせてよ',
+    },
+  ]);
+});
+
+test('parseSubtitleCues keeps ordinary repeated dialogue separate', () => {
+  // A single restyle tag on a repeated line is how ordinary dialogue gets decorated;
+  // it is not phase evidence, whatever the line length.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:00:01.00,0:00:02.00,OPJP,,0,0,0,,{\\c&H111111&}歌詞',
+    'Dialogue: 1,0:00:02.00,0:00:03.00,OPJP,,0,0,0,,{\\c&H222222&}歌詞',
+    'Dialogue: 1,0:00:04.00,0:00:05.00,OPJP,,0,0,0,,{\\c&H333333&}別の歌詞',
+    'Dialogue: 1,0:00:05.00,0:00:06.00,OPJP,,0,0,0,,{\\c&H444444&}別の歌詞',
+    'Dialogue: 8,0:00:07.00,0:00:08.00,Text - JP,,0,0,0,,えっ？',
+    'Dialogue: 8,0:00:08.00,0:00:09.00,Text - JP,,0,0,0,,えっ？',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    { startTime: 1, endTime: 2, text: '歌詞' },
+    { startTime: 2, endTime: 3, text: '歌詞' },
+    { startTime: 4, endTime: 5, text: '別の歌詞' },
+    { startTime: 5, endTime: 6, text: '別の歌詞' },
+    { startTime: 7, endTime: 8, text: 'えっ？' },
+    { startTime: 8, endTime: 9, text: 'えっ？' },
+  ]);
+});
+
+test('parseSubtitleCues keeps separately positioned temporal signs separate', () => {
+  // Two flush signs with the same text but different \move paths are separate authored
+  // occurrences, not phases of one redraw: temporal evidence alone must not merge them.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:02.00,Sign,,0,0,0,,{\\move(100,100,200,100)}立入禁止',
+    'Dialogue: 0,0:00:02.00,0:00:03.00,Sign,,0,0,0,,{\\move(500,400,600,400)}立入禁止',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    { startTime: 1, endTime: 2, text: '立入禁止' },
+    { startTime: 2, endTime: 3, text: '立入禁止' },
+  ]);
+});
+
+test('parseSubtitleCues keeps richly styled ordinary repeats separate', () => {
+  // Blur plus a changing color is still an ordinary restyle. Phase redraws are
+  // recognized by the color/highlight boundary moving *within* the line, which these
+  // leading-block-only events do not have.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:02.00,Dial,,0,0,0,,{\\blur0.4\\c&H111111&}待ってよ',
+    'Dialogue: 0,0:00:02.00,0:00:03.00,Dial,,0,0,0,,{\\blur0.4\\c&H222222&}待ってよ',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    { startTime: 1, endTime: 2, text: '待ってよ' },
+    { startTime: 2, endTime: 3, text: '待ってよ' },
+  ]);
+});
+
+test('parseSubtitleCues keeps canonical metadata when an identical plain cue exists', () => {
+  // A plain dialogue line can share exact timing and text with a recovered canonical
+  // cue from another style. The canonical copy must win the exact-duplicate collapse,
+  // or the live overlay loses the marker it substitutes on.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:05.00,0:00:08.00,Plain,,0,0,0,,ライン',
+    'Comment: 0,0:00:05.00,0:00:08.00,OP,,0,0,0,,ライン',
+    'Dialogue: 0,0:00:05.00,0:00:05.04,OP,,0,0,0,,{\\pos(1,1)\\clip(m 1 1)}ライン',
+    'Dialogue: 0,0:00:05.04,0:00:05.08,OP,,0,0,0,,{\\pos(1,1)\\clip(m 2 2)}ライン',
+    'Dialogue: 0,0:00:05.08,0:00:08.00,OP,,0,0,0,,{\\pos(1,1)\\clip(m 3 3)}ライン',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    {
+      startTime: 5,
+      endTime: 8,
+      text: 'ライン',
+      source: 'canonical-ass',
+      animationStartTime: 5,
+      animationEndTime: 8,
+    },
+  ]);
+});
+
+test('parseSubtitleCues keeps short styled repeats separate even with richer styling', () => {
+  // Two ordinary えっ lines restyled with different colors are two utterances, not two
+  // phases of one lyric: short text never satisfies the changing-override evidence path.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:01.00,0:00:02.00,Dial,,0,0,0,,{\\blur0.4\\c&H111111&}えっ',
+    'Dialogue: 0,0:00:02.00,0:00:03.00,Dial,,0,0,0,,{\\blur0.4\\c&H222222&}えっ',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    { startTime: 1, endTime: 2, text: 'えっ' },
+    { startTime: 2, endTime: 3, text: 'えっ' },
+  ]);
+});
+
 test('parseSubtitleCues keeps back-to-back plain dialogue repeats separate', () => {
   // Several characters greeting in turn: distinct utterances that happen to abut.
   const content = [
@@ -357,6 +473,194 @@ test('parseSubtitleCues collapses exact duplicate cues even without effect tags'
   assert.equal(cues.length, 1);
 });
 
+test('parseSubtitleCues replaces generated glyph animation with its timed canonical comment', () => {
+  // Aegisub automation commonly keeps the authored lyric as a Comment and emits
+  // multiple moving Dialogue layers for every glyph. This mirrors the MyGO ED script:
+  // three entrance copies followed by three exit copies for each character.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Comment: 0,0:00:01.20,0:00:03.80,ED_JP,,0,0,0,,{\\fad(480,480)}今　手にある',
+    'Dialogue: 0,0:00:00.80,0:00:01.50,ED_JP,,0,0,0,,{\\move(10,20,100,200)\\t(0,600,\\fscx100)}今',
+    'Dialogue: 0,0:00:00.80,0:00:01.50,ED_JP,,0,0,0,,{\\move(30,40,100,200)\\t(0,600,\\fscx100)}今',
+    'Dialogue: 0,0:00:00.80,0:00:01.50,ED_JP,,0,0,0,,{\\move(50,60,100,200)\\t(0,600,\\fscx100)}今',
+    'Dialogue: 1,0:00:01.40,0:00:04.20,ED_JP,,0,0,0,,{\\move(100,200,20,30)\\t(2000,2600,\\blur20)}今',
+    'Dialogue: 1,0:00:01.40,0:00:04.20,ED_JP,,0,0,0,,{\\move(100,200,40,50)\\t(2000,2600,\\blur20)}今',
+    'Dialogue: 1,0:00:01.40,0:00:04.20,ED_JP,,0,0,0,,{\\move(100,200,60,70)\\t(2000,2600,\\blur20)}今',
+    'Dialogue: 0,0:00:00.86,0:00:01.56,ED_JP,,0,0,0,,{\\move(10,20,140,200)\\t(0,600,\\fscx100)}手',
+    'Dialogue: 0,0:00:00.86,0:00:01.56,ED_JP,,0,0,0,,{\\move(30,40,140,200)\\t(0,600,\\fscx100)}手',
+    'Dialogue: 0,0:00:00.86,0:00:01.56,ED_JP,,0,0,0,,{\\move(50,60,140,200)\\t(0,600,\\fscx100)}手',
+    'Dialogue: 1,0:00:01.46,0:00:04.26,ED_JP,,0,0,0,,{\\move(140,200,20,30)\\t(2000,2600,\\blur20)}手',
+    'Dialogue: 1,0:00:01.46,0:00:04.26,ED_JP,,0,0,0,,{\\move(140,200,40,50)\\t(2000,2600,\\blur20)}手',
+    'Dialogue: 1,0:00:01.46,0:00:04.26,ED_JP,,0,0,0,,{\\move(140,200,60,70)\\t(2000,2600,\\blur20)}手',
+    'Dialogue: 0,0:00:00.92,0:00:01.62,ED_JP,,0,0,0,,{\\move(10,20,180,200)\\t(0,600,\\fscx100)}にある',
+    'Dialogue: 0,0:00:00.92,0:00:01.62,ED_JP,,0,0,0,,{\\move(30,40,180,200)\\t(0,600,\\fscx100)}にある',
+    'Dialogue: 0,0:00:00.92,0:00:01.62,ED_JP,,0,0,0,,{\\move(50,60,180,200)\\t(0,600,\\fscx100)}にある',
+    'Dialogue: 1,0:00:01.52,0:00:04.32,ED_JP,,0,0,0,,{\\move(180,200,20,30)\\t(2000,2600,\\blur20)}にある',
+    'Dialogue: 1,0:00:01.52,0:00:04.32,ED_JP,,0,0,0,,{\\move(180,200,40,50)\\t(2000,2600,\\blur20)}にある',
+    'Dialogue: 1,0:00:01.52,0:00:04.32,ED_JP,,0,0,0,,{\\move(180,200,60,70)\\t(2000,2600,\\blur20)}にある',
+    'Dialogue: 0,0:00:06.00,0:00:08.00,Dial_JP,,0,0,0,,普通の会話',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.deepEqual(cues, [
+    {
+      startTime: 1.2,
+      endTime: 3.8,
+      text: '今　手にある',
+      source: 'canonical-ass',
+      // Entrance frames start before and exit frames end after the authored timing.
+      animationStartTime: 0.8,
+      animationEndTime: 4.32,
+    },
+    { startTime: 6, endTime: 8, text: '普通の会話' },
+  ]);
+});
+
+test('parseSubtitleCues recovers a full Dialogue line surrounding generated fragments', () => {
+  // Some scripts do not retain the authored line as a Comment. Instead, brief entrance
+  // and exit events contain the complete line around a long run of generated syllables.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:00:01.00,0:00:01.15,ED Romaji,,0,0,0,fx,{\\move(100,40,60,40)}toki yo ugokidase',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(0,300,\\c&HFFFFFF&)}to',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(300,500,\\c&HFFFFFF&)}ki',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(500,700,\\c&HFFFFFF&)}yo',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(700,900,\\c&HFFFFFF&)}u',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(900,1100,\\c&HFFFFFF&)}go',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(1100,1300,\\c&HFFFFFF&)}ki',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(1300,1500,\\c&HFFFFFF&)}da',
+    'Dialogue: 1,0:00:01.15,0:00:03.00,ED Romaji,,0,0,0,fx,{\\t(1500,1800,\\c&HFFFFFF&)}se',
+    'Dialogue: 1,0:00:03.00,0:00:03.15,ED Romaji,,0,0,0,fx,{\\move(60,40,20,40)}toki yo ugokidase',
+    'Dialogue: 0,0:00:06.00,0:00:08.00,Default,,0,0,0,,Ordinary dialogue',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    {
+      startTime: 1,
+      endTime: 3.15,
+      text: 'toki yo ugokidase',
+      source: 'canonical-ass',
+      animationStartTime: 1,
+      animationEndTime: 3.15,
+    },
+    { startTime: 6, endTime: 8, text: 'Ordinary dialogue' },
+  ]);
+});
+
+test('parseSubtitleCues does not promote a short animated fragment as a complete line', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:00:01.00,0:00:03.00,OP English,,0,0,0,,{\\pos(100,100)\\t(0,100,\\fscx120)}my',
+    'Dialogue: 2,0:00:01.00,0:00:03.00,OP English,,0,0,0,,{\\pos(100,100)\\t(0,100,\\fscx120)}my',
+    'Dialogue: 1,0:00:01.00,0:00:03.00,OP English,,0,0,0,,{\\pos(100,100)\\t(0,100,\\fscx120)}m',
+    'Dialogue: 2,0:00:01.00,0:00:03.00,OP English,,0,0,0,,{\\pos(100,100)\\t(0,100,\\fscx120)}m',
+    'Dialogue: 1,0:00:01.00,0:00:03.00,OP English,,0,0,0,,{\\pos(120,100)\\t(20,120,\\fscx120)}y',
+    'Dialogue: 2,0:00:01.00,0:00:03.00,OP English,,0,0,0,,{\\pos(120,100)\\t(20,120,\\fscx120)}y',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.equal(
+    cues.some((cue) => cue.source === 'canonical-ass'),
+    false,
+  );
+});
+
+test('parseSubtitleCues ignores timed comments without a matching animated dialogue cluster', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Comment: 0,0:00:01.00,0:00:03.00,Dial_JP,,0,0,0,,編集メモ',
+    'Comment: 0,0:00:04.00,0:00:06.00,Dial_JP,,0,0,0,,別案の字幕',
+    'Dialogue: 0,0:00:01.00,0:00:03.00,Dial_JP,,0,0,0,,通常の字幕',
+    'Dialogue: 0,0:00:04.00,0:00:06.00,Dial_JP,,0,0,0,,別案の字幕',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.deepEqual(cues, [
+    { startTime: 1, endTime: 3, text: '通常の字幕' },
+    { startTime: 4, endTime: 6, text: '別案の字幕' },
+  ]);
+});
+
+test('parseAssCues returns recovered canonical cues in chronological order', () => {
+  // Recovery appends recovered cues after surviving dialogue; the bare parseAssCues
+  // export must still come back time-ordered.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 0,0:00:06.00,0:00:08.00,Dial,,0,0,0,,あとのセリフ',
+    'Comment: 0,0:00:01.20,0:00:03.80,OP,,0,0,0,,雨が上がっても',
+    'Dialogue: 0,0:00:01.20,0:00:01.24,OP,,0,0,0,,{\\pos(1,1)\\clip(m 1 1)}雨が上がっても',
+    'Dialogue: 0,0:00:01.24,0:00:01.28,OP,,0,0,0,,{\\pos(1,1)\\clip(m 2 2)}雨が上がっても',
+    'Dialogue: 0,0:00:01.28,0:00:03.80,OP,,0,0,0,,{\\pos(1,1)\\clip(m 3 3)}雨が上がっても',
+  ].join('\n');
+
+  assert.deepEqual(
+    parseAssCues(content).map((cue) => cue.startTime),
+    [1.2, 6],
+  );
+});
+
+test('parseSubtitleCues withdraws a recovery whose owner is claimed by a later candidate', () => {
+  // The exit boundary event appears first in the file and recovers a canonical cue from
+  // its own small cluster. The entrance candidate then proves that exit event was a
+  // generated frame of the full animation; the earlier recovery is a duplicate of the
+  // same authored line and must not survive alongside it.
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Dialogue: 1,0:00:14.00,0:00:14.20,ED,,0,0,0,,{\\move(100,200,20,30)}ABCDEFGH',
+    'Dialogue: 1,0:00:13.50,0:00:14.50,ED,,0,0,0,,{\\t(0,300,\\c&HFFFFFF&)}ABC',
+    'Dialogue: 1,0:00:13.50,0:00:14.50,ED,,0,0,0,,{\\t(300,600,\\c&HFFFFFF&)}DEF',
+    'Dialogue: 1,0:00:13.50,0:00:14.50,ED,,0,0,0,,{\\t(600,900,\\c&HFFFFFF&)}GH',
+    'Dialogue: 0,0:00:10.00,0:00:10.20,ED,,0,0,0,,{\\move(10,20,100,200)}ABCDEFGH',
+    'Dialogue: 0,0:00:10.00,0:00:12.00,ED,,0,0,0,,{\\t(0,300,\\fscx100)}ABC',
+    'Dialogue: 0,0:00:10.00,0:00:12.00,ED,,0,0,0,,{\\t(300,600,\\fscx100)}DEF',
+    'Dialogue: 0,0:00:10.00,0:00:13.40,ED,,0,0,0,,{\\t(600,900,\\fscx100)}GH',
+  ].join('\n');
+
+  assert.deepEqual(parseSubtitleCues(content, 'test.ass'), [
+    {
+      startTime: 10,
+      endTime: 14.2,
+      text: 'ABCDEFGH',
+      source: 'canonical-ass',
+      animationStartTime: 10,
+      animationEndTime: 14.2,
+    },
+  ]);
+});
+
+test('parseSubtitleCues recovers canonical comments from generated clip frames', () => {
+  const content = [
+    '[Events]',
+    'Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text',
+    'Comment: 0,0:00:01.00,0:00:03.00,OP_JP,,0,0,0,,雨が上がっても',
+    'Dialogue: 0,0:00:01.00,0:00:01.04,OP_JP,,0,0,0,,{\\pos(960,1068)\\clip(m 1 1)}雨が上がっても',
+    'Dialogue: 0,0:00:01.04,0:00:01.08,OP_JP,,0,0,0,,{\\pos(960,1068)\\clip(m 2 2)}雨が上がっても',
+    'Dialogue: 0,0:00:01.08,0:00:03.00,OP_JP,,0,0,0,,{\\pos(960,1068)\\clip(m 3 3)}雨が上がっても',
+  ].join('\n');
+
+  const cues = parseSubtitleCues(content, 'test.ass');
+
+  assert.deepEqual(cues, [
+    {
+      startTime: 1,
+      endTime: 3,
+      text: '雨が上がっても',
+      source: 'canonical-ass',
+      animationStartTime: 1,
+      animationEndTime: 3,
+    },
+  ]);
+});
+
 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', ''];

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

@@ -6,12 +6,21 @@ import {
   type AssEffectKind,
   type AssOverrideCommand,
 } from './ass-text';
-import { mergeDuplicateCues } from './subtitle-cue-dedup';
+import { hasAssAnimationEvidence, mergeDuplicateCues } from './subtitle-cue-dedup';
 
 export interface SubtitleCue {
   startTime: number;
   endTime: number;
   text: string;
+  /** A complete authored line recovered from matching generated ASS animation events. */
+  source?: 'canonical-ass';
+  /**
+   * Full span of the generated animation events a canonical cue replaced. Entrance and
+   * exit frames routinely run past the authored `startTime`/`endTime`, so live-text
+   * matching must use this envelope while display and history keep the authored timing.
+   */
+  animationStartTime?: number;
+  animationEndTime?: number;
 }
 
 /**
@@ -19,7 +28,8 @@ export interface SubtitleCue {
  * 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}`.
+ * outside the parser, so the public API exposes only timing, text, and the optional
+ * canonical-source marker used by live subtitle consumers.
  */
 export interface AnnotatedSubtitleCue extends SubtitleCue {
   /** Text exactly as authored, override blocks and all. */
@@ -70,7 +80,11 @@ function sanitizeSubtitleCueText(text: string): string {
 }
 
 function toPublicCues(cues: AnnotatedSubtitleCue[]): SubtitleCue[] {
-  return cues.map(({ startTime, endTime, text }) => ({ startTime, endTime, text }));
+  return cues.map(({ startTime, endTime, text, source, animationStartTime, animationEndTime }) =>
+    source
+      ? { startTime, endTime, text, source, animationStartTime, animationEndTime }
+      : { startTime, endTime, text },
+  );
 }
 
 function parseAnnotatedSrtCues(content: string): AnnotatedSubtitleCue[] {
@@ -138,7 +152,13 @@ export function parseSrtCues(content: string): SubtitleCue[] {
 const ASS_TIMING_PATTERN = /^(\d+):(\d{2}):(\d{2})\.(\d{1,2})$/;
 const ASS_FORMAT_PREFIX = 'Format:';
 const ASS_DIALOGUE_PREFIX = 'Dialogue:';
+const ASS_COMMENT_PREFIX = 'Comment:';
 const ASS_NAME_FIELD_ALIASES = ['name', 'actor'];
+const CANONICAL_MATCH_MARGIN_SECONDS = 1;
+const MIN_CANONICAL_ANIMATION_EVENTS = 3;
+// A tiny animated fragment can itself be composed from still smaller glyph events. It is
+// not enough evidence that the fragment represents an authored line boundary.
+const MIN_CANONICAL_DIALOGUE_TEXT_LENGTH = 4;
 
 function parseAssTimestamp(raw: string): number | null {
   const match = ASS_TIMING_PATTERN.exec(raw.trim());
@@ -166,10 +186,333 @@ function findFieldIndex(formatFields: string[], aliases: string[]): number {
   return -1;
 }
 
-function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
+interface ParsedAssEvents {
+  dialogue: AnnotatedSubtitleCue[];
+  comments: AnnotatedSubtitleCue[];
+}
+
+// Every candidate line re-reads the compacted text of each event in its window, so on
+// fragment-heavy scripts the same event compacts thousands of times without this cache.
+const compactMatchTextCache = new WeakMap<AnnotatedSubtitleCue, string>();
+
+function compactAssMatchText(text: string): string {
+  return text.replace(/\s+/gu, '');
+}
+
+function compactCueMatchText(cue: AnnotatedSubtitleCue): string {
+  let compact = compactMatchTextCache.get(cue);
+  if (compact === undefined) {
+    compact = compactAssMatchText(cue.text);
+    compactMatchTextCache.set(cue, compact);
+  }
+  return compact;
+}
+
+function assEventGroupKey(cue: AnnotatedSubtitleCue): string {
+  return `${cue.style}\0${cue.name}`;
+}
+
+/**
+ * Windowed lookup over one style/name group. Every candidate line queries its time
+ * neighborhood, and fragment-heavy scripts put thousands of candidates in one group, so
+ * a linear rescan per candidate is quadratic in practice. Events are sorted by start
+ * once; `prefixMaxEnd` lets the backward walk stop as soon as no earlier event can still
+ * reach the window.
+ */
+interface AssEventGroupIndex {
+  byStart: AnnotatedSubtitleCue[];
+  prefixMaxEnd: number[];
+}
+
+function buildAssEventGroupIndex(events: readonly AnnotatedSubtitleCue[]): AssEventGroupIndex {
+  const byStart = [...events].sort((a, b) => a.startTime - b.startTime || a.order - b.order);
+  const prefixMaxEnd: number[] = [];
+  let maxEnd = -Infinity;
+  for (const event of byStart) {
+    maxEnd = Math.max(maxEnd, event.endTime);
+    prefixMaxEnd.push(maxEnd);
+  }
+  return { byStart, prefixMaxEnd };
+}
+
+/** Group events overlapping `[startTime, endTime]`, returned in source order. */
+function eventsOverlappingWindow(
+  index: AssEventGroupIndex,
+  startTime: number,
+  endTime: number,
+): AnnotatedSubtitleCue[] {
+  const { byStart, prefixMaxEnd } = index;
+  let low = 0;
+  let high = byStart.length;
+  while (low < high) {
+    const mid = (low + high) >>> 1;
+    if (byStart[mid]!.startTime <= endTime) {
+      low = mid + 1;
+    } else {
+      high = mid;
+    }
+  }
+  const matches: AnnotatedSubtitleCue[] = [];
+  for (let i = low - 1; i >= 0 && prefixMaxEnd[i]! >= startTime; i -= 1) {
+    if (byStart[i]!.endTime >= startTime) {
+      matches.push(byStart[i]!);
+    }
+  }
+  return matches.sort((a, b) => a.order - b.order);
+}
+
+interface FragmentGroup {
+  text: string;
+  events: AnnotatedSubtitleCue[];
+}
+
+function fragmentPlacementAnchors(event: AnnotatedSubtitleCue): Set<string> {
+  const anchors = new Set<string>();
+  for (const command of event.overrides) {
+    const name = command.name.toLowerCase();
+    const args = command.args.split(',').map((value) => value.trim());
+    if (name === 'pos' && args.length >= 2) {
+      anchors.add(`pos:${args[0]},${args[1]}`);
+    } else if (name === 'move' && args.length >= 4) {
+      anchors.add(`move:${args[0]},${args[1]}`);
+      anchors.add(`move:${args[2]},${args[3]}`);
+    }
+  }
+  return anchors;
+}
+
+function isRepeatedFragmentCopy(
+  previous: AnnotatedSubtitleCue,
+  current: AnnotatedSubtitleCue,
+): boolean {
+  const previousAnchors = fragmentPlacementAnchors(previous);
+  if ([...fragmentPlacementAnchors(current)].some((anchor) => previousAnchors.has(anchor))) {
+    return true;
+  }
+  return (
+    previous.startTime === current.startTime &&
+    previous.endTime === current.endTime &&
+    previous.overrideSignature === current.overrideSignature
+  );
+}
+
+function groupConsecutiveAssFragments(events: readonly AnnotatedSubtitleCue[]): FragmentGroup[] {
+  const groups: FragmentGroup[] = [];
+  for (const event of events) {
+    const text = compactCueMatchText(event);
+    if (!text) {
+      continue;
+    }
+    const previous = groups.at(-1);
+    if (
+      previous?.text === text &&
+      previous.events.some((previousEvent) => isRepeatedFragmentCopy(previousEvent, event))
+    ) {
+      previous.events.push(event);
+    } else {
+      groups.push({ text, events: [event] });
+    }
+  }
+  return groups;
+}
+
+function findCanonicalFragmentEvents(
+  events: readonly AnnotatedSubtitleCue[],
+  canonicalText: string,
+): AnnotatedSubtitleCue[] {
+  const groups = groupConsecutiveAssFragments(events);
+  const matches = new Set<AnnotatedSubtitleCue>();
+
+  for (let start = 0; start < groups.length; start += 1) {
+    let combined = '';
+    for (let end = start; end < groups.length; end += 1) {
+      const group = groups[end]!;
+      // A complete rendered copy cannot prove that the neighboring events are its
+      // fragments. Exact full-line animation is handled separately for comments.
+      if (group.text.length >= canonicalText.length) {
+        break;
+      }
+      const next = combined + group.text;
+      if (!canonicalText.startsWith(next)) {
+        break;
+      }
+      combined = next;
+      if (combined !== canonicalText) {
+        continue;
+      }
+      for (let index = start; index <= end; index += 1) {
+        for (const event of groups[index]!.events) {
+          matches.add(event);
+        }
+      }
+      start = end;
+      break;
+    }
+  }
+
+  return [...matches];
+}
+
+function matchingAssAnimationEvents(options: {
+  candidate: AnnotatedSubtitleCue;
+  group: AssEventGroupIndex;
+  allowFullLineFrames: boolean;
+}): AnnotatedSubtitleCue[] {
+  const canonicalText = compactCueMatchText(options.candidate);
+  // The group index already restricts to the candidate's style and name.
+  const nearby = eventsOverlappingWindow(
+    options.group,
+    options.candidate.startTime - CANONICAL_MATCH_MARGIN_SECONDS,
+    options.candidate.endTime + CANONICAL_MATCH_MARGIN_SECONDS,
+  );
+  const fragments = findCanonicalFragmentEvents(nearby, canonicalText);
+  if (fragments.length >= MIN_CANONICAL_ANIMATION_EVENTS && hasAssAnimationEvidence(fragments)) {
+    return fragments;
+  }
+
+  if (!options.allowFullLineFrames) {
+    return [];
+  }
+  const fullLineFrames = nearby.filter((cue) => compactCueMatchText(cue) === canonicalText);
+  return fullLineFrames.length >= MIN_CANONICAL_ANIMATION_EVENTS &&
+    hasAssAnimationEvidence(fullLineFrames)
+    ? fullLineFrames
+    : [];
+}
+
+// Reductions rather than `Math.min(...events)`: one generated line can carry an
+// unbounded number of events, and spreading them all as arguments risks the engine's
+// argument-count limit.
+function earliestStartTime(events: readonly AnnotatedSubtitleCue[], seed = Infinity): number {
+  return events.reduce((earliest, event) => Math.min(earliest, event.startTime), seed);
+}
+
+function latestEndTime(events: readonly AnnotatedSubtitleCue[], seed = -Infinity): number {
+  return events.reduce((latest, event) => Math.max(latest, event.endTime), seed);
+}
+
+function includeCanonicalBoundaryEvents(options: {
+  candidate: AnnotatedSubtitleCue;
+  group: AssEventGroupIndex;
+  animationEvents: readonly AnnotatedSubtitleCue[];
+}): AnnotatedSubtitleCue[] {
+  const canonicalText = compactCueMatchText(options.candidate);
+  const startTime = earliestStartTime(options.animationEvents);
+  const endTime = latestEndTime(options.animationEvents);
+  return eventsOverlappingWindow(
+    options.group,
+    startTime - CANONICAL_MATCH_MARGIN_SECONDS,
+    endTime + CANONICAL_MATCH_MARGIN_SECONDS,
+  ).filter((cue) => compactCueMatchText(cue) === canonicalText);
+}
+
+function recoverCanonicalAssEvents({
+  dialogue,
+  comments,
+}: ParsedAssEvents): AnnotatedSubtitleCue[] {
+  const recovered: AnnotatedSubtitleCue[] = [];
+  const suppressed = new Set<AnnotatedSubtitleCue>();
+  // A recovery is only as good as its owning event. When a later candidate proves that
+  // an earlier candidate was itself a generated frame of its animation, the earlier
+  // recovery is a duplicate of the same authored line and must be withdrawn.
+  const recoveredByOwner = new Map<AnnotatedSubtitleCue, AnnotatedSubtitleCue>();
+  const withdrawn = new Set<AnnotatedSubtitleCue>();
+  const eventsByGroup = new Map<string, AnnotatedSubtitleCue[]>();
+  for (const cue of dialogue) {
+    const key = assEventGroupKey(cue);
+    const group = eventsByGroup.get(key);
+    if (group) {
+      group.push(cue);
+    } else {
+      eventsByGroup.set(key, [cue]);
+    }
+  }
+  const indexByGroup = new Map<string, AssEventGroupIndex>();
+  for (const [key, events] of eventsByGroup) {
+    indexByGroup.set(key, buildAssEventGroupIndex(events));
+  }
+  const emptyGroupIndex: AssEventGroupIndex = { byStart: [], prefixMaxEnd: [] };
+  const candidates = [
+    ...comments.map((cue) => ({ cue, kind: 'comment' as const })),
+    ...dialogue
+      .filter(
+        (cue) =>
+          compactCueMatchText(cue).length >= MIN_CANONICAL_DIALOGUE_TEXT_LENGTH &&
+          hasAssAnimationEvidence([cue]),
+      )
+      .sort((left, right) => right.text.length - left.text.length || left.order - right.order)
+      .map((cue) => ({ cue, kind: 'dialogue' as const })),
+  ];
+
+  for (const { cue: candidate, kind } of candidates) {
+    if (candidate.endTime <= candidate.startTime || suppressed.has(candidate)) {
+      continue;
+    }
+    const canonicalText = compactCueMatchText(candidate);
+    if (!canonicalText) {
+      continue;
+    }
+
+    const group = indexByGroup.get(assEventGroupKey(candidate)) ?? emptyGroupIndex;
+    const animationEvents = matchingAssAnimationEvents({
+      candidate,
+      group,
+      allowFullLineFrames: kind === 'comment',
+    });
+    if (animationEvents.length === 0) {
+      continue;
+    }
+
+    const boundaryEvents = includeCanonicalBoundaryEvents({
+      candidate,
+      group,
+      animationEvents,
+    });
+    const generatedEvents = [...new Set([...animationEvents, ...boundaryEvents])];
+    const animationStartTime = earliestStartTime(generatedEvents, candidate.startTime);
+    const animationEndTime = latestEndTime(generatedEvents, candidate.endTime);
+    const startTime = kind === 'comment' ? candidate.startTime : animationStartTime;
+    const endTime = kind === 'comment' ? candidate.endTime : animationEndTime;
+    const recoveredCue: AnnotatedSubtitleCue = {
+      ...candidate,
+      startTime,
+      endTime,
+      animationStartTime,
+      animationEndTime,
+      source: 'canonical-ass',
+    };
+    recovered.push(recoveredCue);
+    recoveredByOwner.set(candidate, recoveredCue);
+    for (const event of generatedEvents) {
+      suppressed.add(event);
+      if (event === candidate) {
+        continue;
+      }
+      const priorRecovery = recoveredByOwner.get(event);
+      if (priorRecovery) {
+        // No text is lost by withdrawing: a fragment claim means the withdrawn line is
+        // a contiguous piece of this candidate's text, and a boundary claim means the
+        // texts are equal, so the surviving canonical cue always contains it.
+        withdrawn.add(priorRecovery);
+      }
+    }
+  }
+
+  const survivingRecovered = recovered.filter((cue) => !withdrawn.has(cue));
+  if (survivingRecovered.length === 0) {
+    return dialogue;
+  }
+  return [...dialogue.filter((cue) => !suppressed.has(cue)), ...survivingRecovered].sort(
+    (a, b) => a.startTime - b.startTime || a.endTime - b.endTime || a.order - b.order,
+  );
+}
+
+function parseAnnotatedAssEvents(content: string): ParsedAssEvents {
   const cues: AnnotatedSubtitleCue[] = [];
+  const comments: AnnotatedSubtitleCue[] = [];
   const lines = content.split(/\r?\n/);
   let inEventsSection = false;
+  let eventOrder = 0;
   const fieldIndex = {
     start: -1,
     end: -1,
@@ -222,7 +565,12 @@ function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
       continue;
     }
 
-    if (!trimmed.startsWith(ASS_DIALOGUE_PREFIX)) {
+    const eventPrefix = trimmed.startsWith(ASS_DIALOGUE_PREFIX)
+      ? ASS_DIALOGUE_PREFIX
+      : trimmed.startsWith(ASS_COMMENT_PREFIX)
+        ? ASS_COMMENT_PREFIX
+        : null;
+    if (!eventPrefix) {
       continue;
     }
 
@@ -230,7 +578,7 @@ function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
       continue;
     }
 
-    const fields = trimmed.slice(ASS_DIALOGUE_PREFIX.length).split(',');
+    const fields = trimmed.slice(eventPrefix.length).split(',');
     if (
       fieldIndex.start >= fields.length ||
       fieldIndex.end >= fields.length ||
@@ -254,7 +602,7 @@ function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
     const effect = readField(fields, fieldIndex.effect);
     const layer = Number(readField(fields, fieldIndex.layer));
     const overrides = collectAssOverrideCommands(rawText);
-    cues.push({
+    const cue: AnnotatedSubtitleCue = {
       startTime,
       endTime,
       text,
@@ -266,11 +614,21 @@ function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
       effectKind: parseAssEffectField(effect),
       overrides,
       overrideSignature: assOverrideSignature(overrides),
-      order: cues.length,
-    });
+      order: eventOrder,
+    };
+    eventOrder += 1;
+    if (eventPrefix === ASS_COMMENT_PREFIX) {
+      comments.push(cue);
+    } else {
+      cues.push(cue);
+    }
   }
 
-  return cues;
+  return { dialogue: cues, comments };
+}
+
+function parseAnnotatedAssCues(content: string): AnnotatedSubtitleCue[] {
+  return recoverCanonicalAssEvents(parseAnnotatedAssEvents(content));
 }
 
 export function parseAssCues(content: string): SubtitleCue[] {

src/main.ts

@@ -235,6 +235,7 @@ import {
   createCycleSecondarySubModeRuntimeHandler,
 } from './main/runtime/domains/mpv';
 import { buildSubtitleTrackDiagnostics } from './main/runtime/mpv-track-diagnostics';
+import { resolveCanonicalPrimarySubtitle } from './main/runtime/primary-subtitle-text';
 import {
   createBuildCopyCurrentSubtitleMainDepsHandler,
   createBuildHandleMineSentenceDigitMainDepsHandler,
@@ -1806,10 +1807,42 @@ async function openYoutubeTrackPickerFromPlayback(): Promise<void> {
 let appTray: Tray | null = null;
 let tokenizeSubtitleDeferred: ((text: string) => Promise<SubtitleData>) | null = null;
 function withCurrentSubtitleTiming(payload: SubtitleData): SubtitleData {
+  const canonical = resolveCanonicalPrimarySubtitle({
+    liveText: payload.text,
+    currentTimeSec: Number(appState.mpvClient?.currentTimePos),
+    cues: appState.activeParsedSubtitleCues,
+  });
   return {
     ...payload,
-    startTime: appState.mpvClient?.currentSubStart ?? null,
-    endTime: appState.mpvClient?.currentSubEnd ?? null,
+    startTime: canonical?.startTime ?? appState.mpvClient?.currentSubStart ?? null,
+    endTime: canonical?.endTime ?? appState.mpvClient?.currentSubEnd ?? null,
+  };
+}
+
+function captureCurrentPrimarySubtitleMiningContext(): SubtitleMiningContext | null {
+  const canonical = resolveCanonicalPrimarySubtitle({
+    liveText: appState.mpvClient?.currentSubText ?? '',
+    currentTimeSec: Number(appState.mpvClient?.currentTimePos),
+    cues: appState.activeParsedSubtitleCues,
+  });
+  // Same validity bar as the live capture path: an unusable canonical span must fall
+  // back rather than hand mining an empty line or an inverted range.
+  const canonicalText = canonical?.text.trim();
+  if (
+    !canonical ||
+    !canonicalText ||
+    !Number.isFinite(canonical.startTime) ||
+    !Number.isFinite(canonical.endTime) ||
+    canonical.endTime <= canonical.startTime
+  ) {
+    return captureLiveSubtitleMiningContext(appState.mpvClient);
+  }
+  return {
+    source: 'overlay',
+    text: canonicalText,
+    startTime: canonical.startTime,
+    endTime: canonical.endTime,
+    capturedAtMs: Date.now(),
   };
 }
 function emitSubtitlePayload(payload: SubtitleData, options?: { resumePrefetch?: boolean }): void {
@@ -5231,6 +5264,7 @@ const markLastCardAsAudioCardHandler = createMarkLastCardAsAudioCardHandler(
 const buildMineSentenceCardMainDepsHandler = createBuildMineSentenceCardMainDepsHandler({
   getAnkiIntegration: () => appState.ankiIntegration,
   getMpvClient: () => appState.mpvClient,
+  getPrimarySubtitle: () => captureCurrentPrimarySubtitleMiningContext(),
   showMpvOsd: (text) => overlayNotificationsRuntime.showConfiguredStatusNotification(text),
   mineSentenceCardCore,
   recordCardsMined: (count, noteIds) => {
@@ -5544,9 +5578,7 @@ const { registerIpcRuntimeHandlers } = composeIpcRuntimeHandlers({
       // live mpv sub timings at lookup time so media generation clips the mined line even
       // when extraction finishes long after playback has moved on.
       recordSubtitleMiningContext: (context) =>
-        recordSubtitleMiningContext(
-          context ?? captureLiveSubtitleMiningContext(appState.mpvClient),
-        ),
+        recordSubtitleMiningContext(context ?? captureCurrentPrimarySubtitleMiningContext()),
       quitApp: () => requestAppQuit(),
       toggleVisibleOverlay: () => toggleVisibleOverlay(),
       tokenizeCurrentSubtitle: async () => {

src/main/runtime/anki-actions-main-deps.test.ts

@@ -64,11 +64,17 @@ test('anki action main deps builders map callbacks', async () => {
   const mine = createBuildMineSentenceCardMainDepsHandler({
     getAnkiIntegration: () => ({ enabled: true }),
     getMpvClient: () => ({ connected: true }),
+    getPrimarySubtitle: () => ({ text: '正式な字幕', startTime: 1, endTime: 3 }),
     showMpvOsd: (text) => calls.push(`mine:${text}`),
     mineSentenceCardCore: async () => true,
     recordCardsMined: (count) => calls.push(`cards:${count}`),
   })();
   assert.deepEqual(mine.getMpvClient(), { connected: true });
+  assert.deepEqual(mine.getPrimarySubtitle?.(), {
+    text: '正式な字幕',
+    startTime: 1,
+    endTime: 3,
+  });
   mine.showMpvOsd('m');
   await mine.mineSentenceCardCore({
     ankiIntegration: { enabled: true },

src/main/runtime/anki-actions-main-deps.ts

@@ -1,4 +1,4 @@
-import type { createRefreshKnownWordCacheHandler } from './anki-actions';
+import type { createRefreshKnownWordCacheHandler, PrimarySubtitle } from './anki-actions';
 
 type RefreshKnownWordCacheMainDeps = Parameters<typeof createRefreshKnownWordCacheHandler>[0];
 
@@ -72,10 +72,12 @@ export function createBuildMarkLastCardAsAudioCardMainDepsHandler<TAnki>(deps: {
 export function createBuildMineSentenceCardMainDepsHandler<TAnki, TMpv>(deps: {
   getAnkiIntegration: () => TAnki;
   getMpvClient: () => TMpv;
+  getPrimarySubtitle?: () => PrimarySubtitle | null;
   showMpvOsd: (text: string) => void;
   mineSentenceCardCore: (options: {
     ankiIntegration: TAnki;
     mpvClient: TMpv;
+    primarySubtitle?: PrimarySubtitle;
     showMpvOsd: (text: string) => void;
   }) => Promise<boolean>;
   recordCardsMined: (count: number, noteIds?: number[]) => void;
@@ -83,10 +85,14 @@ export function createBuildMineSentenceCardMainDepsHandler<TAnki, TMpv>(deps: {
   return () => ({
     getAnkiIntegration: () => deps.getAnkiIntegration(),
     getMpvClient: () => deps.getMpvClient(),
+    ...(deps.getPrimarySubtitle
+      ? { getPrimarySubtitle: () => deps.getPrimarySubtitle?.() ?? null }
+      : {}),
     showMpvOsd: (text: string) => deps.showMpvOsd(text),
     mineSentenceCardCore: (options: {
       ankiIntegration: TAnki;
       mpvClient: TMpv;
+      primarySubtitle?: PrimarySubtitle;
       showMpvOsd: (text: string) => void;
     }) => deps.mineSentenceCardCore(options),
     recordCardsMined: (count: number, noteIds?: number[]) => deps.recordCardsMined(count, noteIds),

src/main/runtime/anki-actions.test.ts

@@ -87,3 +87,20 @@ test('mine sentence handler records mined cards only when core returns true', as
   await mineSentenceCard();
   assert.deepEqual(calls, ['osd:mine', 'osd:mine', 'cards:1']);
 });
+
+test('mine sentence handler forwards the canonical primary subtitle snapshot', async () => {
+  const primarySubtitle = { text: '正式な字幕', startTime: 1, endTime: 3 };
+  const mineSentenceCard = createMineSentenceCardHandler({
+    getAnkiIntegration: () => ({}),
+    getMpvClient: () => ({}),
+    getPrimarySubtitle: () => primarySubtitle,
+    showMpvOsd: () => {},
+    mineSentenceCardCore: async (options) => {
+      assert.equal(options.primarySubtitle, primarySubtitle);
+      return true;
+    },
+    recordCardsMined: () => {},
+  });
+
+  await mineSentenceCard();
+});

src/main/runtime/anki-actions.ts

@@ -2,6 +2,12 @@ type AnkiIntegrationLike = {
   refreshKnownWordCache: () => Promise<void>;
 };
 
+export type PrimarySubtitle = {
+  text: string;
+  startTime: number;
+  endTime: number;
+};
+
 export function createUpdateLastCardFromClipboardHandler<TAnki>(deps: {
   getAnkiIntegration: () => TAnki;
   readClipboardText: () => string;
@@ -69,18 +75,22 @@ export function createMarkLastCardAsAudioCardHandler<TAnki>(deps: {
 export function createMineSentenceCardHandler<TAnki, TMpv>(deps: {
   getAnkiIntegration: () => TAnki;
   getMpvClient: () => TMpv;
+  getPrimarySubtitle?: () => PrimarySubtitle | null;
   showMpvOsd: (text: string) => void;
   mineSentenceCardCore: (options: {
     ankiIntegration: TAnki;
     mpvClient: TMpv;
+    primarySubtitle?: PrimarySubtitle;
     showMpvOsd: (text: string) => void;
   }) => Promise<boolean>;
   recordCardsMined: (count: number, noteIds?: number[]) => void;
 }) {
   return async (): Promise<void> => {
+    const primarySubtitle = deps.getPrimarySubtitle?.();
     const created = await deps.mineSentenceCardCore({
       ankiIntegration: deps.getAnkiIntegration(),
       mpvClient: deps.getMpvClient(),
+      ...(primarySubtitle ? { primarySubtitle } : {}),
       showMpvOsd: deps.showMpvOsd,
     });
     if (created) {

src/main/runtime/autoplay-subtitle-priming-runtime.ts

@@ -1,6 +1,7 @@
 import type { SubtitleCue, SubtitleData } from '../../types';
 import { selectAutoplayStartupCue } from './autoplay-subtitle-primer';
 import { primeVisibleOverlaySubtitleFromMpv } from './current-subtitle-snapshot';
+import { resolvePrimarySubtitleText } from './primary-subtitle-text';
 import { resolveSubtitleSourcePath } from './subtitle-prefetch-source';
 
 const AUTOPLAY_SUBTITLE_PRIME_LOOKAHEAD_SECONDS = 2;
@@ -141,6 +142,16 @@ export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimi
     return true;
   }
 
+  function resolveLivePrimarySubtitleText(text: string): string {
+    const client = deps.getMpvClient();
+    const currentTimeSec = Number(client?.currentTimePos ?? deps.getLastObservedTimePos());
+    return resolvePrimarySubtitleText({
+      liveText: text,
+      currentTimeSec,
+      cues: deps.getActiveParsedSubtitleCues(),
+    });
+  }
+
   async function primeCurrentSubtitleForAutoplay(mediaPath: string): Promise<void> {
     const client = deps.getMpvClient();
     if (!client?.connected || !isCurrentAutoplayMediaPath(mediaPath)) {
@@ -155,7 +166,8 @@ export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimi
       );
       return null;
     });
-    const text = typeof subTextRaw === 'string' ? subTextRaw : '';
+    const liveText = typeof subTextRaw === 'string' ? subTextRaw : '';
+    const text = resolveLivePrimarySubtitleText(liveText);
     if (emitAutoplayPrimedSubtitle(mediaPath, text)) {
       return;
     }
@@ -175,6 +187,7 @@ export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimi
   async function primeCurrentSubtitleForVisibleOverlay(): Promise<void> {
     await primeVisibleOverlaySubtitleFromMpv({
       getMpvClient: () => deps.getMpvClient(),
+      resolvePrimarySubtitleText: (text) => resolveLivePrimarySubtitleText(text),
       setCurrentSubText: (text) => {
         deps.setCurrentSubText(text);
       },

src/main/runtime/current-subtitle-snapshot.ts

@@ -46,6 +46,7 @@ export async function resolveCurrentSubtitleForRenderer(deps: {
 export async function primeVisibleOverlaySubtitleFromMpv(deps: {
   getMpvClient: () => CurrentSubtitleMpvClient | null;
   setCurrentSubText: (text: string) => void;
+  resolvePrimarySubtitleText?: (text: string) => string;
   getCurrentSubtitleData: () => SubtitleData | null;
   consumeCachedSubtitle: (text: string) => SubtitleData | null;
   onSubtitleChange: (text: string) => void;
@@ -73,7 +74,8 @@ export async function primeVisibleOverlaySubtitleFromMpv(deps: {
     return;
   }
 
-  const text = typeof subTextRaw === 'string' ? subTextRaw : '';
+  const liveText = typeof subTextRaw === 'string' ? subTextRaw : '';
+  const text = deps.resolvePrimarySubtitleText?.(liveText) ?? liveText;
   deps.setCurrentSubText(text);
 
   const primeSecondarySubtitle = async (): Promise<void> => {

src/main/runtime/mpv-main-event-actions.test.ts

@@ -26,6 +26,29 @@ test('subtitle change handler updates state and forwards uncached text without r
   assert.deepEqual(calls, ['set:line', 'process:line', 'presence']);
 });
 
+test('subtitle change handler consistently forwards resolved canonical text', () => {
+  const calls: string[] = [];
+  const handler = createHandleMpvSubtitleChangeHandler({
+    resolveSubtitleText: () => '今　手にある物差しでは',
+    setCurrentSubText: (text) => calls.push(`set:${text}`),
+    getImmediateSubtitlePayload: (text) => {
+      calls.push(`lookup:${text}`);
+      return null;
+    },
+    broadcastSubtitle: () => {},
+    onSubtitleChange: (text) => calls.push(`process:${text}`),
+    refreshDiscordPresence: () => {},
+  });
+
+  handler({ text: '今今今手手手ににに' });
+
+  assert.deepEqual(calls, [
+    'set:今　手にある物差しでは',
+    'lookup:今　手にある物差しでは',
+    'process:今　手にある物差しでは',
+  ]);
+});
+
 test('subtitle change handler clears immediately for empty subtitle text', () => {
   const calls: string[] = [];
   const handler = createHandleMpvSubtitleChangeHandler({

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

@@ -4,7 +4,8 @@ type AnilistPostWatchRunOptions = {
   watchedSeconds?: number;
 };
 
-const SEEK_LIKE_TIME_DELTA_SECONDS = 2.5;
+/** Jump size that marks a time-pos change as a seek rather than normal playback. */
+export const SEEK_LIKE_TIME_DELTA_SECONDS = 2.5;
 
 function isSeekLikeTimeChange(previousTime: number | null, nextTime: number): boolean {
   if (previousTime === null || !Number.isFinite(previousTime) || !Number.isFinite(nextTime)) {
@@ -14,6 +15,7 @@ function isSeekLikeTimeChange(previousTime: number | null, nextTime: number): bo
 }
 
 export function createHandleMpvSubtitleChangeHandler(deps: {
+  resolveSubtitleText?: (text: string) => string;
   setCurrentSubText: (text: string) => void;
   getImmediateSubtitlePayload?: (text: string) => SubtitleData | null;
   emitImmediateSubtitle?: (payload: SubtitleData) => void;
@@ -22,7 +24,8 @@ export function createHandleMpvSubtitleChangeHandler(deps: {
   refreshDiscordPresence: () => void;
   logDebug?: (message: string) => void;
 }) {
-  return ({ text }: { text: string }): void => {
+  return ({ text: liveText }: { text: string }): void => {
+    const text = deps.resolveSubtitleText?.(liveText) ?? liveText;
     deps.setCurrentSubText(text);
     const immediatePayload = deps.getImmediateSubtitlePayload?.(text) ?? null;
     if (immediatePayload) {

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

@@ -43,6 +43,7 @@ export function createBindMpvMainEventHandlersHandler(deps: {
   logSubtitleTimingError: (message: string, error: unknown) => void;
 
   setCurrentSubText: (text: string) => void;
+  resolveSubtitleText?: (text: string) => string;
   getImmediateSubtitlePayload?: (text: string) => SubtitleData | null;
   emitImmediateSubtitle?: (payload: SubtitleData) => void;
   broadcastSubtitle: (payload: SubtitleData) => void;
@@ -117,6 +118,7 @@ export function createBindMpvMainEventHandlersHandler(deps: {
       logError: (message, error) => deps.logSubtitleTimingError(message, error),
     });
     const handleMpvSubtitleChange = createHandleMpvSubtitleChangeHandler({
+      resolveSubtitleText: deps.resolveSubtitleText,
       setCurrentSubText: (text) => deps.setCurrentSubText(text),
       getImmediateSubtitlePayload: (text) => deps.getImmediateSubtitlePayload?.(text) ?? null,
       emitImmediateSubtitle: deps.emitImmediateSubtitle

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

@@ -387,3 +387,170 @@ test('subtitle-track transitions ignore stale parsed cues until replacement cues
   handlers.recordImmersionSubtitleLine('飛び上がる', 20.04, 20.08);
   assert.deepEqual(recordedStarts.slice(-1), [20]);
 });
+
+test('canonical ASS cues replace live glyph spam for display, history, and immersion', () => {
+  const immersion: Array<{ text: string; start: number; end: number }> = [];
+  const timing: Array<{ text: string; start: number; end: number }> = [];
+  const handlers = createBuildBindMpvMainEventHandlersMainDepsHandler({
+    appState: {
+      initialArgs: null,
+      overlayRuntimeInitialized: true,
+      mpvClient: { currentTimePos: 2 },
+      immersionTracker: {
+        recordSubtitleLine: (text: string, start: number, end: number) =>
+          immersion.push({ text, start, end }),
+      },
+      subtitleTimingTracker: {
+        recordSubtitle: (text: string, start: number, end: number) =>
+          timing.push({ text, start, end }),
+      },
+      activeParsedSubtitleCues: [
+        {
+          startTime: 1.2,
+          endTime: 3.8,
+          text: '今　手にある物差しでは',
+          source: 'canonical-ass',
+        },
+        {
+          startTime: 3,
+          endTime: 6,
+          text: '飛び越えてみたくて',
+          source: 'canonical-ass',
+        },
+      ],
+      currentMediaPath: '/video.mkv',
+      currentSubText: '',
+      currentSubAssText: '',
+      playbackPaused: null,
+      previousSecondarySubVisibility: false,
+    },
+    getQuitOnDisconnectArmed: () => false,
+    scheduleQuitCheck: () => {},
+    quitApp: () => {},
+    reportJellyfinRemoteStopped: () => {},
+    syncOverlayMpvSubtitleSuppression: () => {},
+    maybeRunAnilistPostWatchUpdate: async () => {},
+    logSubtitleTimingError: () => {},
+    broadcastToOverlayWindows: () => {},
+    onSubtitleChange: () => {},
+    ensureImmersionTrackerInitialized: () => {},
+    updateCurrentMediaPath: () => {},
+    restoreMpvSubVisibility: () => {},
+    getCurrentAnilistMediaKey: () => null,
+    resetAnilistMediaTracking: () => {},
+    maybeProbeAnilistDuration: () => {},
+    ensureAnilistMediaGuess: () => {},
+    syncImmersionMediaState: () => {},
+    updateCurrentMediaTitle: () => {},
+    resetAnilistMediaGuessState: () => {},
+    reportJellyfinRemoteProgress: () => {},
+    updateSubtitleRenderMetrics: () => {},
+    refreshDiscordPresence: () => {},
+  })();
+
+  assert.equal(handlers.resolveSubtitleText?.('今\n今\n今\n手\n手\n手'), '今　手にある物差しでは');
+  handlers.recordImmersionSubtitleLine('今', 0.8, 1.5);
+  handlers.recordImmersionSubtitleLine('手', 0.86, 1.56);
+  handlers.recordSubtitleTiming('今', 0.8, 1.5);
+
+  assert.deepEqual(immersion, [{ text: '今　手にある物差しでは', start: 1.2, end: 3.8 }]);
+  assert.deepEqual(timing, [{ text: '今　手にある物差しでは', start: 1.2, end: 3.8 }]);
+
+  // Concurrent dialogue during the song is not part of the animation: it must be
+  // recorded as itself -- without the fragment lines beside it -- and must not cause
+  // the song line to be recorded again when the animation frames resume.
+  assert.equal(handlers.resolveSubtitleText?.('普通のセリフ\n今\n手'), '普通のセリフ\n今\n手');
+  handlers.recordImmersionSubtitleLine('普通のセリフ\n今\n手', 1.9, 3.2);
+  handlers.recordImmersionSubtitleLine('にある', 2.1, 2.9);
+  handlers.recordSubtitleTiming('次のセリフ', 3.9, 5.0);
+
+  assert.deepEqual(immersion.slice(1), [{ text: '普通のセリフ', start: 1.9, end: 3.2 }]);
+  assert.deepEqual(timing.slice(1), [{ text: '次のセリフ', start: 3.9, end: 5 }]);
+
+  // Overlapping canonical lines resolve as shifting subsets (A, then A+B, then A).
+  // Every recorded cue is remembered, so each authored line still records exactly once.
+  handlers.recordImmersionSubtitleLine('飛び越えて', 3.2, 3.4);
+  handlers.recordImmersionSubtitleLine('手にある', 3.5, 3.7);
+  handlers.recordSubtitleTiming('飛び越えて', 3.2, 3.4);
+  handlers.recordSubtitleTiming('手にある', 3.5, 3.7);
+
+  assert.deepEqual(immersion.slice(2), [{ text: '飛び越えてみたくて', start: 3, end: 6 }]);
+  assert.deepEqual(timing.slice(2), [{ text: '飛び越えてみたくて', start: 3, end: 6 }]);
+
+  // A backward seek means the user is rewatching: the timing history (a viewing log)
+  // records the revisited line again, while immersion stays once-per-media.
+  handlers.onTimePosUpdate?.(30);
+  handlers.onTimePosUpdate?.(2);
+  handlers.recordSubtitleTiming('今', 0.8, 1.5);
+  handlers.recordImmersionSubtitleLine('今', 0.8, 1.5);
+
+  assert.deepEqual(timing.slice(3), [{ text: '今　手にある物差しでは', start: 1.2, end: 3.8 }]);
+  assert.equal(immersion.length, 3);
+
+  // A jump of exactly the seek threshold counts as a seek, matching the time-pos
+  // handler's own `>=` boundary.
+  handlers.onTimePosUpdate?.(4.5);
+  handlers.onTimePosUpdate?.(2);
+  handlers.recordSubtitleTiming('今', 0.8, 1.5);
+
+  assert.deepEqual(timing.slice(4), [{ text: '今　手にある物差しでは', start: 1.2, end: 3.8 }]);
+});
+
+test('subtitle-track changes stop stale canonical cues from substituting immediately', () => {
+  const appState = {
+    initialArgs: null,
+    overlayRuntimeInitialized: true,
+    mpvClient: { currentTimePos: 2 },
+    immersionTracker: { recordSubtitleLine: () => {} },
+    subtitleTimingTracker: { recordSubtitle: () => {} },
+    activeParsedSubtitleCues: [
+      {
+        startTime: 1.2,
+        endTime: 3.8,
+        text: '今　手にある物差しでは',
+        source: 'canonical-ass' as const,
+      },
+    ] as Array<{ startTime: number; endTime: number; text: string; source?: 'canonical-ass' }>,
+    activeParsedSubtitleSource: 'track-a.ass' as string | null,
+    currentMediaPath: '/video.mkv',
+    currentSubText: '',
+    currentSubAssText: '',
+    playbackPaused: null,
+    previousSecondarySubVisibility: false,
+  };
+  const handlers = createBuildBindMpvMainEventHandlersMainDepsHandler({
+    appState,
+    getQuitOnDisconnectArmed: () => false,
+    scheduleQuitCheck: () => {},
+    quitApp: () => {},
+    reportJellyfinRemoteStopped: () => {},
+    syncOverlayMpvSubtitleSuppression: () => {},
+    maybeRunAnilistPostWatchUpdate: async () => {},
+    logSubtitleTimingError: () => {},
+    broadcastToOverlayWindows: () => {},
+    onSubtitleChange: () => {},
+    ensureImmersionTrackerInitialized: () => {},
+    updateCurrentMediaPath: () => {},
+    restoreMpvSubVisibility: () => {},
+    getCurrentAnilistMediaKey: () => null,
+    resetAnilistMediaTracking: () => {},
+    maybeProbeAnilistDuration: () => {},
+    ensureAnilistMediaGuess: () => {},
+    syncImmersionMediaState: () => {},
+    updateCurrentMediaTitle: () => {},
+    resetAnilistMediaGuessState: () => {},
+    reportJellyfinRemoteProgress: () => {},
+    updateSubtitleRenderMetrics: () => {},
+    refreshDiscordPresence: () => {},
+  })();
+
+  assert.equal(handlers.resolveSubtitleText?.('今\n手にある'), '今　手にある物差しでは');
+
+  // The new track's cues arrive only after an async re-parse; until then, the old
+  // track's canonical lyric must not replace the new track's live text.
+  handlers.onSubtitleTrackChange?.(2);
+
+  assert.deepEqual(appState.activeParsedSubtitleCues, []);
+  assert.equal(appState.activeParsedSubtitleSource, null);
+  assert.equal(handlers.resolveSubtitleText?.('今\n手にある'), '今\n手にある');
+});

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

@@ -1,5 +1,11 @@
 import { createSubtitleLineDedupGate } from '../../core/services/subtitle-line-dedup-gate';
 import type { MergedToken, SubtitleCue, SubtitleData } from '../../types';
+import { SEEK_LIKE_TIME_DELTA_SECONDS } from './mpv-main-event-actions';
+import {
+  resolveCanonicalPrimarySubtitle,
+  resolvePrimarySubtitleText,
+  stripCanonicalFragmentLines,
+} from './primary-subtitle-text';
 
 type AnilistPostWatchRunOptions = {
   watchedSeconds?: number;
@@ -36,6 +42,8 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
       recordSubtitle?: (text: string, start: number, end: number, secondaryText?: string) => void;
     } | null;
     activeParsedSubtitleCues?: SubtitleCue[] | null;
+    /** Cache key of the source the cues were parsed from; cleared with the cues. */
+    activeParsedSubtitleSource?: string | null;
     currentMediaPath?: string | null;
     currentSubText: string;
     currentSubAssText: string;
@@ -93,6 +101,38 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
   const immersionLineDedupGate = createSubtitleLineDedupGate({
     getParsedCues: () => deps.appState.activeParsedSubtitleCues,
   });
+  // One seen-set per consumer: canonical cues overlap, so live samples resolve to
+  // shifting subsets (A, then A+B, then B). Remembering every recorded cue -- not just
+  // the previous sample -- keeps each authored line recorded exactly once per source.
+  const recordedImmersionCanonicalKeys = new Set<string>();
+  const recordedTimingCanonicalKeys = new Set<string>();
+  // Bumped on track/media changes so an immersion record whose tokenization resolves
+  // after the change is dropped instead of landing in the next session.
+  let subtitleSessionEpoch = 0;
+  let lastTimePosForTimingReset: number | null = null;
+  const canonicalCueKey = (cue: SubtitleCue): string =>
+    `${cue.startTime}|${cue.endTime}|${cue.text}`;
+  const resetSubtitleDeduplication = (): void => {
+    immersionLineDedupGate.reset();
+    recordedImmersionCanonicalKeys.clear();
+    recordedTimingCanonicalKeys.clear();
+    subtitleSessionEpoch += 1;
+    lastTimePosForTimingReset = null;
+  };
+  const resolveCanonicalSample = (liveText: string, startSec: number) =>
+    resolveCanonicalPrimarySubtitle({
+      liveText,
+      currentTimeSec: startSec,
+      cues: deps.appState.activeParsedSubtitleCues,
+    });
+  // When substitution declined because dialogue shares the screen with a song, record
+  // the dialogue alone rather than the combined dialogue-plus-fragments stack.
+  const stripFragmentsForRecording = (liveText: string, startSec: number) =>
+    stripCanonicalFragmentLines({
+      liveText,
+      currentTimeSec: startSec,
+      cues: deps.appState.activeParsedSubtitleCues,
+    });
   const hasInitialPlaybackQuitOnDisconnectArg = (): boolean =>
     Boolean(
       deps.appState.initialArgs?.managedPlayback ||
@@ -111,45 +151,99 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
     scheduleQuitCheck: (callback: () => void) => deps.scheduleQuitCheck(callback),
     isMpvConnected: () => Boolean(deps.appState.mpvClient?.connected),
     quitApp: () => deps.quitApp(),
+    resolveSubtitleText: (liveText: string) =>
+      resolvePrimarySubtitleText({
+        liveText,
+        currentTimeSec: Number(deps.appState.mpvClient?.currentTimePos),
+        cues: deps.appState.activeParsedSubtitleCues,
+      }),
     recordImmersionSubtitleLine: (text: string, start: number, end: number) => {
       deps.ensureImmersionTrackerInitialized();
       const tracker = deps.appState.immersionTracker;
       if (!tracker?.recordSubtitleLine) {
         return;
       }
+      const recordLine = (lineText: string, startSec: number, endSec: number): void => {
+        const secondaryText = deps.appState.mpvClient?.currentSecondarySubText || null;
+        const cachedTokens =
+          deps.appState.currentSubtitleData?.text === lineText
+            ? deps.appState.currentSubtitleData.tokens
+            : null;
+        if (cachedTokens) {
+          tracker.recordSubtitleLine?.(lineText, startSec, endSec, cachedTokens, secondaryText);
+          return;
+        }
+        if (!deps.tokenizeSubtitleForImmersion) {
+          tracker.recordSubtitleLine?.(lineText, startSec, endSec, null, secondaryText);
+          return;
+        }
+        const epochAtRecord = subtitleSessionEpoch;
+        void deps
+          .tokenizeSubtitleForImmersion(lineText)
+          .then((payload) => {
+            if (subtitleSessionEpoch !== epochAtRecord) {
+              return;
+            }
+            tracker.recordSubtitleLine?.(
+              lineText,
+              startSec,
+              endSec,
+              payload?.tokens ?? null,
+              secondaryText,
+            );
+          })
+          .catch(() => {
+            if (subtitleSessionEpoch !== epochAtRecord) {
+              return;
+            }
+            tracker.recordSubtitleLine?.(lineText, startSec, endSec, null, secondaryText);
+          });
+      };
+      const canonical = resolveCanonicalSample(text, start);
+      if (canonical) {
+        for (const cue of canonical.cues) {
+          const key = canonicalCueKey(cue);
+          if (recordedImmersionCanonicalKeys.has(key)) {
+            continue;
+          }
+          recordedImmersionCanonicalKeys.add(key);
+          recordLine(cue.text, cue.startTime, cue.endTime);
+        }
+        return;
+      }
+      text = stripFragmentsForRecording(text, start);
       if (!immersionLineDedupGate.shouldRecord({ text, startSec: start, endSec: end })) {
         return;
       }
-      const secondaryText = deps.appState.mpvClient?.currentSecondarySubText || null;
-      const cachedTokens =
-        deps.appState.currentSubtitleData?.text === text
-          ? deps.appState.currentSubtitleData.tokens
-          : null;
-      if (cachedTokens) {
-        tracker.recordSubtitleLine(text, start, end, cachedTokens, secondaryText);
-        return;
-      }
-      if (!deps.tokenizeSubtitleForImmersion) {
-        tracker.recordSubtitleLine(text, start, end, null, secondaryText);
-        return;
-      }
-      void deps
-        .tokenizeSubtitleForImmersion(text)
-        .then((payload) => {
-          tracker.recordSubtitleLine?.(text, start, end, payload?.tokens ?? null, secondaryText);
-        })
-        .catch(() => {
-          tracker.recordSubtitleLine?.(text, start, end, null, secondaryText);
-        });
+      recordLine(text, start, end);
     },
     hasSubtitleTimingTracker: () => Boolean(deps.appState.subtitleTimingTracker),
-    recordSubtitleTiming: (text: string, start: number, end: number) =>
-      deps.appState.subtitleTimingTracker?.recordSubtitle?.(
-        text,
-        start,
-        end,
-        deps.appState.mpvClient?.currentSecondarySubText || undefined,
-      ),
+    recordSubtitleTiming: (text: string, start: number, end: number) => {
+      const secondaryText = deps.appState.mpvClient?.currentSecondarySubText || undefined;
+      const canonical = resolveCanonicalSample(text, start);
+      if (!canonical) {
+        deps.appState.subtitleTimingTracker?.recordSubtitle?.(
+          stripFragmentsForRecording(text, start),
+          start,
+          end,
+          secondaryText,
+        );
+        return;
+      }
+      for (const cue of canonical.cues) {
+        const key = canonicalCueKey(cue);
+        if (recordedTimingCanonicalKeys.has(key)) {
+          continue;
+        }
+        recordedTimingCanonicalKeys.add(key);
+        deps.appState.subtitleTimingTracker?.recordSubtitle?.(
+          cue.text,
+          cue.startTime,
+          cue.endTime,
+          secondaryText,
+        );
+      }
+    },
     maybeRunAnilistPostWatchUpdate: (options?: AnilistPostWatchRunOptions) =>
       deps.maybeRunAnilistPostWatchUpdate(options),
     logSubtitleTimingError: (message: string, error: unknown) =>
@@ -170,7 +264,14 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
       ? (message: string) => deps.logSubtitleProcessingDebug!(message)
       : undefined,
     onSubtitleTrackChange: (sid: number | null) => {
-      immersionLineDedupGate.reset();
+      resetSubtitleDeduplication();
+      // The replacement track's cues arrive only after an async re-read and re-parse.
+      // Clearing synchronously keeps the previous track's canonical cues from
+      // substituting into, or recording against, the new track's live text. The source
+      // key is cleared with the cues so cue-list consumers (the sidebar snapshot)
+      // re-parse on demand instead of trusting the stale pairing.
+      deps.appState.activeParsedSubtitleCues = [];
+      deps.appState.activeParsedSubtitleSource = null;
       deps.onSubtitleTrackChange?.(sid);
     },
     onSubtitleTrackListChange: deps.onSubtitleTrackListChange
@@ -185,7 +286,7 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
     broadcastSecondarySubtitle: (text: string) =>
       deps.broadcastToOverlayWindows('secondary-subtitle:set', text),
     updateCurrentMediaPath: (path: string) => {
-      immersionLineDedupGate.reset();
+      resetSubtitleDeduplication();
       deps.updateCurrentMediaPath(path);
     },
     restoreMpvSubVisibility: () => deps.restoreMpvSubVisibility(),
@@ -217,9 +318,22 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
     },
     reportJellyfinRemoteProgress: (forceImmediate: boolean) =>
       deps.reportJellyfinRemoteProgress(forceImmediate),
-    onTimePosUpdate: deps.onTimePosUpdate
-      ? (time: number) => deps.onTimePosUpdate!(time)
-      : undefined,
+    onTimePosUpdate: (time: number) => {
+      // Timing history is a viewing log: after a real backward seek, a rewatched
+      // canonical line should enter it again. Immersion stats keep their
+      // once-per-media deduplication and are not reset here.
+      if (
+        Number.isFinite(time) &&
+        lastTimePosForTimingReset !== null &&
+        time <= lastTimePosForTimingReset - SEEK_LIKE_TIME_DELTA_SECONDS
+      ) {
+        recordedTimingCanonicalKeys.clear();
+      }
+      if (Number.isFinite(time)) {
+        lastTimePosForTimingReset = time;
+      }
+      deps.onTimePosUpdate?.(time);
+    },
     onFullscreenChange: deps.onFullscreenChange
       ? (fullscreen: boolean) => deps.onFullscreenChange!(fullscreen)
       : undefined,

src/main/runtime/primary-subtitle-text.test.ts

@@ -0,0 +1,277 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import {
+  resolveCanonicalPrimarySubtitle,
+  resolvePrimarySubtitleText,
+  stripCanonicalFragmentLines,
+} from './primary-subtitle-text';
+
+test('resolvePrimarySubtitleText prefers an active canonical cue over flattened mpv glyphs', () => {
+  // mpv renders each simultaneously active ASS event on its own sub-text line.
+  const text = resolvePrimarySubtitleText({
+    liveText: '今\n今\n今\n手\n手\n手\nにある\nにある\nにある',
+    currentTimeSec: 2,
+    cues: [
+      {
+        startTime: 1.2,
+        endTime: 3.8,
+        text: '今　手にある',
+        source: 'canonical-ass',
+      },
+    ],
+  });
+
+  assert.equal(text, '今　手にある');
+});
+
+test('resolvePrimarySubtitleText preserves live text outside canonical cue timing', () => {
+  const text = resolvePrimarySubtitleText({
+    liveText: '通常の会話',
+    currentTimeSec: 8,
+    cues: [
+      {
+        startTime: 1.2,
+        endTime: 3.8,
+        text: '今　手にある',
+        source: 'canonical-ass',
+      },
+    ],
+  });
+
+  assert.equal(text, '通常の会話');
+});
+
+test('resolvePrimarySubtitleText keeps concurrent dialogue that is not part of the animation', () => {
+  // An insert song's canonical window can overlap real dialogue on the same track.
+  const text = resolvePrimarySubtitleText({
+    liveText: '普通のセリフ\n今\n手にある',
+    currentTimeSec: 2,
+    cues: [
+      {
+        startTime: 1.2,
+        endTime: 3.8,
+        text: '今　手にある',
+        source: 'canonical-ass',
+      },
+    ],
+  });
+
+  assert.equal(text, '普通のセリフ\n今\n手にある');
+});
+
+test('resolvePrimarySubtitleText keeps a fresh line starting just after the animation ended', () => {
+  const text = resolvePrimarySubtitleText({
+    liveText: '次のセリフ',
+    currentTimeSec: 4.1,
+    cues: [
+      {
+        startTime: 1.2,
+        endTime: 3.8,
+        text: '今　手にある',
+        source: 'canonical-ass',
+      },
+    ],
+  });
+
+  assert.equal(text, '次のセリフ');
+});
+
+test('resolvePrimarySubtitleText survives overlapping frames of consecutive karaoke lines', () => {
+  // Near a line boundary the previous line's exit frames and the next line's entrance
+  // frames render together; neither line alone explains every live segment.
+  const cues = [
+    { startTime: 1.2, endTime: 3.8, text: '今　手にある', source: 'canonical-ass' as const },
+    { startTime: 3.8, endTime: 6.4, text: '物差しでは', source: 'canonical-ass' as const },
+  ];
+
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: '手にある\n物差し\nでは',
+      currentTimeSec: 3.6,
+      cues,
+    }),
+    '今　手にある',
+  );
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: '手にある\n物差し\nでは',
+      currentTimeSec: 3.9,
+      cues,
+    }),
+    '物差しでは',
+  );
+});
+
+test('resolvePrimarySubtitleText combines simultaneous canonical cues in source order', () => {
+  const text = resolvePrimarySubtitleText({
+    liveText: 'fir\nst\nsecond',
+    currentTimeSec: 2,
+    cues: [
+      { startTime: 1, endTime: 3, text: 'first', source: 'canonical-ass' },
+      { startTime: 1.5, endTime: 2.5, text: 'second', source: 'canonical-ass' },
+    ],
+  });
+
+  assert.equal(text, 'first\nsecond');
+});
+
+test('resolveCanonicalPrimarySubtitle covers a nearby generated animation edge', () => {
+  const cue = {
+    startTime: 1.2,
+    endTime: 3.8,
+    text: '今　手にある',
+    source: 'canonical-ass' as const,
+  };
+  const resolved = resolveCanonicalPrimarySubtitle({
+    liveText: '今\n手にある',
+    currentTimeSec: 0.8,
+    cues: [cue],
+  });
+
+  assert.deepEqual(resolved, {
+    text: '今　手にある',
+    startTime: 1.2,
+    endTime: 3.8,
+    cues: [cue],
+  });
+});
+
+test('resolveCanonicalPrimarySubtitle covers exit frames that outlive the authored timing', () => {
+  // Real generated animations keep exit fragments on screen well past the authored
+  // comment window; the recorded animation envelope is what makes them resolvable.
+  const cue = {
+    startTime: 1.2,
+    endTime: 3.8,
+    text: '今　手にある',
+    source: 'canonical-ass' as const,
+    animationStartTime: 0.8,
+    animationEndTime: 5.6,
+  };
+  const resolved = resolveCanonicalPrimarySubtitle({
+    liveText: '今\n手にある',
+    currentTimeSec: 5.4,
+    cues: [cue],
+  });
+
+  assert.deepEqual(resolved, {
+    text: '今　手にある',
+    startTime: 1.2,
+    endTime: 3.8,
+    cues: [cue],
+  });
+});
+
+test('resolvePrimarySubtitleText handles late exit frames overlapping the next active line', () => {
+  // The previous line's exit fragments can persist more than a second into the next
+  // authored line. The next line supplies the text; the previous line's envelope
+  // explains its lingering fragments.
+  const cues = [
+    {
+      startTime: 1.2,
+      endTime: 3.8,
+      text: '今　手にある',
+      source: 'canonical-ass' as const,
+      animationStartTime: 0.8,
+      animationEndTime: 5.6,
+    },
+    {
+      startTime: 3.8,
+      endTime: 6.4,
+      text: '物差しでは',
+      source: 'canonical-ass' as const,
+      animationStartTime: 3.4,
+      animationEndTime: 7.0,
+    },
+  ];
+
+  assert.equal(
+    resolvePrimarySubtitleText({
+      liveText: '手にある\n手にある\n物差し\nでは',
+      currentTimeSec: 5.2,
+      cues,
+    }),
+    '物差しでは',
+  );
+});
+
+test('resolveCanonicalPrimarySubtitle rejects unrelated live text at the animation edge', () => {
+  const resolved = resolveCanonicalPrimarySubtitle({
+    liveText: '次のセリフ',
+    currentTimeSec: 4.1,
+    cues: [
+      {
+        startTime: 1.2,
+        endTime: 3.8,
+        text: '今　手にある',
+        source: 'canonical-ass',
+      },
+    ],
+  });
+
+  assert.equal(resolved, null);
+});
+
+test('stripCanonicalFragmentLines drops fragment lines but keeps concurrent dialogue', () => {
+  const cues = [
+    {
+      startTime: 1.2,
+      endTime: 3.8,
+      text: '今　手にある',
+      source: 'canonical-ass' as const,
+    },
+  ];
+
+  assert.equal(
+    stripCanonicalFragmentLines({
+      liveText: '普通のセリフ\n今\n手にある',
+      currentTimeSec: 2,
+      cues,
+    }),
+    '普通のセリフ',
+  );
+  // No canonical cue nearby: nothing to strip.
+  assert.equal(
+    stripCanonicalFragmentLines({ liveText: '普通のセリフ\n今', currentTimeSec: 30, cues }),
+    '普通のセリフ\n今',
+  );
+  // Everything matched (defensive): return the input rather than empty text.
+  assert.equal(
+    stripCanonicalFragmentLines({ liveText: '今\n手にある', currentTimeSec: 2, cues }),
+    '今\n手にある',
+  );
+});
+
+test('resolveCanonicalPrimarySubtitle picks the cue its fragments spell, not the nearest', () => {
+  // In the gap between two authored spans, the next line sits closer in time while only
+  // the previous line's exit fragments are on screen: the fragments decide.
+  const cues = [
+    {
+      startTime: 1,
+      endTime: 3,
+      text: '今　手にある',
+      source: 'canonical-ass' as const,
+      animationStartTime: 0.6,
+      animationEndTime: 3.9,
+    },
+    {
+      startTime: 4,
+      endTime: 6,
+      text: '物差しでは',
+      source: 'canonical-ass' as const,
+      animationStartTime: 3.5,
+      animationEndTime: 6.4,
+    },
+  ];
+
+  assert.equal(
+    resolveCanonicalPrimarySubtitle({ liveText: '手にある', currentTimeSec: 3.8, cues })?.text,
+    '今　手にある',
+  );
+  // Fragments of both lines in the gap: both envelopes cover the moment (distance 0),
+  // and the earlier line wins the tie while it is still animating out.
+  assert.equal(
+    resolveCanonicalPrimarySubtitle({ liveText: '手にある\n物差し', currentTimeSec: 3.8, cues })
+      ?.text,
+    '今　手にある',
+  );
+});

src/main/runtime/primary-subtitle-text.ts

@@ -0,0 +1,160 @@
+import type { SubtitleCue } from '../../types';
+
+// Slack on top of each cue's recorded animation envelope, for time-pos observation
+// staleness and small user sub-delay offsets. The envelope itself covers how far
+// entrance/exit frames actually run past the authored timing.
+const CANONICAL_ANIMATION_EDGE_TOLERANCE_SECONDS = 1;
+
+export interface ResolvedPrimarySubtitle {
+  text: string;
+  startTime: number;
+  endTime: number;
+  /** The canonical cues behind `text`, for consumers that record lines individually. */
+  cues: SubtitleCue[];
+}
+
+function animationSpan(cue: SubtitleCue): { start: number; end: number } {
+  return {
+    start: cue.animationStartTime ?? cue.startTime,
+    end: cue.animationEndTime ?? cue.endTime,
+  };
+}
+
+function nearbyCanonicalCues(
+  cues: readonly SubtitleCue[] | null | undefined,
+  currentTimeSec: number,
+): SubtitleCue[] {
+  return (cues ?? []).filter((cue) => {
+    if (cue.source !== 'canonical-ass') {
+      return false;
+    }
+    const span = animationSpan(cue);
+    return (
+      span.end >= currentTimeSec - CANONICAL_ANIMATION_EDGE_TOLERANCE_SECONDS &&
+      span.start <= currentTimeSec + CANONICAL_ANIMATION_EDGE_TOLERANCE_SECONDS
+    );
+  });
+}
+
+function compactWhitespace(text: string): string {
+  return text.replace(/\s+/gu, '');
+}
+
+/**
+ * mpv's `sub-text` renders each simultaneously active ASS event on its own line, so
+ * while a generated animation plays every live line is a contiguous piece of the
+ * authored text. A line that is not -- concurrent dialogue during an insert song, or a
+ * fresh line starting just after the animation ended -- proves the live text is not this
+ * animation, and substituting the canonical line would swallow real dialogue.
+ */
+function liveTextIsFromCues(liveText: string, cues: readonly SubtitleCue[]): boolean {
+  const compactCues = cues.map((cue) => compactWhitespace(cue.text));
+  const segments = liveText.split('\n').map(compactWhitespace).filter(Boolean);
+  return (
+    segments.length > 0 &&
+    segments.every((segment) => compactCues.some((cueText) => cueText.includes(segment)))
+  );
+}
+
+export function resolveCanonicalPrimarySubtitle(options: {
+  liveText: string;
+  currentTimeSec: number;
+  cues: readonly SubtitleCue[] | null | undefined;
+}): ResolvedPrimarySubtitle | null {
+  if (!Number.isFinite(options.currentTimeSec)) {
+    return null;
+  }
+
+  // Consecutive karaoke lines overlap: one line's exit frames are still on screen while
+  // the next line's entrance frames appear. The fragment check therefore runs against
+  // every canonical cue whose animation envelope reaches the current time, while only
+  // the active (or single nearest) cue supplies the displayed text.
+  const nearby = nearbyCanonicalCues(options.cues, options.currentTimeSec);
+  const active = nearby.filter(
+    (cue) => cue.startTime <= options.currentTimeSec && cue.endTime > options.currentTimeSec,
+  );
+  const liveSegments = options.liveText.split('\n').map(compactWhitespace).filter(Boolean);
+  const selected =
+    active.length > 0
+      ? active
+      : nearby
+          // Between authored spans, proximity alone can pick the wrong neighbor: the
+          // next line can sit closer while only the previous line's exit fragments are
+          // on screen. Only cues that explain at least one live line may be selected.
+          .filter((cue) => {
+            const cueText = compactWhitespace(cue.text);
+            return liveSegments.some((segment) => cueText.includes(segment));
+          })
+          .map((cue) => {
+            const span = animationSpan(cue);
+            const distance =
+              options.currentTimeSec < span.start
+                ? span.start - options.currentTimeSec
+                : Math.max(0, options.currentTimeSec - span.end);
+            return { cue, distance };
+          })
+          .sort((a, b) => a.distance - b.distance || a.cue.startTime - b.cue.startTime)
+          .slice(0, 1)
+          .map(({ cue }) => cue);
+  if (selected.length === 0 || !liveTextIsFromCues(options.liveText, nearby)) {
+    return null;
+  }
+
+  const texts: string[] = [];
+  const seen = new Set<string>();
+  for (const cue of selected) {
+    if (!seen.has(cue.text)) {
+      seen.add(cue.text);
+      texts.push(cue.text);
+    }
+  }
+  return {
+    text: texts.join('\n'),
+    startTime: Math.min(...selected.map((cue) => cue.startTime)),
+    endTime: Math.max(...selected.map((cue) => cue.endTime)),
+    cues: selected,
+  };
+}
+
+/**
+ * Live text with generated-animation fragment lines removed. Recording paths use this
+ * when full canonical substitution declined -- concurrent dialogue during an insert
+ * song: the dialogue is worth recording, the glyph fragments beside it are not. Returns
+ * the input unchanged when no canonical cue is near or nothing non-fragment remains.
+ */
+export function stripCanonicalFragmentLines(options: {
+  liveText: string;
+  currentTimeSec: number;
+  cues: readonly SubtitleCue[] | null | undefined;
+}): string {
+  if (!Number.isFinite(options.currentTimeSec)) {
+    return options.liveText;
+  }
+  const nearby = nearbyCanonicalCues(options.cues, options.currentTimeSec);
+  if (nearby.length === 0) {
+    return options.liveText;
+  }
+  const compactCues = nearby.map((cue) => compactWhitespace(cue.text));
+  const kept = options.liveText.split('\n').filter((line) => {
+    const compact = compactWhitespace(line);
+    return compact && !compactCues.some((cueText) => cueText.includes(compact));
+  });
+  return kept.length > 0 ? kept.join('\n') : options.liveText;
+}
+
+export function resolvePrimarySubtitleText(options: {
+  liveText: string;
+  currentTimeSec: number;
+  cues: readonly SubtitleCue[] | null | undefined;
+}): string {
+  if (!options.liveText.trim()) {
+    return options.liveText;
+  }
+  return (
+    resolveCanonicalPrimarySubtitle({
+      liveText: options.liveText,
+      currentTimeSec: options.currentTimeSec,
+      cues: options.cues,
+    })?.text ?? options.liveText
+  );
+}

src/renderer/subtitle-render.test.ts

@@ -1424,6 +1424,19 @@ test('subtitle annotation CSS underlines JLPT tokens without changing token colo
   );
 });
 
+test('prepareSecondarySubtitleLines drops layered duplicate lines in short stacks', () => {
+  // A word-level animation stacks one event per layer copy; the stack is too short for
+  // the karaoke heuristic but the duplicates are still never distinct content.
+  assert.deepEqual(prepareSecondarySubtitleLines('Your\\NYour\\NYour\\NYour\\Nmosaic'), [
+    'Your',
+    'mosaic',
+  ]);
+  assert.deepEqual(prepareSecondarySubtitleLines('One line\\NAnother line'), [
+    'One line',
+    'Another line',
+  ]);
+});
+
 test('prepareSecondarySubtitleLines collapses karaoke syllable spam into one deduped line', () => {
   // Karaoke-typeset OP/ED: one ASS event per syllable, duplicated across layers,
   // joined with \N by mpv's secondary-sub-text.

src/renderer/subtitle-render.ts

@@ -677,10 +677,12 @@ export function prepareSecondarySubtitleLines(text: string): string[] {
     .split('\n')
     .map((line) => line.trim())
     .filter((line) => line.length > 0);
-  if (!isKaraokeLikeLineSet(lines)) {
-    return lines;
-  }
 
+  // Identical lines in one render are layered copies of the same event (animation
+  // scripts stack several per glyph), never distinct content -- always drop them, so a
+  // short stack like "Your ×4 / mosaic" collapses without needing the karaoke
+  // heuristic. Karaoke-likeness is still judged on the raw stack, where the layered
+  // repetition is the signal.
   const seen = new Set<string>();
   const unique: string[] = [];
   for (const line of lines) {
@@ -688,6 +690,10 @@ export function prepareSecondarySubtitleLines(text: string): string[] {
     seen.add(line);
     unique.push(line);
   }
+  if (!isKaraokeLikeLineSet(lines)) {
+    return unique;
+  }
+
   return [unique.join(' ')];
 }