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

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
+  );
+}