fix(streaming): keep subtitle tokenization prefetch warm for full episodes (#183)

src/core/services/subtitle-prefetch.test.ts

@@ -74,7 +74,6 @@ test('prefetch service tokenizes priority window cues and caches them', async ()
     preCacheTokenization: (text, data) => {
       cached.set(text, data);
     },
-    isCacheFull: () => false,
     priorityWindowSize: 3,
   });
 
@@ -91,32 +90,38 @@ test('prefetch service tokenizes priority window cues and caches them', async ()
   assert.ok(cached.has('line-2'));
 });
 
-test('prefetch service stops when cache is full', async () => {
+test('prefetch service warms every cue even when the cache evicts along the way', async () => {
   const cues = makeCues(20);
-  let tokenizeCalls = 0;
-  let cacheSize = 0;
+  const tokenized: string[] = [];
+  // Stand-in for the LRU: only the last 5 entries survive, so later cues evict earlier ones.
+  const cache = new Set<string>();
 
   const service = createSubtitlePrefetchService({
     cues,
     tokenizeSubtitle: async (text) => {
-      tokenizeCalls += 1;
+      tokenized.push(text);
       return { text, tokens: [] };
     },
-    preCacheTokenization: () => {
-      cacheSize += 1;
+    preCacheTokenization: (text) => {
+      cache.add(text);
+      while (cache.size > 5) {
+        const oldest = cache.values().next().value;
+        if (oldest === undefined) break;
+        cache.delete(oldest);
+      }
     },
-    isCacheFull: () => cacheSize >= 5,
+    hasCachedTokenization: (text) => cache.has(text),
     priorityWindowSize: 3,
   });
 
   service.start(0);
-  for (let i = 0; i < 30; i += 1) {
+  for (let i = 0; i < 60; i += 1) {
     await flushMicrotasks();
   }
   service.stop();
 
-  // Should have stopped at 5 (cache full), not tokenized all 20
-  assert.ok(tokenizeCalls <= 6, `Expected <= 6 tokenize calls, got ${tokenizeCalls}`);
+  assert.equal(tokenized.length, 20, `Expected all 20 cues warmed, got ${tokenized.length}`);
+  assert.equal(new Set(tokenized).size, 20, 'Each cue is tokenized at most once per run');
 });
 
 test('prefetch service can be stopped mid-flight', async () => {
@@ -130,7 +135,6 @@ test('prefetch service can be stopped mid-flight', async () => {
       return { text, tokens: [] };
     },
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     priorityWindowSize: 3,
   });
 
@@ -159,7 +163,6 @@ test('prefetch service onSeek re-prioritizes from new position', async () => {
     preCacheTokenization: (text) => {
       cachedTexts.push(text);
     },
-    isCacheFull: () => false,
     priorityWindowSize: 3,
   });
 
@@ -183,7 +186,7 @@ test('prefetch service onSeek re-prioritizes from new position', async () => {
   assert.ok(hasPostSeekCue, 'Should have cached cues after seek position');
 });
 
-test('prefetch service still warms the priority window when cache is full', async () => {
+test('prefetch service warms the priority window ahead of the rest of the file', async () => {
   const cues = makeCues(20);
   const cachedTexts: string[] = [];
 
@@ -193,7 +196,6 @@ test('prefetch service still warms the priority window when cache is full', asyn
     preCacheTokenization: (text) => {
       cachedTexts.push(text);
     },
-    isCacheFull: () => true,
     priorityWindowSize: 3,
   });
 
@@ -217,7 +219,6 @@ test('prefetch service pause/resume halts and continues tokenization', async ()
       return { text, tokens: [] };
     },
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     priorityWindowSize: 3,
   });
 
@@ -255,7 +256,6 @@ test('prefetch service skips cues already present in tokenization cache', async
     },
     preCacheTokenization: () => {},
     hasCachedTokenization: (text) => text === 'line-0' || text === 'line-1',
-    isCacheFull: () => false,
     priorityWindowSize: 3,
   });
 
@@ -285,7 +285,6 @@ test('prefetch service deduplicates repeated cue text within a run', async () =>
       return { text, tokens: [] };
     },
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     priorityWindowSize: 3,
   });
 

src/core/services/subtitle-prefetch.ts

@@ -7,7 +7,6 @@ export interface SubtitlePrefetchServiceDeps {
   tokenizeSubtitle: (text: string) => Promise<SubtitleData | null>;
   preCacheTokenization: (text: string, data: SubtitleData) => void;
   hasCachedTokenization?: (text: string) => boolean;
-  isCacheFull: () => boolean;
   priorityWindowSize?: number;
 }
 
@@ -57,11 +56,14 @@ export function createSubtitlePrefetchService(
   let paused = false;
   let currentRunId = 0;
 
+  // A run is a single bounded pass over one file's cues, deduped by `warmedKeys` and by
+  // `hasCachedTokenization`, so the worst case is one tokenization per cue. The cache is
+  // an LRU and bounds its own memory, so a full cache is not a reason to stop warming;
+  // stopping there used to leave the tail of longer media permanently uncached.
   async function tokenizeCueList(
     cuesToProcess: SubtitleCue[],
     runId: number,
     warmedKeys: Set<string>,
-    options: { allowWhenCacheFull?: boolean } = {},
   ): Promise<void> {
     for (const cue of cuesToProcess) {
       if (stopped || runId !== currentRunId) {
@@ -77,10 +79,6 @@ export function createSubtitlePrefetchService(
         return;
       }
 
-      if (!options.allowWhenCacheFull && deps.isCacheFull()) {
-        return;
-      }
-
       const cacheKey = normalizeSubtitleCacheKey(cue.text);
       if (!cacheKey || warmedKeys.has(cacheKey) || deps.hasCachedTokenization?.(cue.text)) {
         if (cacheKey) {
@@ -110,7 +108,7 @@ export function createSubtitlePrefetchService(
 
     // Phase 1: Priority window
     const priorityCues = computePriorityWindow(cues, currentTimeSeconds, windowSize);
-    await tokenizeCueList(priorityCues, runId, warmedKeys, { allowWhenCacheFull: true });
+    await tokenizeCueList(priorityCues, runId, warmedKeys);
 
     if (stopped || runId !== currentRunId) {
       return;

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

@@ -308,25 +308,54 @@ test('hasCachedSubtitle checks prefetched entries without consuming them', async
   assert.equal(controller.hasCachedSubtitle('猫\nです'), false);
 });
 
-test('isCacheFull returns false when cache is below limit', () => {
+test('cache keeps every entry while below the limit', () => {
   const controller = createSubtitleProcessingController({
-    tokenizeSubtitle: async (text) => ({ text, tokens: null }),
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
     emitSubtitle: () => {},
+    cacheLimit: 8,
   });
 
-  assert.equal(controller.isCacheFull(), false);
+  for (let i = 0; i < 8; i += 1) {
+    controller.preCacheTokenization(`line-${i}`, { text: `line-${i}`, tokens: [] });
+  }
+
+  assert.deepEqual(
+    Array.from({ length: 8 }, (_, i) => controller.hasCachedSubtitle(`line-${i}`)),
+    Array.from({ length: 8 }, () => true),
+  );
 });
 
-test('isCacheFull returns true when cache reaches limit', async () => {
+test('cache evicts least recently used entries once the limit is reached', () => {
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: () => {},
+    cacheLimit: 3,
+  });
+
+  for (const line of ['a', 'b', 'c']) {
+    controller.preCacheTokenization(line, { text: line, tokens: [] });
+  }
+  // Touching 'a' makes 'b' the eviction candidate.
+  controller.consumeCachedSubtitle('a');
+  controller.preCacheTokenization('d', { text: 'd', tokens: [] });
+
+  assert.equal(controller.hasCachedSubtitle('b'), false);
+  assert.deepEqual(
+    ['a', 'c', 'd'].map((line) => controller.hasCachedSubtitle(line)),
+    [true, true, true],
+  );
+});
+
+test('default cache limit covers a full-length title without evicting', () => {
   const controller = createSubtitleProcessingController({
     tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
     emitSubtitle: () => {},
   });
 
-  // Fill cache to the 256 limit
-  for (let i = 0; i < 256; i += 1) {
+  for (let i = 0; i < 2000; i += 1) {
     controller.preCacheTokenization(`line-${i}`, { text: `line-${i}`, tokens: [] });
   }
 
-  assert.equal(controller.isCacheFull(), true);
+  assert.equal(controller.hasCachedSubtitle('line-0'), true);
+  assert.equal(controller.hasCachedSubtitle('line-1999'), true);
 });

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

@@ -5,8 +5,17 @@ export interface SubtitleProcessingControllerDeps {
   emitSubtitle: (payload: SubtitleData) => void;
   logDebug?: (message: string) => void;
   now?: () => number;
+  cacheLimit?: number;
 }
 
+/**
+ * Pure memory bound on the LRU, not a coverage limit: prefetching runs to the end of a
+ * file regardless of cache pressure. Sized to hold a feature-length title (a 24-minute
+ * episode runs 300-400 lines, a 2-hour film ~2000) plus room for lines that repeat across
+ * episodes of a series, so openings and endings stay warm between titles.
+ */
+export const DEFAULT_SUBTITLE_TOKENIZATION_CACHE_LIMIT = 2500;
+
 export interface SubtitleProcessingController {
   onSubtitleChange: (text: string) => void;
   refreshCurrentSubtitle: (textOverride?: string) => void;
@@ -14,7 +23,6 @@ export interface SubtitleProcessingController {
   preCacheTokenization: (text: string, data: SubtitleData) => void;
   consumeCachedSubtitle: (text: string) => SubtitleData | null;
   hasCachedSubtitle: (text: string) => boolean;
-  isCacheFull: () => boolean;
 }
 
 export function normalizeSubtitleCacheKey(text: string): string {
@@ -24,7 +32,10 @@ export function normalizeSubtitleCacheKey(text: string): string {
 export function createSubtitleProcessingController(
   deps: SubtitleProcessingControllerDeps,
 ): SubtitleProcessingController {
-  const SUBTITLE_TOKENIZATION_CACHE_LIMIT = 256;
+  const SUBTITLE_TOKENIZATION_CACHE_LIMIT =
+    deps.cacheLimit && deps.cacheLimit > 0
+      ? deps.cacheLimit
+      : DEFAULT_SUBTITLE_TOKENIZATION_CACHE_LIMIT;
   let latestText = '';
   let lastEmittedText = '';
   let cacheGeneration = 0;
@@ -174,8 +185,5 @@ export function createSubtitleProcessingController(
     hasCachedSubtitle: (text: string) => {
       return tokenizationCache.has(normalizeSubtitleCacheKey(text));
     },
-    isCacheFull: () => {
-      return tokenizationCache.size >= SUBTITLE_TOKENIZATION_CACHE_LIMIT;
-    },
   };
 }

src/main.ts

@@ -1955,7 +1955,6 @@ const subtitlePrefetchInitController = createSubtitlePrefetchInitController({
     subtitleProcessingController.preCacheTokenization(text, data);
   },
   hasCachedTokenization: (text) => subtitleProcessingController.hasCachedSubtitle(text),
-  isCacheFull: () => subtitleProcessingController.isCacheFull(),
   logInfo: (message) => logger.info(message),
   logWarn: (message) => logger.warn(message),
   onParsedSubtitleCuesChanged: (cues, sourceKey) => {
@@ -1989,7 +1988,11 @@ const refreshSubtitlePrefetchFromActiveTrackHandler =
   createRefreshSubtitlePrefetchFromActiveTrackHandler({
     getMpvClient: () => appState.mpvClient,
     getLastObservedTimePos: () => lastObservedTimePos,
-    shouldKeepExistingCuesOnMissingSource: (videoPath) => isYoutubeMediaPath(videoPath),
+    // Remote media has no extractable on-disk track to fall back to, so a transient
+    // resolve miss (sid briefly 'no', a cycle onto an embedded stream track) would
+    // otherwise drop a working cue list for the rest of the episode.
+    shouldKeepExistingCuesOnMissingSource: (videoPath) =>
+      isYoutubeMediaPath(videoPath) || isRemoteMediaPath(videoPath),
     subtitlePrefetchInitController,
     resolveActiveSubtitleSidebarSource: (input) => resolveActiveSubtitleSidebarSourceHandler(input),
     logDebug: (message) => logger.debug(message),
@@ -2997,6 +3000,8 @@ const {
         streamIndex,
         delaySeconds,
       }),
+    initSubtitlePrefetch: (sourcePath) =>
+      subtitlePrefetchRuntime.refreshSubtitleSidebarFromSource(sourcePath),
     logDebug: (message, error) => {
       logger.debug(message, error);
     },

src/main/main-wiring.test.ts

@@ -176,6 +176,29 @@ test('subtitle sidebar media path tag is assigned after prefetch succeeds', () =
   );
 });
 
+test('remote media keeps parsed cues when the active subtitle source cannot be resolved', () => {
+  const source = readMainSource();
+  const actionBlock = source.match(
+    /createRefreshSubtitlePrefetchFromActiveTrackHandler\(\{(?<body>[\s\S]*?)\n  \}\);/,
+  )?.groups?.body;
+
+  assert.ok(actionBlock);
+  assert.match(actionBlock, /isYoutubeMediaPath\(videoPath\) \|\| isRemoteMediaPath\(videoPath\)/);
+});
+
+test('jellyfin subtitle preload seeds the tokenization prefetch directly', () => {
+  const source = readMainSource();
+  const actionBlock = source.match(
+    /preloadJellyfinExternalSubtitlesMainDeps:\s*\{(?<body>[\s\S]*?)\n  \},/,
+  )?.groups?.body;
+
+  assert.ok(actionBlock);
+  assert.match(
+    actionBlock,
+    /initSubtitlePrefetch: \(sourcePath\) =>\s*subtitlePrefetchRuntime\.refreshSubtitleSidebarFromSource\(sourcePath\),/,
+  );
+});
+
 test('update overlay notification action triggers install flow', () => {
   const source = readMainSource();
   const runtimeSource = readSource('src/main/runtime/overlay-notifications-runtime.ts');

src/main/runtime/jellyfin-subtitle-preload-main-deps.ts

@@ -28,6 +28,9 @@ export function createBuildPreloadJellyfinExternalSubtitlesMainDepsHandler(
       ? (itemId, streamIndex, delaySeconds) =>
           deps.saveSubtitleDelay!(itemId, streamIndex, delaySeconds)
       : undefined,
+    initSubtitlePrefetch: deps.initSubtitlePrefetch
+      ? (sourcePath) => deps.initSubtitlePrefetch!(sourcePath)
+      : undefined,
     logDebug: (message: string, error: unknown) => deps.logDebug(message, error),
   });
 }

src/main/runtime/jellyfin-subtitle-preload.test.ts

@@ -40,6 +40,9 @@ function makeDeps(overrides: {
   >[0]['setActiveSubtitleDelayKey'];
   loadSubtitleSourceText?: (source: string) => Promise<string>;
   saveSubtitleDelay?: (itemId: string, streamIndex: number, delaySeconds: number) => void;
+  initSubtitlePrefetch?: Parameters<
+    typeof createPreloadJellyfinExternalSubtitlesHandler
+  >[0]['initSubtitlePrefetch'];
   logDebug?: Parameters<typeof createPreloadJellyfinExternalSubtitlesHandler>[0]['logDebug'];
 }) {
   return {
@@ -58,6 +61,7 @@ function makeDeps(overrides: {
     setActiveSubtitleDelayKey: overrides.setActiveSubtitleDelayKey,
     loadSubtitleSourceText: overrides.loadSubtitleSourceText,
     saveSubtitleDelay: overrides.saveSubtitleDelay,
+    initSubtitlePrefetch: overrides.initSubtitlePrefetch,
     logDebug: overrides.logDebug ?? (() => {}),
   };
 }
@@ -134,6 +138,92 @@ test('preload jellyfin subtitles caches external tracks locally and chooses japa
   ]);
 });
 
+test('preload jellyfin subtitles starts prefetch for the selected japanese track', async () => {
+  const prefetched: string[] = [];
+  const preload = createPreloadJellyfinExternalSubtitlesHandler(
+    makeDeps({
+      listJellyfinSubtitleTracks: async () => [
+        { index: 0, language: 'jpn', title: 'Japanese', deliveryUrl: 'https://sub/a.srt' },
+        { index: 1, language: 'eng', title: 'English', deliveryUrl: 'https://sub/b.srt' },
+      ],
+      getMpvClient: () => ({
+        requestProperty: async () => [
+          {
+            type: 'sub',
+            id: 5,
+            lang: 'jpn',
+            title: 'Japanese',
+            external: true,
+            'external-filename': '/tmp/subminer-jellyfin-subtitles/0.srt',
+          },
+          {
+            type: 'sub',
+            id: 6,
+            lang: 'eng',
+            title: 'English',
+            external: true,
+            'external-filename': '/tmp/subminer-jellyfin-subtitles/1.srt',
+          },
+        ],
+      }),
+      cacheSubtitleTrack: async (track) => ({
+        path: `/tmp/subminer-jellyfin-subtitles/${track.index}.srt`,
+        cleanupDir: '/tmp/subminer-jellyfin-subtitles',
+      }),
+      initSubtitlePrefetch: (sourcePath) => {
+        prefetched.push(sourcePath);
+      },
+    }),
+  );
+
+  await preload({ session, clientInfo, itemId: 'item-1' });
+  await new Promise((resolve) => setImmediate(resolve));
+
+  assert.deepEqual(prefetched, ['/tmp/subminer-jellyfin-subtitles/0.srt']);
+});
+
+test('preload jellyfin subtitles survives prefetch start failures', async () => {
+  const logs: string[] = [];
+  const commands: Array<Array<string | number>> = [];
+  const preload = createPreloadJellyfinExternalSubtitlesHandler(
+    makeDeps({
+      listJellyfinSubtitleTracks: async () => [
+        { index: 0, language: 'jpn', title: 'Japanese', deliveryUrl: 'https://sub/a.srt' },
+      ],
+      getMpvClient: () => ({
+        requestProperty: async () => [
+          {
+            type: 'sub',
+            id: 5,
+            lang: 'jpn',
+            title: 'Japanese',
+            external: true,
+            'external-filename': '/tmp/subminer-jellyfin-subtitles/0.srt',
+          },
+        ],
+      }),
+      sendMpvCommand: (command) => commands.push(command),
+      cacheSubtitleTrack: async (track) => ({
+        path: `/tmp/subminer-jellyfin-subtitles/${track.index}.srt`,
+        cleanupDir: '/tmp/subminer-jellyfin-subtitles',
+      }),
+      initSubtitlePrefetch: async () => {
+        throw new Error('parse failed');
+      },
+      logDebug: (message) => logs.push(message),
+    }),
+  );
+
+  await preload({ session, clientInfo, itemId: 'item-1' });
+  await new Promise((resolve) => setImmediate(resolve));
+
+  assert.deepEqual(logs, ['Failed to start subtitle prefetch for Jellyfin subtitle']);
+  assert.ok(
+    commands.some((command) => command[0] === 'set_property' && command[1] === 'sid'),
+    'subtitle selection still happens when prefetch start fails',
+  );
+});
+
 test('preload jellyfin subtitles stages tracks without temporary subtitle selection', async () => {
   const commands: Array<Array<string | number>> = [];
   const preload = createPreloadJellyfinExternalSubtitlesHandler(

src/main/runtime/jellyfin-subtitle-preload.ts

@@ -320,6 +320,7 @@ export function createPreloadJellyfinExternalSubtitlesHandler(deps: {
   setActiveSubtitleDelayKey?: (key: JellyfinSubtitleDelayKey | null) => void;
   loadSubtitleSourceText?: (source: string) => Promise<string>;
   saveSubtitleDelay?: (itemId: string, streamIndex: number, delaySeconds: number) => boolean | void;
+  initSubtitlePrefetch?: (sourcePath: string) => void | Promise<void>;
   logDebug: (message: string, error: unknown) => void;
 }): PreloadJellyfinExternalSubtitlesHandler {
   const activeCacheDirs = new Set<string>();
@@ -329,6 +330,18 @@ export function createPreloadJellyfinExternalSubtitlesHandler(deps: {
     deps.sendMpvCommand(['set_property', 'sub-delay', 0]);
   }
 
+  // mpv's sid property-change is the only thing that normally starts prefetching, so a
+  // coalesced or missed event leaves the whole episode uncached. The downloaded path is
+  // known here, so seed the pipeline directly instead of waiting on the observer.
+  function startSubtitlePrefetchForCachedTrack(sourcePath: string): void {
+    if (!deps.initSubtitlePrefetch) return;
+    void Promise.resolve()
+      .then(() => deps.initSubtitlePrefetch!(sourcePath))
+      .catch((error) => {
+        deps.logDebug('Failed to start subtitle prefetch for Jellyfin subtitle', error);
+      });
+  }
+
   function cleanupActiveCache(): void {
     const dirs = [...activeCacheDirs];
     if (dirs.length === 0) return;
@@ -438,6 +451,7 @@ export function createPreloadJellyfinExternalSubtitlesHandler(deps: {
             }
           }
           deps.sendMpvCommand(['set_property', 'sid', japanesePrimaryId]);
+          startSubtitlePrefetchForCachedTrack(selectedCachedTrack.path);
         } else {
           deps.setActiveSubtitleDelayKey?.(null);
           resetManagedSubtitleDelay();

src/main/runtime/subtitle-prefetch-init.test.ts

@@ -54,7 +54,6 @@ test('latest subtitle prefetch init wins over stale async loads', async () => {
     }),
     tokenizeSubtitle: async () => null,
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     logInfo: () => {},
     logWarn: () => {},
   });
@@ -99,7 +98,6 @@ test('cancelPendingInit prevents an in-flight load from attaching a stale servic
     }),
     tokenizeSubtitle: async () => null,
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     logInfo: () => {},
     logWarn: () => {},
   });
@@ -137,7 +135,6 @@ test('subtitle prefetch init publishes parsed cues and clears them on cancel', a
     }),
     tokenizeSubtitle: async () => null,
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     logInfo: () => {},
     logWarn: () => {},
     onParsedSubtitleCuesChanged: (cues) => {
@@ -181,7 +178,6 @@ test('subtitle prefetch init publishes the provided stable source key instead of
     }),
     tokenizeSubtitle: async () => null,
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     logInfo: () => {},
     logWarn: () => {},
     onParsedSubtitleCuesChanged: (_cues, source) => {
@@ -222,7 +218,6 @@ test('subtitle prefetch init clears parsed cues when initialization fails', asyn
     }),
     tokenizeSubtitle: async () => null,
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     logInfo: () => {},
     logWarn: () => {},
     onParsedSubtitleCuesChanged: (cues) => {
@@ -247,7 +242,6 @@ test('subtitle prefetch init logs a warning when the source parses to zero cues'
     },
     tokenizeSubtitle: async () => null,
     preCacheTokenization: () => {},
-    isCacheFull: () => false,
     logInfo: () => {},
     logWarn: (message) => warnings.push(message),
   });

src/main/runtime/subtitle-prefetch-init.ts

@@ -14,7 +14,6 @@ export interface SubtitlePrefetchInitControllerDeps {
   tokenizeSubtitle: (text: string) => Promise<SubtitleData | null>;
   preCacheTokenization: (text: string, data: SubtitleData) => void;
   hasCachedTokenization?: (text: string) => boolean;
-  isCacheFull: () => boolean;
   logInfo: (message: string) => void;
   logWarn: (message: string) => void;
   onParsedSubtitleCuesChanged?: (cues: SubtitleCue[] | null, sourceKey: string | null) => void;
@@ -72,7 +71,6 @@ export function createSubtitlePrefetchInitController(
         tokenizeSubtitle: (text) => deps.tokenizeSubtitle(text),
         preCacheTokenization: (text, data) => deps.preCacheTokenization(text, data),
         hasCachedTokenization: (text) => deps.hasCachedTokenization?.(text) ?? false,
-        isCacheFull: () => deps.isCacheFull(),
       });
 
       if (revision !== initRevision) {