fix(tokenizer): bound retry ladder and cache, fix name-match edge cases

- Cap only blind shrinking-window retries (not backend-guided shrinks); a line that exhausts the cap escalates to one parseText fallback instead of dropping to raw text
- Bound the cross-line termsFind cache by retained dictionary-entry weight, not just key count, and re-check it when a lookup resolves
- Keep halfwidth katakana character names in the greedy pre-pass, and let a generic word beat a name it fully contains
- Share one Han code-point table between the character dictionary and the scanner's name pre-pass; narrow the mob-disambiguator filter to the split letters, not every one-character term
- Release the subtitle prefetch pause on a new onProcessingSettled signal instead of the tokenized emit, so duplicate/suppressed/failed lines no longer pause prefetch indefinitely
- Split the Yomitan scan runtime's injected helper script into its own file

bun.lock

@@ -26,7 +26,7 @@
         "eslint": "^10.8.0",
         "prettier": "^3.8.1",
         "typescript": "^5.9.3",
-        "undici": "7.28.0",
+        "undici": "7.29.0",
       },
     },
   },
@@ -36,8 +36,9 @@
   "overrides": {
     "@xmldom/xmldom": "0.8.13",
     "app-builder-lib": "26.15.3",
-    "brace-expansion": "5.0.8",
+    "brace-expansion": "5.0.9",
     "electron-builder-squirrel-windows": "26.15.3",
+    "fast-uri": "3.1.5",
     "form-data": "4.0.6",
     "ip-address": "10.2.0",
     "js-yaml": "4.3.0",
@@ -46,6 +47,7 @@
     "picomatch": "4.0.4",
     "tar": "7.5.21",
     "tmp": "0.2.7",
+    "undici": "7.29.0",
   },
   "packages": {
     "@discordjs/collection": ["@discordjs/collection@2.1.1", "", {}, "sha512-LiSusze9Tc7qF03sLCujF5iZp7K+vRNEDBZ86FT9aQAv3vxMLihUvKvpsCWiQ2DJq1tVckopKm1rxomgNUc9hg=="],
@@ -266,7 +268,7 @@
 
     "boolean": ["boolean@3.2.0", "", {}, "sha512-d0II/GO9uf9lfUHH2BQsjxzRJZBdsjgsBiW4BvhWk/3qoKwQFjIDVN19PfX8F2D/r9PCMTtLWjYVCFrpeYUzsw=="],
 
-    "brace-expansion": ["brace-expansion@5.0.8", "", { "dependencies": { "balanced-match": "^4.0.2" } }, "sha512-JZyDyq3D4AUifKTPOB7DELf6XsB3WdPuNxCtob1vFXPsSXhdAiHBWJ/tJ8HAc9aH84BK+5JFZLNkJKx3G9kzQg=="],
+    "brace-expansion": ["brace-expansion@5.0.9", "", { "dependencies": { "balanced-match": "^4.0.2" } }, "sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg=="],
 
     "buffer-from": ["buffer-from@1.1.2", "", {}, "sha512-E+XQCRwSbaaiChtv6k6Dwgc+bx+Bs6vuKJHHl5kox/BaKbhiXzqQOwK4cO22yElGp2OCmjwVhT3HmxgyPGnJfQ=="],
 
@@ -404,7 +406,7 @@
 
     "fast-levenshtein": ["fast-levenshtein@2.0.6", "", {}, "sha512-DCXu6Ifhqcks7TZKY3Hxp3y6qphY5SJZmrWMDrKcERSOXWQdMhU9Ig/PYrzyw/ul9jOIyh0N4M0tbC5hodg8dw=="],
 
-    "fast-uri": ["fast-uri@3.1.4", "", {}, "sha512-8JnbkQ4juDyvYs4mgFGQqg4yCYtFDtUtmp2QIQq11ZZe5CFQ5wcqm1rqDgAh/QdMySuBnPzMUiJUNZG5N/AiQw=="],
+    "fast-uri": ["fast-uri@3.1.5", "", {}, "sha512-gHwA1O9LDIcKunMKhObS/HimwtehO1nPUECKAu5TpKgaO19fcWEl4bliWe1jWxVFvIXztJjjQ4L8XQ1EU9f7Jw=="],
 
     "fdir": ["fdir@6.5.0", "", { "peerDependencies": { "picomatch": "^3 || ^4" }, "optionalPeers": ["picomatch"] }, "sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg=="],
 
@@ -714,7 +716,7 @@
 
     "typescript": ["typescript@5.9.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw=="],
 
-    "undici": ["undici@7.28.0", "", {}, "sha512-cRZYrTDwWznlnRiPjggAGxZXanty6M8RV1ff8Wm4LWXBp7/IG8v5DnOm74DtUBp9OONpK75YlPnIjQqX0dBDtA=="],
+    "undici": ["undici@7.29.0", "", {}, "sha512-IDxfleLmmbSskfWSUATiN1nfn2rDuvnMOqb5CWR92iIfojA0Ud+ulOAAEQ57LPr9rWmsreUyf5lwyao+7GNNVw=="],
 
     "undici-types": ["undici-types@7.16.0", "", {}, "sha512-Zz+aZWSj8LE6zoxD+xrjh4VfkIG8Ya6LvYkZqtUQGJPZjYl53ypCaUwWqo7eI0x66KBGeRo+mlBEkMSeSZ38Nw=="],
 
@@ -754,8 +756,6 @@
 
     "@discordjs/rest/discord-api-types": ["discord-api-types@0.38.43", "", {}, "sha512-sSoBf/nK6m7BGtw65mi+QBuvEWaHE8MMziFLqWL+gT6ME/BLg34dRSVKS3Husx40uU06bvxUc3/X+D9Y6/zAbw=="],
 
-    "@discordjs/rest/undici": ["undici@6.27.0", "", {}, "sha512-YmfV3YnEDzXRC5lZ2jWtWWHKGUm1zIt8AhesR1tens+HTNv+YZlN/dp6G727LOvMJ8xjP9Be7Y2Sdr96LDm+pg=="],
-
     "@discordjs/util/discord-api-types": ["discord-api-types@0.38.43", "", {}, "sha512-sSoBf/nK6m7BGtw65mi+QBuvEWaHE8MMziFLqWL+gT6ME/BLg34dRSVKS3Husx40uU06bvxUc3/X+D9Y6/zAbw=="],
 
     "@electron/asar/commander": ["commander@5.1.0", "", {}, "sha512-P0CysNDQ7rtVw4QIQtm+MRxV66vKFSvlsQvGYXZWR3qFU0jlMKHZZZgw8e+8DSah4UDKMqnknRDQz+xuQXQ/Zg=="],
@@ -808,8 +808,6 @@
 
     "node-gyp/env-paths": ["env-paths@2.2.1", "", {}, "sha512-+h1lkLKhZMTYjog1VEpJNG7NZJWcuc2DDk/qsqSTRRCOXiLjeQ1d1/udrUGhqMxUgAlwKNZ0cf2uqan5GLuS2A=="],
 
-    "node-gyp/undici": ["undici@6.27.0", "", {}, "sha512-YmfV3YnEDzXRC5lZ2jWtWWHKGUm1zIt8AhesR1tens+HTNv+YZlN/dp6G727LOvMJ8xjP9Be7Y2Sdr96LDm+pg=="],
-
     "node-gyp/which": ["which@6.0.1", "", { "dependencies": { "isexe": "^4.0.0" }, "bin": { "node-which": "bin/which.js" } }, "sha512-oGLe46MIrCRqX7ytPUf66EAYvdeMIZYn3WaocqqKZAxrBpkqHfL/qvTyJ/bTk5+AqHCjXmrv3CEWgy368zhRUg=="],
 
     "pkijs/@noble/hashes": ["@noble/hashes@1.4.0", "", {}, "sha512-V1JJ1WTRUqHHrOSh597hURcMqVKVGL/ea3kv0gSnEdsEZ0/+VyPghM1lMNGc00z7CIQorSvbKpuJkxvuHbvdbg=="],

changes/background-config-log-level.md

@@ -0,0 +1,4 @@
+type: fixed
+area: logging
+
+- Background startup now respects the configured logging level when `--log-level` is not explicitly provided.

changes/security-audit-high-advisories.md

@@ -0,0 +1,4 @@
+type: internal
+area: dependencies
+
+- Patched three high-severity dependency advisories flagged by `bun audit`: `undici` (cross-user information disclosure via degenerate private cache directives), `brace-expansion` (denial of service via unbounded intermediate arrays), and `fast-uri` (host confusion via backslash authority introducer).

changes/stream-subtitle-tokenization-cache.md

@@ -0,0 +1,7 @@
+type: fixed
+area: streaming
+
+- Jellyfin playback now seeds the subtitle tokenization prefetch straight from the subtitle file it downloads, instead of waiting on an mpv track-selection event that could be missed or coalesced and leave a whole episode tokenizing line by line.
+- Streamed media no longer drops its parsed subtitle cues when the active subtitle track briefly cannot be resolved, such as when cycling onto a subtitle track embedded in the stream.
+- Subtitle prefetching now runs to the end of a file instead of stopping as soon as the tokenization cache fills, so the back half of an episode no longer gets tokenized line by line during playback. Previously the cache was also never cleared between episodes, so the stall carried over to every later title in a session.
+- Raised the tokenization cache from 256 to 2500 lines. It is now purely a memory bound rather than a limit on how much gets prefetched, and it leaves room for lines that repeat across episodes so openings and endings stay warm between titles.

changes/subtitle-plain-line-on-cache-miss.md

@@ -0,0 +1,4 @@
+type: fixed
+area: overlay
+
+- Subtitle lines no longer wait for tokenization to finish before appearing, even when the previous line is still being processed. On a tokenization cache miss, the plain line is shown immediately at its cue time and upgrades in place once tokens and annotations are ready; stale results cannot replace newer cues, and the basic plain-text websocket no longer receives a duplicate event for the annotation-only upgrade. A failed tokenization is no longer cached as the plain line, so a repeated line gets another chance at annotations instead of staying plain for the rest of the session.

changes/subtitle-tokenization-performance.md

@@ -0,0 +1,18 @@
+type: changed
+area: subtitles
+
+- Subtitle tokenization no longer runs a duplicate full `parseText` pass per line: the termsFind scanner walk is now the only tokenizer and emits its own hoverable filler runs for unmatched text (parseText is kept only as an error fallback). This roughly halves the dictionary work per line.
+- The Yomitan scanning helpers are now installed once per parser window (`__subminerYomitanScan`) instead of re-shipping and re-parsing a ~500-line script for every subtitle line; each line only evaluates a tiny call.
+- termsFind lookups are cached across subtitle lines in a window-persistent LRU keyed by substring, so repeated particles and verb forms stop costing backend round trips. The cache invalidates on dictionary/settings changes and window reloads.
+- The scanner walk now skips lookups at punctuation and whitespace positions (latin letters and digits still look up, e.g. Tシャツ). The shrinking-window retry ladder keeps following the consumed lengths the backend reports, and only blind guesses (windows the backend consumed whole, which tell it nothing) are capped at four per position. A line that hits that cap escalates to a single `parseText` for the whole line, so a hard line still resolves to dictionary tokens instead of an unparsed run, without letting the ladder run to one lookup per window length.
+- Tokenizer runtime dependencies are built once instead of per line, fixing a JLPT lookup cache that never hit (it was keyed on a per-call closure identity and leaked a Map per line) and a `which mecab` availability check that re-ran synchronously on every line when MeCab is absent.
+- Subtitle changes no longer restart the prefetch run per line (which discarded in-flight tokenization work); prefetch now only pauses for the live line and restarts on real seeks, cache invalidation, or option changes. Prefetch also stays paused for the whole time the subtitle processing controller is working on the line, including the provisional raw emit that precedes tokenization, so it never competes with the on-screen line for the parser window. The pause is released when the controller reports it has settled, which also covers the lines that finish without an emit (a suppressed duplicate or a failed tokenization) and used to leave prefetching paused indefinitely.
+- Added per-stage debug timings (`scanMs`, `mecabMs`, `frequencyMs`, `annotateMs`) to the subtitle tokenization pipeline log.
+- Fixed a reading that stopped covering its surface when an unmatched kana run extended the preceding token (for example a trailing る on 待ち合わせ), which silently disabled the known-word reading fallback for those tokens.
+- Subtitle prefetching no longer stays paused for the rest of a cue when the same subtitle text is reported twice and there is nothing to tokenize. This covers the startup and overlay priming paths as well as ordinary subtitle changes.
+- Character name and image lookups are now refreshed centrally whenever a character dictionary sync changes its content, so a newly added name can no longer be skipped by a stale candidate list.
+- A subtitle that was on screen when its annotations were invalidated (by mining a card, for example) is now re-annotated instead of staying plain for the rest of the line.
+- Character name annotations no longer cost a dictionary lookup at every position in a line. The scanner now knows which name forms the current title's character dictionary actually contains and only checks where one can start, which removes the whole overhead of having the character dictionary enabled (measured: 21 lookups per line down to 10, the same as with it disabled). Titles with no cached character data keep the previous exhaustive scan, so a missing snapshot costs speed rather than a missing name.
+- The cross-line termsFind cache is now bounded by the number of retained dictionary entries as well as by key count, so a run of lookups that each carry hundreds of entries with full glossaries cannot grow the parser window's memory without limit. The budget is re-checked when a lookup resolves, so a single oversized response is dropped rather than parked in the cache and reused.
+- The unnamed-mob disambiguator filter (Girl A / Girl B) now targets the letters those labels are split into, instead of every one-character term: a character whose name really is one character (𠮷, or a single kana) keeps it. The character dictionary and the scanner's name pre-pass also share one Han code-point table now, so a name the dictionary accepts is a name the scanner will look for.
+- A character name written in halfwidth katakana takes part in the greedy name pre-pass again, so a longer generic word can no longer swallow the start of it.

docs-site/usage.md

@@ -149,7 +149,7 @@ Once Jellyfin is configured, the tray menu includes `Jellyfin Discovery` for sta
 
 - `--log-level` controls logger verbosity.
 - `--dev` and `--debug` are app/dev-mode switches; they are not log-level aliases.
-- `--background` defaults to quieter logging (`warn`) unless `--log-level` is set.
+- `--background` starts at the default quieter logging level (`warn`), then follows `logging.level` after config loads. An explicit `--log-level` remains the override.
 - `--background` launched from a terminal detaches and returns the prompt; stop it with tray Quit or `SubMiner.AppImage --stop` (`SubMiner.exe --stop` on Windows).
 - Linux desktop launcher starts SubMiner with `--background` by default (via electron-builder `linux.executableArgs`).
 - On Hyprland and other Wayland compositors, the tray icon appears only when your panel provides a StatusNotifier/AppIndicator tray host.

docs-site/websocket-texthooker-api.md

@@ -64,7 +64,7 @@ Use the basic subtitle websocket when you only need the current subtitle line as
 - **Client auth:** none
 - **Reconnects:** client-managed
 
-When a client connects, SubMiner immediately sends the latest subtitle payload if one is available. After that, it pushes a new message each time the current subtitle changes.
+When a client connects, SubMiner immediately sends the latest subtitle payload if one is available. After that, it pushes a new message each time the current subtitle changes. Annotation-only upgrades do not repeat the same line on this basic stream.
 
 #### Message shape
 
@@ -96,6 +96,8 @@ Use the annotation websocket for custom clients that want the same structured to
 
 In practice, if you are building a new client, prefer `annotationWebsocket` unless you specifically need compatibility with an existing `websocket` consumer.
 
+On a tokenization cache miss, this stream first sends the cue as plain text with an empty `tokens` array, then sends the annotated replacement when tokenization finishes. Treat each message as the complete current state, replacing the previous payload.
+
 #### Message shape
 
 ```json

docs/architecture/subtitle-overlay-priming.md

@@ -3,7 +3,7 @@
 # Subtitle Overlay Priming
 
 Status: active
-Last verified: 2026-06-14
+Last verified: 2026-08-04
 Owner: Kyle Yasuda
 Read when: debugging subtitle state or blank Linux/X11 overlay windows when the visible overlay is shown or recreated
 
@@ -47,18 +47,39 @@ subtitles do not draw.
    `emitSubtitle(payload)` and `refreshCurrentSubtitle(text)`, then prime secondary subtitles.
 6. Tokenization cache hit: call `consumeCachedSubtitle(text)`, `onSubtitleChange(text)`, and
    `emitSubtitle(cachedPayload)`, then prime secondary subtitles.
-7. Cache miss: call `refreshCurrentSubtitle(text)` and let normal tokenization emit the final
+7. Cache miss: call `refreshCurrentSubtitle(text)`. Normal processing emits a plain payload
-   payload.
+   synchronously, then replaces it with the tokenized payload when ready.
 
-In `src/main.ts`, both `onSubtitleChange` and `refreshCurrentSubtitle` pause
+Both `onSubtitleChange` and `refreshCurrentSubtitle` pause `subtitlePrefetchService` and then call
-`subtitlePrefetchService`, notify it with `onSeek(lastObservedTimePos)`, and then call the matching
+the matching `subtitleProcessingController` method, giving the visible overlay priority over
-`subtitleProcessingController` method. This gives the visible overlay priority over background
+background prefetch work. Prefetch is not re-centered here: restarting the run per line
-prefetch work and re-centers prefetch around the live playback time.
+(`onSeek`) discarded the in-flight tokenization every time the subtitle changed, so only real
+seeks restart it (see `onTimePosUpdate` in `src/main.ts`).
+
+The pause is released by the controller's `onProcessingSettled` callback, which fires once it has
+no work left. Emits do not release it: the first emit for an uncached line is the plain payload
+that precedes tokenization, and a run can finish without emitting at all (a suppressed duplicate,
+a failed tokenization). Both controller methods return whether processing is now pending, and the
+caller resumes immediately when it is not — a repeated subtitle schedules no work, so no settle is
+coming and prefetching would otherwise idle for the rest of the cue.
+
+## Live Cue Delivery
+
+- A tokenization cache miss emits the plain cue synchronously. Tokenization remains serialized so
+  live work does not contend for Yomitan state.
+- If a newer cue arrives while an older line is still tokenizing, the newer plain cue or empty
+  clear payload is emitted immediately. The older tokenization result is dropped before it can
+  replace the current cue.
+- The current cue upgrades in place when its tokens and annotations are ready. This can reflow text
+  or character images, but cue visibility does not wait for that work.
 
 ## Emitted State
 
-- `emitSubtitle(payload)` maps to `emitSubtitlePayload(payload)`, which sends the normal
+- `emitSubtitle(payload)` maps to `emitSubtitlePayload(payload)`. Overlay windows and annotation
-  annotated subtitle payload to overlay windows and subtitle websocket listeners.
+  websocket listeners receive both the immediate plain cue and its later annotation upgrade.
+- The basic subtitle websocket receives the immediate plain cue only. Because its serialized
+  payload discards annotations, the later upgrade would be an identical duplicate and is skipped
+  when text and cue timing match.
 - Secondary priming reads mpv `secondary-sub-text`, stores it in
   `mpvClient.currentSecondarySubText`, and broadcasts `secondary-subtitle:set` to overlay windows.
 - If secondary `requestProperty` fails, the primary flow stays complete and only a debug line is

package.json

@@ -84,8 +84,9 @@
   "overrides": {
     "@xmldom/xmldom": "0.8.13",
     "app-builder-lib": "26.15.3",
-    "brace-expansion": "5.0.8",
+    "brace-expansion": "5.0.9",
     "electron-builder-squirrel-windows": "26.15.3",
+    "fast-uri": "3.1.5",
     "form-data": "4.0.6",
     "ip-address": "10.2.0",
     "js-yaml": "4.3.0",
@@ -93,7 +94,8 @@
     "minimatch": "10.2.5",
     "picomatch": "4.0.4",
     "tar": "7.5.21",
-    "tmp": "0.2.7"
+    "tmp": "0.2.7",
+    "undici": "7.29.0"
   },
   "keywords": [
     "anki",
@@ -125,7 +127,7 @@
     "@types/ws": "^8.18.1",
     "electron": "42.6.0",
     "electron-builder": "26.15.3",
-    "undici": "7.28.0",
+    "undici": "7.29.0",
     "esbuild": "^0.25.12",
     "eslint": "^10.8.0",
     "prettier": "^3.8.1",

src/core/services/index.ts

@@ -1,5 +1,9 @@
 export { Texthooker } from './texthooker';
-export { hasMpvWebsocketPlugin, SubtitleWebSocket } from './subtitle-ws';
+export {
+  hasMpvWebsocketPlugin,
+  isSubtitleAnnotationUpgrade,
+  SubtitleWebSocket,
+} from './subtitle-ws';
 export { registerGlobalShortcuts } from './shortcut';
 export { createIpcDepsRuntime, registerIpcHandlers } from './ipc';
 export { shortcutMatchesInputForLocalFallback } from './shortcut-fallback';

src/core/services/startup-bootstrap.test.ts

@@ -205,7 +205,7 @@ test('runStartupBootstrapRuntime skips lifecycle when generate-config flow handl
   assert.deepEqual(calls, ['setLog:warn:cli', 'forceX11', 'enforceWayland']);
 });
 
-test('runStartupBootstrapRuntime enables quiet background mode by default', () => {
+test('runStartupBootstrapRuntime lets config govern background log level by default', () => {
   const calls: string[] = [];
   const args = makeArgs({ background: true });
 
@@ -222,7 +222,7 @@ test('runStartupBootstrapRuntime enables quiet background mode by default', () =
   });
 
   assert.equal(result.backgroundMode, true);
-  assert.deepEqual(calls, ['setLog:warn:cli', 'forceX11', 'enforceWayland', 'startLifecycle']);
+  assert.deepEqual(calls, ['forceX11', 'enforceWayland', 'startLifecycle']);
 });
 
 test('runStartupBootstrapRuntime enables quiet update mode by default', () => {

src/core/services/startup.ts

@@ -45,7 +45,7 @@ export function runStartupBootstrapRuntime(
 
   if (initialArgs.logLevel) {
     deps.setLogLevel(initialArgs.logLevel, 'cli');
-  } else if (initialArgs.background || initialArgs.update) {
+  } else if (initialArgs.update) {
     deps.setLogLevel('warn', 'cli');
   }
 

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;
+  const tokenized: string[] = [];
-  let cacheSize = 0;
+  // 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: () => {
+    preCacheTokenization: (text) => {
-      cacheSize += 1;
+      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.equal(tokenized.length, 20, `Expected all 20 cues warmed, got ${tokenized.length}`);
-  assert.ok(tokenizeCalls <= 6, `Expected <= 6 tokenize calls, got ${tokenizeCalls}`);
+  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-race.test.ts

@@ -0,0 +1,159 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import type { SubtitleData } from '../../types';
+import { createSubtitleProcessingController } from './subtitle-processing-controller';
+
+function flushMicrotasks(): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, 0));
+}
+
+test('new subtitle emits plain immediately without parallel tokenization or a stale overwrite', async () => {
+  const emitted: SubtitleData[] = [];
+  const resolvers = new Map<string, (value: SubtitleData | null) => void>();
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) =>
+      await new Promise<SubtitleData | null>((resolve) => {
+        resolvers.set(text, resolve);
+      }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('first');
+  controller.onSubtitleChange('second');
+
+  assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
+    { text: 'second', tokens: null },
+  ]);
+  assert.equal(resolvers.has('second'), false);
+
+  const resolveFirst = resolvers.get('first');
+  assert.ok(resolveFirst);
+  resolveFirst({ text: 'first', tokens: [] });
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
+    { text: 'second', tokens: null },
+  ]);
+  assert.equal(resolvers.has('second'), true);
+
+  const resolveSecond = resolvers.get('second');
+  assert.ok(resolveSecond);
+  resolveSecond({ text: 'second', tokens: [] });
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
+    { text: 'second', tokens: null },
+    { text: 'second', tokens: [] },
+  ]);
+});
+
+test('subtitle clears immediately while previous tokenization remains pending', async () => {
+  const emitted: SubtitleData[] = [];
+  let resolveTokenization: ((value: SubtitleData | null) => void) | undefined;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async () =>
+      await new Promise<SubtitleData | null>((resolve) => {
+        resolveTokenization = resolve;
+      }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('first');
+  controller.onSubtitleChange('');
+
+  assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
+    { text: '', tokens: null },
+  ]);
+
+  assert.ok(resolveTokenization);
+  resolveTokenization({ text: 'first', tokens: [] });
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
+    { text: '', tokens: null },
+  ]);
+});
+
+test('returning to an uncached completed line emits it while another line is pending', async () => {
+  const emitted: SubtitleData[] = [];
+  let resolvePending: ((value: SubtitleData | null) => void) | undefined;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      if (text === 'A') {
+        return { text, tokens: [] };
+      }
+      return await new Promise<SubtitleData | null>((resolve) => {
+        resolvePending = resolve;
+      });
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('A');
+  await flushMicrotasks();
+  controller.invalidateTokenizationCache();
+  controller.onSubtitleChange('B');
+  controller.onSubtitleChange('A');
+
+  assert.deepEqual(emitted.at(-1), { text: 'A', tokens: null });
+  assert.ok(resolvePending);
+  resolvePending({ text: 'B', tokens: [] });
+});
+
+test('ABA subtitle changes reuse the matching first tokenization only after A is current again', async () => {
+  const emitted: SubtitleData[] = [];
+  const tokenizeCalls: string[] = [];
+  const resolvers: Array<(value: SubtitleData | null) => void> = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      tokenizeCalls.push(text);
+      return await new Promise<SubtitleData | null>((resolve) => {
+        resolvers.push(resolve);
+      });
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('A');
+  controller.onSubtitleChange('B');
+  controller.onSubtitleChange('A');
+  const resolveFirst = resolvers[0];
+  assert.ok(resolveFirst);
+  resolveFirst({ text: 'A', tokens: [{ value: 1 } as never] });
+  await flushMicrotasks();
+
+  assert.deepEqual(tokenizeCalls, ['A']);
+  assert.deepEqual(emitted, [
+    { text: 'A', tokens: null },
+    { text: 'B', tokens: null },
+    { text: 'A', tokens: null },
+    { text: 'A', tokens: [{ value: 1 } as never] },
+  ]);
+});
+
+test('cached next subtitle does not downgrade to plain while processing is busy', async () => {
+  const emitted: SubtitleData[] = [];
+  let resolveTokenization: ((value: SubtitleData | null) => void) | undefined;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) =>
+      await new Promise<SubtitleData | null>((resolve) => {
+        resolveTokenization = () => resolve({ text, tokens: [] });
+      }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.preCacheTokenization('cached', { text: 'cached', tokens: [] });
+  controller.onSubtitleChange('pending');
+  controller.onSubtitleChange('cached');
+
+  assert.deepEqual(emitted, [{ text: 'pending', tokens: null }]);
+  assert.ok(resolveTokenization);
+  resolveTokenization({ text: 'pending', tokens: [] });
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: 'pending', tokens: null },
+    { text: 'cached', tokens: [] },
+  ]);
+});

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

@@ -7,18 +7,138 @@ function flushMicrotasks(): Promise<void> {
   return new Promise((resolve) => setTimeout(resolve, 0));
 }
 
-test('subtitle processing emits tokenized payload when tokenization succeeds', async () => {
+test('subtitle processing emits plain payload immediately on cache miss, then tokenized payload', async () => {
   const emitted: SubtitleData[] = [];
   const controller = createSubtitleProcessingController({
     tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
     emitSubtitle: (payload) => emitted.push(payload),
   });
 
+  controller.onSubtitleChange('字幕');
+  assert.deepEqual(emitted, [{ text: '字幕', tokens: null }]);
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: '字幕', tokens: null },
+    { text: '字幕', tokens: [] },
+  ]);
+});
+
+test('cache invalidation during pending tokenization does not re-emit the plain payload', async () => {
+  const emitted: SubtitleData[] = [];
+  const resolvers: Array<(value: SubtitleData | null) => void> = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) =>
+      await new Promise<SubtitleData | null>((resolve) => {
+        resolvers.push(() => resolve({ text, tokens: [{ value: resolvers.length } as never] }));
+      }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('行');
+  assert.deepEqual(emitted, [{ text: '行', tokens: null }]);
+
+  controller.invalidateTokenizationCache();
+  resolvers[0]?.({ text: '行', tokens: [] });
+  await flushMicrotasks();
+  // Retry for the new generation is now pending; still no duplicate plain emit.
+  assert.deepEqual(emitted, [{ text: '行', tokens: null }]);
+
+  resolvers[1]?.({ text: '行', tokens: [] });
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: '行', tokens: null },
+    { text: '行', tokens: [{ value: 2 } as never] },
+  ]);
+});
+
+test('failed refresh does not downgrade an already emitted tokenized subtitle', async () => {
+  const emitted: SubtitleData[] = [];
+  let tokenizeCalls = 0;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      tokenizeCalls += 1;
+      if (tokenizeCalls > 1) {
+        throw new Error('tokenizer gone');
+      }
+      return { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('行');
+  await flushMicrotasks();
+  controller.invalidateTokenizationCache();
+  controller.refreshCurrentSubtitle();
+  await flushMicrotasks();
+
+  assert.equal(tokenizeCalls, 2);
+  assert.deepEqual(emitted, [
+    { text: '行', tokens: null },
+    { text: '行', tokens: [] },
+  ]);
+});
+
+test('null-tokenization refresh does not downgrade an already emitted tokenized subtitle', async () => {
+  const emitted: SubtitleData[] = [];
+  let tokenizeCalls = 0;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      tokenizeCalls += 1;
+      return tokenizeCalls > 1 ? null : { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('行');
+  await flushMicrotasks();
+  controller.invalidateTokenizationCache();
+  controller.refreshCurrentSubtitle();
+  await flushMicrotasks();
+
+  assert.equal(tokenizeCalls, 2);
+  assert.deepEqual(emitted, [
+    { text: '行', tokens: null },
+    { text: '行', tokens: [] },
+  ]);
+});
+
+test('subtitle processing does not emit plain payload for cached lines', async () => {
+  const emitted: SubtitleData[] = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.preCacheTokenization('字幕', { text: '字幕', tokens: [] });
   controller.onSubtitleChange('字幕');
   await flushMicrotasks();
   assert.deepEqual(emitted, [{ text: '字幕', tokens: [] }]);
 });
 
+test('subtitle processing shows plain line while tokenization is still pending', async () => {
+  const emitted: SubtitleData[] = [];
+  let resolveTokenization: ((value: SubtitleData | null) => void) | undefined;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) =>
+      await new Promise<SubtitleData | null>((resolve) => {
+        resolveTokenization = () => resolve({ text, tokens: [] });
+      }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('遅い行');
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [{ text: '遅い行', tokens: null }]);
+
+  assert.ok(resolveTokenization);
+  resolveTokenization({ text: '遅い行', tokens: [] });
+  await flushMicrotasks();
+  assert.deepEqual(emitted, [
+    { text: '遅い行', tokens: null },
+    { text: '遅い行', tokens: [] },
+  ]);
+});
+
 test('subtitle processing drops stale tokenization and delivers latest subtitle only once', async () => {
   const emitted: SubtitleData[] = [];
   let firstResolve: ((value: SubtitleData | null) => void) | undefined;
@@ -41,7 +161,11 @@ test('subtitle processing drops stale tokenization and delivers latest subtitle
   await flushMicrotasks();
   await flushMicrotasks();
 
-  assert.deepEqual(emitted, [{ text: 'second', tokens: [] }]);
+  assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
+    { text: 'second', tokens: null },
+    { text: 'second', tokens: [] },
+  ]);
 });
 
 test('subtitle processing skips duplicate subtitle emission', async () => {
@@ -60,7 +184,10 @@ test('subtitle processing skips duplicate subtitle emission', async () => {
   controller.onSubtitleChange('same');
   await flushMicrotasks();
 
-  assert.equal(emitted.length, 1);
+  assert.deepEqual(emitted, [
+    { text: 'same', tokens: null },
+    { text: 'same', tokens: [] },
+  ]);
   assert.equal(tokenizeCalls, 1);
 });
 
@@ -84,7 +211,9 @@ test('subtitle processing reuses cached tokenization for repeated subtitle text'
 
   assert.equal(tokenizeCalls, 2);
   assert.deepEqual(emitted, [
+    { text: 'first', tokens: null },
     { text: 'first', tokens: [] },
+    { text: 'second', tokens: null },
     { text: 'second', tokens: [] },
     { text: 'first', tokens: [] },
   ]);
@@ -100,7 +229,48 @@ test('subtitle processing falls back to plain subtitle when tokenization returns
   controller.onSubtitleChange('fallback');
   await flushMicrotasks();
 
+  assert.deepEqual(
+    emitted,
+    [{ text: 'fallback', tokens: null }],
+    'plain payload should not be re-emitted when tokenization yields nothing new',
+  );
+});
+
+test('null tokenization is not cached and a later cue retries tokenization', async () => {
+  const emitted: SubtitleData[] = [];
+  const callsByText = new Map<string, number>();
+  let failNext = true;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      callsByText.set(text, (callsByText.get(text) ?? 0) + 1);
+      if (text === 'fallback' && failNext) {
+        failNext = false;
+        return null;
+      }
+      return { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('fallback');
+  await flushMicrotasks();
+
+  assert.equal(callsByText.get('fallback'), 1);
+  assert.equal(
+    controller.hasCachedSubtitle('fallback'),
+    false,
+    'plain fallback must not be cached when tokenization yields nothing',
+  );
   assert.deepEqual(emitted, [{ text: 'fallback', tokens: null }]);
+
+  controller.onSubtitleChange('other');
+  await flushMicrotasks();
+  controller.onSubtitleChange('fallback');
+  await flushMicrotasks();
+
+  assert.equal(callsByText.get('fallback'), 2, 'later cue should retry tokenization');
+  assert.equal(controller.hasCachedSubtitle('fallback'), true);
+  assert.deepEqual(emitted.at(-1), { text: 'fallback', tokens: [] });
 });
 
 test('subtitle processing ignores duplicate current subtitle refresh without cache invalidation', async () => {
@@ -120,7 +290,10 @@ test('subtitle processing ignores duplicate current subtitle refresh without cac
   await flushMicrotasks();
 
   assert.equal(tokenizeCalls, 1);
-  assert.deepEqual(emitted, [{ text: 'same', tokens: [] }]);
+  assert.deepEqual(emitted, [
+    { text: 'same', tokens: null },
+    { text: 'same', tokens: [] },
+  ]);
 });
 
 test('subtitle processing coalesces refresh requests while current subtitle is processing', async () => {
@@ -146,7 +319,10 @@ test('subtitle processing coalesces refresh requests while current subtitle is p
   await flushMicrotasks();
 
   assert.equal(tokenizeCalls, 1);
-  assert.deepEqual(emitted, [{ text: 'same', tokens: [] }]);
+  assert.deepEqual(emitted, [
+    { text: 'same', tokens: null },
+    { text: 'same', tokens: [] },
+  ]);
 });
 
 test('subtitle processing refresh re-tokenizes after cache invalidation', async () => {
@@ -168,6 +344,7 @@ test('subtitle processing refresh re-tokenizes after cache invalidation', async
 
   assert.equal(tokenizeCalls, 2);
   assert.deepEqual(emitted, [
+    { text: 'same', tokens: null },
     { text: 'same', tokens: [{ value: 1 } as never] },
     { text: 'same', tokens: [{ value: 2 } as never] },
   ]);
@@ -183,7 +360,10 @@ test('subtitle processing refresh can use explicit text override', async () => {
   controller.refreshCurrentSubtitle('initial');
   await flushMicrotasks();
 
-  assert.deepEqual(emitted, [{ text: 'initial', tokens: [] }]);
+  assert.deepEqual(emitted, [
+    { text: 'initial', tokens: null },
+    { text: 'initial', tokens: [] },
+  ]);
 });
 
 test('subtitle processing cache invalidation only affects future subtitle events', async () => {
@@ -205,10 +385,10 @@ test('subtitle processing cache invalidation only affects future subtitle events
   await flushMicrotasks();
 
   assert.equal(callsByText.get('same'), 1);
-  assert.equal(emitted.length, 3);
+  assert.equal(emitted.length, 5);
 
   controller.invalidateTokenizationCache();
-  assert.equal(emitted.length, 3);
+  assert.equal(emitted.length, 5);
 
   controller.onSubtitleChange('different');
   await flushMicrotasks();
@@ -308,25 +488,161 @@ 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 < 2000; i += 1) {
-  for (let i = 0; i < 256; 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);
+});
+
+test('onSubtitleChange reports whether processing was scheduled', async () => {
+  const emitted: SubtitleData[] = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  // New text schedules work, so an emit (and anything gated on it) will follow.
+  assert.equal(controller.onSubtitleChange('字幕'), true);
+  await flushMicrotasks();
+
+  // A repeat emits nothing, so callers must not wait on an emit that is never
+  // coming (subtitle prefetching would stay paused for the rest of the cue).
+  const emittedCount = emitted.length;
+  assert.equal(controller.onSubtitleChange('字幕'), false);
+  await flushMicrotasks();
+  assert.equal(emitted.length, emittedCount);
+});
+
+test('refreshCurrentSubtitle reports the empty-text emit that an in-flight run will deliver', async () => {
+  const emitted: SubtitleData[] = [];
+  let resolveFirst: ((value: SubtitleData | null) => void) | undefined;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      if (text === '字幕') {
+        return await new Promise<SubtitleData | null>((resolve) => {
+          resolveFirst = resolve;
+        });
+      }
+      return { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  controller.onSubtitleChange('字幕');
+  await flushMicrotasks();
+
+  // Clearing the subtitle while tokenization is in flight: the running loop
+  // picks the empty text up and emits it, so callers gated on that emit (the
+  // prefetch pause) must be told one is coming.
+  assert.equal(controller.refreshCurrentSubtitle(''), true);
+
+  resolveFirst?.({ text: '字幕', tokens: [] });
+  await flushMicrotasks();
+  await flushMicrotasks();
+  // '字幕' is the provisional plain emit the in-flight run already made before
+  // the refresh; '' is the emit the refresh promised.
+  assert.deepEqual(
+    emitted.map((payload) => payload.text),
+    ['字幕', ''],
+  );
+});
+
+test('onProcessingSettled fires once after the queue drains, including runs that emit nothing', async () => {
+  const events: string[] = [];
+  let resolveFirst: ((value: SubtitleData | null) => void) | undefined;
+  let tokenizationFails = false;
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => {
+      if (tokenizationFails) {
+        return null;
+      }
+      if (text === '一行目') {
+        return await new Promise<SubtitleData | null>((resolve) => {
+          resolveFirst = resolve;
+        });
+      }
+      return { text, tokens: [] };
+    },
+    emitSubtitle: (payload) => events.push(`emit:${payload.text}`),
+    onProcessingSettled: () => events.push('settled'),
+  });
+
+  controller.onSubtitleChange('一行目');
+  await flushMicrotasks();
+  // A second line arrives before the first finishes: the controller still has
+  // work, so it must not report itself settled between the two.
+  controller.onSubtitleChange('二行目');
+  resolveFirst?.({ text: '一行目', tokens: [] });
+  await flushMicrotasks();
+  await flushMicrotasks();
+
+  assert.deepEqual(events, ['emit:一行目', 'emit:二行目', 'emit:二行目', 'settled']);
+
+  // Tokenization failure on a line already shown plain: nothing is emitted, and
+  // the settle signal is the only way a caller learns the work is over.
+  events.length = 0;
+  tokenizationFails = true;
+  controller.invalidateTokenizationCache();
+  assert.equal(controller.refreshCurrentSubtitle('二行目'), true);
+  await flushMicrotasks();
+  await flushMicrotasks();
+  assert.deepEqual(events, ['settled']);
+});
+
+test('refreshCurrentSubtitle reports no emit for empty text when nothing is running', async () => {
+  const emitted: SubtitleData[] = [];
+  const controller = createSubtitleProcessingController({
+    tokenizeSubtitle: async (text) => ({ text, tokens: [] }),
+    emitSubtitle: (payload) => emitted.push(payload),
+  });
+
+  assert.equal(controller.refreshCurrentSubtitle(''), false);
+  await flushMicrotasks();
+  assert.deepEqual(emitted, []);
 });

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

@@ -3,18 +3,41 @@ import type { SubtitleData } from '../../types';
 export interface SubtitleProcessingControllerDeps {
   tokenizeSubtitle: (text: string) => Promise<SubtitleData | null>;
   emitSubtitle: (payload: SubtitleData) => void;
+  /**
+   * Fires when the controller runs out of work: every scheduled line has been
+   * processed, whether it ended in an emit, a suppressed duplicate, or a
+   * tokenizer failure. Callers that hold a resource for the duration of
+   * processing (prefetch pausing) release it here rather than on an emit,
+   * which is not guaranteed to happen.
+   */
+  onProcessingSettled?: () => void;
   logDebug?: (message: string) => void;
   now?: () => number;
+  cacheLimit?: number;
 }
 
+/**
+ * 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;
+   * Returns whether processing is now scheduled or already in flight for this
+   * event. A false return means the controller is idle and will do nothing, so
+   * onProcessingSettled will not fire; callers that pause work for the duration
+   * of processing (such as subtitle prefetching) must release it themselves.
+   */
+  onSubtitleChange: (text: string) => boolean;
+  /** Same contract as onSubtitleChange: whether processing is pending. */
+  refreshCurrentSubtitle: (textOverride?: string) => boolean;
   invalidateTokenizationCache: () => void;
   preCacheTokenization: (text: string, data: SubtitleData) => void;
   consumeCachedSubtitle: (text: string) => SubtitleData | null;
   hasCachedSubtitle: (text: string) => boolean;
-  isCacheFull: () => boolean;
 }
 
 export function normalizeSubtitleCacheKey(text: string): string {
@@ -24,9 +47,15 @@ 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 = '';
+  // Tracks the latest provisional plain emit across rapid changes and loop retries
+  // so the same line is never shown plain twice.
+  let lastPlainEmittedText: string | null = null;
   let cacheGeneration = 0;
   let lastEmittedGeneration = 0;
   let processing = false;
@@ -70,9 +99,12 @@ export function createSubtitleProcessingController(
         const startedAtMs = now();
 
         if (!text.trim()) {
+          if (lastPlainEmittedText !== text) {
             deps.emitSubtitle({ text, tokens: null });
+          }
           lastEmittedText = text;
           lastEmittedGeneration = generation;
+          lastPlainEmittedText = null;
           break;
         }
 
@@ -82,11 +114,25 @@ export function createSubtitleProcessingController(
           if (cachedTokenized) {
             output = cachedTokenized;
           } else {
+            // Cache miss: show the plain line on time; the tokenized payload
+            // upgrades it once ready. Skipped on refreshes of an already
+            // emitted line so downstream consumers never see a downgrade.
+            if (text !== lastEmittedText && text !== lastPlainEmittedText) {
+              deps.emitSubtitle({ text, tokens: null });
+              lastPlainEmittedText = text;
+            }
             const tokenized = await deps.tokenizeSubtitle(text);
+            // A null result is a transient tokenizer failure, not a verdict on
+            // the line: caching the plain fallback would pin it untokenized for
+            // every later occurrence.
             if (tokenized) {
               output = tokenized;
+              // A result computed before an invalidation must not repopulate the
+              // fresh cache, or the retry below would serve the stale entry.
+              if (generation === cacheGeneration) {
+                setCachedTokenization(text, tokenized);
+              }
             }
-            setCachedTokenization(text, output);
           }
         } catch (error) {
           deps.logDebug?.(`Subtitle tokenization failed: ${(error as Error).message}`);
@@ -107,9 +153,16 @@ export function createSubtitleProcessingController(
           continue;
         }
 
+        // An untokenized result adds nothing when this line was already shown,
+        // either provisionally or as an earlier full emit (failed refresh) —
+        // emitting it would duplicate or downgrade what is on screen.
+        const plainAlreadyShown = lastPlainEmittedText === text || lastEmittedText === text;
+        if (!(output.tokens === null && output.text === text && plainAlreadyShown)) {
           deps.emitSubtitle(output);
+        }
         lastEmittedText = text;
         lastEmittedGeneration = generation;
+        lastPlainEmittedText = null;
         deps.logDebug?.(
           `Subtitle tokenization delivered; elapsed=${now() - startedAtMs}ms, staleDrops=${staleDropCount}`,
         );
@@ -126,32 +179,50 @@ export function createSubtitleProcessingController(
           (latestText.trim() && cacheGeneration !== lastEmittedGeneration)
         ) {
           processLatest();
+          return;
         }
+        // Nothing left to do: signal completion even when this run emitted
+        // nothing (suppressed duplicate, tokenizer failure), or callers waiting
+        // on the controller would wait forever.
+        deps.onProcessingSettled?.();
       });
   };
 
   return {
     onSubtitleChange: (text: string) => {
       if (text === latestText) {
-        return;
+        // A run already in flight for this text will still emit for it.
+        return processing;
       }
       latestText = text;
+      if (
+        processing &&
+        text !== lastPlainEmittedText &&
+        !tokenizationCache.has(normalizeSubtitleCacheKey(text))
+      ) {
+        deps.emitSubtitle({ text, tokens: null });
+        lastPlainEmittedText = text;
+      }
       processLatest();
+      return true;
     },
     refreshCurrentSubtitle: (textOverride?: string) => {
       if (typeof textOverride === 'string') {
         latestText = textOverride;
       }
       if (!latestText.trim()) {
-        return;
+        // A run in flight will pick this up and emit the empty subtitle, so
+        // the caller is still waiting on an emit.
+        return processing;
       }
-      if (
+      if (processing) {
-        processing ||
+        return true;
-        (latestText === lastEmittedText && cacheGeneration === lastEmittedGeneration)
+      }
-      ) {
+      if (latestText === lastEmittedText && cacheGeneration === lastEmittedGeneration) {
-        return;
+        return false;
       }
       processLatest();
+      return true;
     },
     invalidateTokenizationCache: () => {
       tokenizationCache.clear();
@@ -169,13 +240,11 @@ export function createSubtitleProcessingController(
       latestText = text;
       lastEmittedText = text;
       lastEmittedGeneration = cacheGeneration;
+      lastPlainEmittedText = null;
       return cached;
     },
     hasCachedSubtitle: (text: string) => {
       return tokenizationCache.has(normalizeSubtitleCacheKey(text));
     },
-    isCacheFull: () => {
-      return tokenizationCache.size >= SUBTITLE_TOKENIZATION_CACHE_LIMIT;
-    },
   };
 }

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

@@ -1,6 +1,7 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
 import {
+  isSubtitleAnnotationUpgrade,
   serializeInitialSubtitleWebsocketMessage,
   serializeSubtitleMarkup,
   serializeSubtitleWebsocketMessage,
@@ -13,6 +14,40 @@ const frequencyOptions = {
   mode: 'banded' as const,
 };
 
+test('annotation upgrade requires matching text and cue timing', () => {
+  const current: SubtitleData = {
+    text: '字幕',
+    tokens: null,
+    startTime: 10,
+    endTime: 12,
+  };
+
+  assert.equal(
+    isSubtitleAnnotationUpgrade(current, {
+      ...current,
+      tokens: [],
+    }),
+    true,
+  );
+  assert.equal(
+    isSubtitleAnnotationUpgrade(current, {
+      ...current,
+      tokens: [],
+      startTime: 11,
+    }),
+    false,
+  );
+  assert.equal(
+    isSubtitleAnnotationUpgrade(current, {
+      ...current,
+      text: '次の字幕',
+      tokens: [],
+    }),
+    false,
+  );
+  assert.equal(isSubtitleAnnotationUpgrade(current, current), false);
+});
+
 test('serializeSubtitleMarkup escapes plain text and preserves line breaks', () => {
   const payload: SubtitleData = {
     text: 'a < b\nx & y',

src/core/services/subtitle-ws.ts

@@ -20,6 +20,20 @@ export type SubtitleWebsocketFrequencyOptions = {
 
 export type SubtitleWebsocketPayloadMode = 'plain' | 'annotated';
 
+export function isSubtitleAnnotationUpgrade(
+  current: SubtitleData | null,
+  next: SubtitleData,
+): boolean {
+  return (
+    current !== null &&
+    current.tokens === null &&
+    next.tokens !== null &&
+    current.text === next.text &&
+    current.startTime === next.startTime &&
+    current.endTime === next.endTime
+  );
+}
+
 type SubtitleWebsocketMessageOptions = {
   payloadMode?: SubtitleWebsocketPayloadMode;
 };

src/core/services/tokenizer.test.ts

@@ -2934,44 +2934,12 @@ test('tokenizeSubtitle preserves Yomitan compound token when MeCab components ar
                 return [];
               }
 
-              if (script.includes('parseText')) {
               return [
                 {
-                    source: 'scanning-parser',
+                  surface: '取り組んで',
-                    index: 0,
-                    content: [
-                      [
-                        {
-                          text: '取り組んで',
                   reading: 'とりくんで',
-                          headwords: [[{ term: '取り組む' }]],
+                  headword: '取り組む',
-                        },
-                      ],
-                      [
-                        {
-                          text: 'もらいます',
-                          reading: 'もらいます',
-                          headwords: [[{ term: 'もらう' }]],
-                        },
-                      ],
-                    ],
-                  },
-                ];
-              }
-
-              return [
-                {
-                  surface: '取り',
-                  reading: 'とり',
-                  headword: '取る',
                   startPos: 0,
-                  endPos: 2,
-                },
-                {
-                  surface: '組んで',
-                  reading: 'くんで',
-                  headword: '組む',
-                  startPos: 2,
                   endPos: 5,
                 },
                 {

src/core/services/tokenizer.ts

@@ -70,6 +70,7 @@ export interface TokenizerServiceDeps {
   getNameMatchImagesEnabled?: () => boolean;
   getCharacterNameImage?: (term: string) => CharacterNameImage | null;
   getCurrentCharacterDictionaryMediaId?: () => number | null;
+  getCharacterNameCandidates?: () => { key: string; forms: string[] } | null;
   getFrequencyDictionaryEnabled?: () => boolean;
   getFrequencyDictionaryMatchMode?: () => FrequencyDictionaryMatchMode;
   getFrequencyRank?: FrequencyDictionaryLookup;
@@ -106,6 +107,7 @@ export interface TokenizerDepsRuntimeOptions {
   getNameMatchImagesEnabled?: () => boolean;
   getCharacterNameImage?: (term: string) => CharacterNameImage | null;
   getCurrentCharacterDictionaryMediaId?: () => number | null;
+  getCharacterNameCandidates?: () => { key: string; forms: string[] } | null;
   getFrequencyDictionaryEnabled?: () => boolean;
   getFrequencyDictionaryMatchMode?: () => FrequencyDictionaryMatchMode;
   getFrequencyRank?: FrequencyDictionaryLookup;
@@ -266,6 +268,7 @@ export function createTokenizerDepsRuntime(
     getNameMatchImagesEnabled: options.getNameMatchImagesEnabled,
     getCharacterNameImage: options.getCharacterNameImage,
     getCurrentCharacterDictionaryMediaId: options.getCurrentCharacterDictionaryMediaId,
+    getCharacterNameCandidates: options.getCharacterNameCandidates,
     getFrequencyDictionaryEnabled: options.getFrequencyDictionaryEnabled,
     getFrequencyDictionaryMatchMode: options.getFrequencyDictionaryMatchMode ?? (() => 'headword'),
     getFrequencyRank: options.getFrequencyRank,
@@ -716,15 +719,30 @@ function getAnnotationOptions(deps: TokenizerServiceDeps): TokenizerAnnotationOp
   };
 }
 
+// Per-line stage durations for the pipeline debug log; every field is filled in
+// by the stage that awaits the corresponding work.
+interface TokenizationStageTimings {
+  scanMs?: number;
+  mecabMs?: number;
+  frequencyMs?: number;
+  annotateMs?: number;
+}
+
 async function parseWithYomitanInternalParser(
   text: string,
   deps: TokenizerServiceDeps,
   options: TokenizerAnnotationOptions,
+  stageTimings?: TokenizationStageTimings,
 ): Promise<MergedToken[] | null> {
+  const scanStartedAtMs = Date.now();
   const selectedTokens = await requestYomitanScanTokens(text, deps, logger, {
     includeNameMatchMetadata: options.nameMatchEnabled,
     currentCharacterDictionaryMediaId: deps.getCurrentCharacterDictionaryMediaId?.() ?? null,
+    nameCandidates: deps.getCharacterNameCandidates?.() ?? null,
   });
+  if (stageTimings) {
+    stageTimings.scanMs = Date.now() - scanStartedAtMs;
+  }
   if (!selectedTokens || selectedTokens.length === 0) {
     return null;
   }
@@ -757,6 +775,7 @@ async function parseWithYomitanInternalParser(
 
   const frequencyRankPromise: Promise<YomitanFrequencyIndex> = options.frequencyEnabled
     ? (async () => {
+        const frequencyStartedAtMs = Date.now();
         const frequencyMatchMode = options.frequencyMatchMode;
         const termReadingList = buildYomitanFrequencyTermReadingList(
           normalizedSelectedTokens,
@@ -767,12 +786,17 @@ async function parseWithYomitanInternalParser(
           deps,
           logger,
         );
-        return buildYomitanFrequencyIndex(yomitanFrequencies);
+        const frequencyIndex = buildYomitanFrequencyIndex(yomitanFrequencies);
+        if (stageTimings) {
+          stageTimings.frequencyMs = Date.now() - frequencyStartedAtMs;
+        }
+        return frequencyIndex;
       })()
     : Promise.resolve({ byPair: new Map(), byTerm: new Map() });
 
   const mecabEnrichmentPromise: Promise<MergedToken[]> = needsMecabPosEnrichment(options)
     ? (async () => {
+        const mecabStartedAtMs = Date.now();
         try {
           const mecabTokens = await deps.tokenizeWithMecab(text);
           const enrichTokensWithMecab = deps.enrichTokensWithMecab ?? enrichTokensWithMecabAsync;
@@ -786,6 +810,10 @@ async function parseWithYomitanInternalParser(
             `textLength=${text.length}`,
           );
           return normalizedSelectedTokens;
+        } finally {
+          if (stageTimings) {
+            stageTimings.mecabMs = Date.now() - mecabStartedAtMs;
+          }
         }
       })()
     : Promise.resolve(normalizedSelectedTokens);
@@ -876,15 +904,35 @@ export async function tokenizeSubtitle(
   const annotationOptions = getAnnotationOptions(deps);
   annotationOptions.sourceText = tokenizeText;
 
-  const yomitanTokens = await parseWithYomitanInternalParser(tokenizeText, deps, annotationOptions);
+  const stageTimings: TokenizationStageTimings = {};
+  const startedAtMs = Date.now();
+  const logStageTimings = (tokenCount: number): void => {
+    logger.debug(
+      `Subtitle tokenization stages; textLength=${tokenizeText.length}, tokenCount=${tokenCount}, ` +
+        `scanMs=${stageTimings.scanMs ?? '-'}, mecabMs=${stageTimings.mecabMs ?? '-'}, ` +
+        `frequencyMs=${stageTimings.frequencyMs ?? '-'}, annotateMs=${stageTimings.annotateMs ?? '-'}, ` +
+        `totalMs=${Date.now() - startedAtMs}`,
+    );
+  };
+
+  const yomitanTokens = await parseWithYomitanInternalParser(
+    tokenizeText,
+    deps,
+    annotationOptions,
+    stageTimings,
+  );
   if (yomitanTokens && yomitanTokens.length > 0) {
+    const annotateStartedAtMs = Date.now();
     const annotatedTokens = await applyAnnotationStage(yomitanTokens, deps, annotationOptions);
+    stageTimings.annotateMs = Date.now() - annotateStartedAtMs;
     const renderedTokens = applyCharacterNameImages(annotatedTokens, deps, annotationOptions);
+    logStageTimings(renderedTokens.length);
     return {
       text: displayText,
       tokens: renderedTokens.length > 0 ? renderedTokens : null,
     };
   }
 
+  logStageTimings(0);
   return { text: displayText, tokens: null };
 }

src/core/services/tokenizer/golden-corpus-harness.ts

@@ -366,8 +366,11 @@ export function createReplayMessageStore(messages: GoldenRecordedMessage[]): Rep
   };
 }
 
-async function runInjectedScriptInVm(script: string, store: ReplayMessageStore): Promise<unknown> {
+// One persistent context per fixture, matching the real parser window: the
-  return await vm.runInNewContext(script, {
+// scan runtime installs itself once into globalThis and later per-line call
+// scripts reuse it.
+function createInjectedScriptVm(store: ReplayMessageStore): (script: string) => Promise<unknown> {
+  const context = vm.createContext({
     chrome: {
       runtime: {
         lastError: null,
@@ -393,6 +396,7 @@ async function runInjectedScriptInVm(script: string, store: ReplayMessageStore):
     Set,
     String,
   });
+  return async (script: string) => await vm.runInContext(script, context);
 }
 
 export function createReplayTokenizerDeps(fixture: GoldenFixture): TokenizerServiceDeps {
@@ -400,13 +404,14 @@ export function createReplayTokenizerDeps(fixture: GoldenFixture): TokenizerServ
   const scriptResults = new Map(
     fixture.recording.scripts.map((entry) => [entry.sha256, entry] as const),
   );
+  const runInjectedScriptInVm = createInjectedScriptVm(store);
 
   const parserWindow = {
     isDestroyed: () => false,
     webContents: {
       executeJavaScript: async (script: string) => {
         try {
-          return await runInjectedScriptInVm(script, store);
+          return await runInjectedScriptInVm(script);
         } catch (vmError) {
           const recorded = scriptResults.get(hashInjectedScript(script));
           if (recorded) {

src/core/services/tokenizer/yomitan-parser-runtime.ts

@@ -3,6 +3,14 @@ import * as fs from 'fs';
 import * as http from 'http';
 import * as path from 'path';
 import { selectYomitanParseTokens } from './parser-selection-stage';
+import {
+  buildYomitanScanCallScript,
+  buildYomitanScanNameCandidatesScript,
+  CHARACTER_DICTIONARY_TITLE_PREFIX,
+  YOMITAN_SCAN_RUNTIME_INSTALL_SCRIPT,
+  YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL,
+  type YomitanFrequencyMode,
+} from './yomitan-scan-runtime-script';
 
 interface LoggerLike {
   error: (message: string, ...args: unknown[]) => void;
@@ -22,8 +30,6 @@ interface YomitanParserRuntimeDeps {
   createYomitanExtensionWindow?: (pageName: string) => Promise<BrowserWindow | null>;
 }
 
-type YomitanFrequencyMode = 'occurrence-based' | 'rank-based';
-
 export interface YomitanDictionaryInfo {
   title: string;
   revision?: string | number;
@@ -74,13 +80,19 @@ export interface YomitanAddNoteResult {
 }
 
 const DEFAULT_YOMITAN_SCAN_LENGTH = 40;
-const CHARACTER_DICTIONARY_TITLE_PREFIX = 'SubMiner Character Dictionary';
 const yomitanProfileMetadataByWindow = new WeakMap<BrowserWindow, YomitanProfileMetadata>();
 const yomitanProfileDiagnosticsLoggedByWindow = new WeakSet<BrowserWindow>();
 const yomitanFrequencyCacheByWindow = new WeakMap<
   BrowserWindow,
   Map<string, YomitanTermFrequency[]>
 >();
+// Epoch passed with every scan request; the in-window termsFind cache clears
+// itself when the epoch changes (dictionary imports, settings changes).
+const yomitanScanCacheEpochByWindow = new WeakMap<BrowserWindow, number>();
+
+function getYomitanScanCacheEpoch(window: BrowserWindow): number {
+  return yomitanScanCacheEpochByWindow.get(window) ?? 0;
+}
 
 function isObject(value: unknown): value is Record<string, unknown> {
   return Boolean(value && typeof value === 'object');
@@ -99,6 +111,7 @@ function isScanTokenArray(value: unknown): value is YomitanScanToken[] {
         typeof entry.startPos === 'number' &&
         typeof entry.endPos === 'number' &&
         (entry.isNameMatch === undefined || typeof entry.isNameMatch === 'boolean') &&
+        (entry.isUnparsedRun === undefined || typeof entry.isUnparsedRun === 'boolean') &&
         (entry.frequencyRank === undefined || typeof entry.frequencyRank === 'number') &&
         (entry.wordClasses === undefined ||
           (Array.isArray(entry.wordClasses) &&
@@ -107,13 +120,9 @@ function isScanTokenArray(value: unknown): value is YomitanScanToken[] {
   );
 }
 
-function scanTokenSpanKey(token: YomitanScanToken): string {
-  return `${token.startPos}:${token.endPos}:${token.surface}`;
-}
-
 // Maps a parse-selected token to the scanner-token shape carried out of the
-// parser runtime. Shared by both selectYomitanParseTokens fallback paths so the
+// parser runtime, used by the parseText fallback path when the in-window
-// projected fields stay in sync as the shape changes.
+// scanner is unavailable.
 function toYomitanScanToken(token: {
   surface: string;
   reading: string;
@@ -132,66 +141,6 @@ function toYomitanScanToken(token: {
   };
 }
 
-// parseText segmentation is authoritative (it emits filler chunks for text the
-// termsFind scanner skips), but only the termsFind scanner carries annotation
-// metadata (isNameMatch, frequencyRank, headwordReading, wordClasses). Graft
-// scanner tokens onto the parseText segmentation per matching span so one
-// unmatched chunk degrades only itself instead of dropping the whole line's
-// metadata.
-//
-// Exception: character-name tokens. The greedy name scan can re-segment text
-// around a name (e.g. とヨータ → と + ヨータ instead of とヨー + タ), so
-// parseText segmentation cannot be authoritative there. Each name span is
-// expanded until it aligns with token boundaries in both segmentations, then
-// the parse tokens inside are replaced with the scanner tokens.
-function mergeScannerTokensIntoParseTokens(
-  parseScanTokens: YomitanScanToken[],
-  scannerTokens: YomitanScanToken[],
-): YomitanScanToken[] {
-  const scannerTokensBySpan = new Map<string, YomitanScanToken>();
-  for (const token of scannerTokens) {
-    scannerTokensBySpan.set(scanTokenSpanKey(token), token);
-  }
-  const graftedTokens = parseScanTokens.map(
-    (token) => scannerTokensBySpan.get(scanTokenSpanKey(token)) ?? token,
-  );
-
-  const nameTokens = scannerTokens.filter((token) => token.isNameMatch === true);
-  if (nameTokens.length === 0) {
-    return graftedTokens;
-  }
-
-  const regions = nameTokens.map((token) => ({ start: token.startPos, end: token.endPos }));
-  const allTokens = [...parseScanTokens, ...scannerTokens];
-  let expanded = true;
-  while (expanded) {
-    expanded = false;
-    for (const region of regions) {
-      for (const token of allTokens) {
-        const overlaps = token.startPos < region.end && token.endPos > region.start;
-        const extendsBeyond = token.startPos < region.start || token.endPos > region.end;
-        if (overlaps && extendsBeyond) {
-          region.start = Math.min(region.start, token.startPos);
-          region.end = Math.max(region.end, token.endPos);
-          expanded = true;
-        }
-      }
-    }
-  }
-
-  const isInsideNameRegion = (token: YomitanScanToken): boolean =>
-    regions.some((region) => token.startPos >= region.start && token.endPos <= region.end);
-
-  const merged = graftedTokens.filter((token) => !isInsideNameRegion(token));
-  for (const token of scannerTokens) {
-    if (isInsideNameRegion(token)) {
-      merged.push(token);
-    }
-  }
-  merged.sort((a, b) => a.startPos - b.startPos || a.endPos - b.endPos);
-  return merged;
-}
-
 function makeTermReadingCacheKey(term: string, reading: string | null): string {
   return `${term}\u0000${reading ?? ''}`;
 }
@@ -208,6 +157,7 @@ function getWindowFrequencyCache(window: BrowserWindow): Map<string, YomitanTerm
 function clearWindowCaches(window: BrowserWindow): void {
   yomitanProfileMetadataByWindow.delete(window);
   yomitanFrequencyCacheByWindow.delete(window);
+  yomitanScanCacheEpochByWindow.set(window, getYomitanScanCacheEpoch(window) + 1);
 }
 export function clearYomitanParserCachesForWindow(window: BrowserWindow): void {
   clearWindowCaches(window);
@@ -704,6 +654,10 @@ async function ensureYomitanParserWindow(
       if (readyPromise) {
         await readyPromise;
       }
+      // Eagerly install the scan runtime so the first subtitle line does not
+      // pay the install round trip; failures fall back to the per-request
+      // install-and-retry path.
+      await installYomitanScanRuntime(parserWindow).catch(() => {});
 
       return true;
     } catch (err) {
@@ -877,668 +831,42 @@ async function serveDictionaryZipOnce<T>(
   }
 }
 
-const YOMITAN_SCANNING_HELPERS = String.raw`
+async function installYomitanScanRuntime(parserWindow: BrowserWindow): Promise<void> {
-      const HIRAGANA_CONVERSION_RANGE = [0x3041, 0x3096];
+  await parserWindow.webContents.executeJavaScript(YOMITAN_SCAN_RUNTIME_INSTALL_SCRIPT, true);
-      const KATAKANA_CONVERSION_RANGE = [0x30a1, 0x30f6];
+  // A fresh runtime has no candidate list; force the next scan to reinstall it.
-      const KANA_PROLONGED_SOUND_MARK_CODE_POINT = 0x30fc;
+  yomitanScanNameCandidateKeyByWindow.delete(parserWindow);
-      const KATAKANA_SMALL_KA_CODE_POINT = 0x30f5;
+}
-      const KATAKANA_SMALL_KE_CODE_POINT = 0x30f6;
-      const KANA_RANGES = [[0x3040, 0x309f], [0x30a0, 0x30ff]];
-      const JAPANESE_RANGES = [[0x3040, 0x30ff], [0x3400, 0x9fff]];
-      function isCodePointInRange(codePoint, range) { return codePoint >= range[0] && codePoint <= range[1]; }
-      function isCodePointInRanges(codePoint, ranges) { return ranges.some((range) => isCodePointInRange(codePoint, range)); }
-      function isCodePointKana(codePoint) { return isCodePointInRanges(codePoint, KANA_RANGES); }
-      function isCodePointJapanese(codePoint) { return isCodePointInRanges(codePoint, JAPANESE_RANGES); }
-      function createFuriganaSegment(text, reading) { return {text, reading}; }
-      function getSegmentReadingContribution(segment) {
-        if (typeof segment.reading === "string" && segment.reading.length > 0) { return segment.reading; }
-        const segmentText = typeof segment.text === "string" ? segment.text : "";
-        const isKanaOnly = segmentText.length > 0 && [...segmentText].every((char) => isCodePointKana(char.codePointAt(0)));
-        return isKanaOnly ? segmentText : "";
-      }
-      function getProlongedHiragana(previousCharacter) {
-        switch (previousCharacter) {
-          case "あ": case "か": case "が": case "さ": case "ざ": case "た": case "だ": case "な": case "は": case "ば": case "ぱ": case "ま": case "や": case "ら": case "わ": case "ぁ": case "ゃ": case "ゎ": return "あ";
-          case "い": case "き": case "ぎ": case "し": case "じ": case "ち": case "ぢ": case "に": case "ひ": case "び": case "ぴ": case "み": case "り": case "ぃ": return "い";
-          case "う": case "く": case "ぐ": case "す": case "ず": case "つ": case "づ": case "ぬ": case "ふ": case "ぶ": case "ぷ": case "む": case "ゆ": case "る": case "ぅ": case "ゅ": return "う";
-          case "え": case "け": case "げ": case "せ": case "ぜ": case "て": case "で": case "ね": case "へ": case "べ": case "ぺ": case "め": case "れ": case "ぇ": return "え";
-          case "お": case "こ": case "ご": case "そ": case "ぞ": case "と": case "ど": case "の": case "ほ": case "ぼ": case "ぽ": case "も": case "よ": case "ろ": case "を": case "ぉ": case "ょ": return "う";
-          default: return null;
-        }
-      }
-      function getFuriganaKanaSegments(text, reading) {
-        const newSegments = [];
-        let start = 0;
-        let state = (reading[0] === text[0]);
-        for (let i = 1; i < text.length; ++i) {
-          const newState = (reading[i] === text[i]);
-          if (state === newState) { continue; }
-          newSegments.push(createFuriganaSegment(text.substring(start, i), state ? '' : reading.substring(start, i)));
-          state = newState;
-          start = i;
-        }
-        newSegments.push(createFuriganaSegment(text.substring(start), state ? '' : reading.substring(start)));
-        return newSegments;
-      }
-      function convertKatakanaToHiragana(text, keepProlongedSoundMarks = false) {
-        let result = '';
-        const offset = (HIRAGANA_CONVERSION_RANGE[0] - KATAKANA_CONVERSION_RANGE[0]);
-        for (let char of text) {
-          const codePoint = char.codePointAt(0);
-          switch (codePoint) {
-            case KATAKANA_SMALL_KA_CODE_POINT:
-            case KATAKANA_SMALL_KE_CODE_POINT:
-              break;
-            case KANA_PROLONGED_SOUND_MARK_CODE_POINT:
-              if (!keepProlongedSoundMarks && result.length > 0) {
-                const char2 = getProlongedHiragana(result[result.length - 1]);
-                if (char2 !== null) { char = char2; }
-              }
-              break;
-            default:
-              if (isCodePointInRange(codePoint, KATAKANA_CONVERSION_RANGE)) {
-                char = String.fromCodePoint(codePoint + offset);
-              }
-              break;
-          }
-          result += char;
-        }
-        return result;
-      }
-      function segmentizeFurigana(reading, readingNormalized, groups, groupsStart) {
-        const groupCount = groups.length - groupsStart;
-        if (groupCount <= 0) { return reading.length === 0 ? [] : null; }
-        const group = groups[groupsStart];
-        const {isKana, text} = group;
-        if (isKana) {
-          if (group.textNormalized !== null && readingNormalized.startsWith(group.textNormalized)) {
-            const segments = segmentizeFurigana(reading.substring(text.length), readingNormalized.substring(text.length), groups, groupsStart + 1);
-            if (segments !== null) {
-              if (reading.startsWith(text)) { segments.unshift(createFuriganaSegment(text, '')); }
-              else { segments.unshift(...getFuriganaKanaSegments(text, reading)); }
-              return segments;
-            }
-          }
-          return null;
-        }
-        let result = null;
-        for (let i = reading.length; i >= text.length; --i) {
-          const segments = segmentizeFurigana(reading.substring(i), readingNormalized.substring(i), groups, groupsStart + 1);
-          if (segments !== null) {
-            if (result !== null) { return null; }
-            segments.unshift(createFuriganaSegment(text, reading.substring(0, i)));
-            result = segments;
-          }
-          if (groupCount === 1) { break; }
-        }
-        return result;
-      }
-      function distributeFurigana(term, reading) {
-        if (reading === term) { return [createFuriganaSegment(term, '')]; }
-        const groups = [];
-        let groupPre = null;
-        let isKanaPre = null;
-        for (const c of term) {
-          const isKana = isCodePointKana(c.codePointAt(0));
-          if (isKana === isKanaPre) { groupPre.text += c; }
-          else {
-            groupPre = {isKana, text: c, textNormalized: null};
-            groups.push(groupPre);
-            isKanaPre = isKana;
-          }
-        }
-        for (const group of groups) {
-          if (group.isKana) { group.textNormalized = convertKatakanaToHiragana(group.text); }
-        }
-        const segments = segmentizeFurigana(reading, convertKatakanaToHiragana(reading), groups, 0);
-        return segments !== null ? segments : [createFuriganaSegment(term, reading)];
-      }
-      function getStemLength(text1, text2) {
-        const minLength = Math.min(text1.length, text2.length);
-        if (minLength === 0) { return 0; }
-        let i = 0;
-        while (true) {
-          const char1 = text1.codePointAt(i);
-          const char2 = text2.codePointAt(i);
-          if (char1 !== char2) { break; }
-          const charLength = String.fromCodePoint(char1).length;
-          i += charLength;
-          if (i >= minLength) {
-            if (i > minLength) { i -= charLength; }
-            break;
-          }
-        }
-        return i;
-      }
-      function distributeFuriganaInflected(term, reading, source) {
-        const termNormalized = convertKatakanaToHiragana(term);
-        const readingNormalized = convertKatakanaToHiragana(reading);
-        const sourceNormalized = convertKatakanaToHiragana(source);
-        let mainText = term;
-        let stemLength = getStemLength(termNormalized, sourceNormalized);
-        const readingStemLength = getStemLength(readingNormalized, sourceNormalized);
-        if (readingStemLength > 0 && readingStemLength >= stemLength) {
-          mainText = reading;
-          stemLength = readingStemLength;
-          reading = source.substring(0, stemLength) + reading.substring(stemLength);
-        }
-        const segments = [];
-        if (stemLength > 0) {
-          mainText = source.substring(0, stemLength) + mainText.substring(stemLength);
-          const segments2 = distributeFurigana(mainText, reading);
-          let consumed = 0;
-          for (const segment of segments2) {
-            const start = consumed;
-            consumed += segment.text.length;
-            if (consumed < stemLength) { segments.push(segment); }
-            else if (consumed === stemLength) { segments.push(segment); break; }
-            else {
-              if (start < stemLength) { segments.push(createFuriganaSegment(mainText.substring(start, stemLength), '')); }
-              break;
-            }
-          }
-        }
-        if (stemLength < source.length) {
-          const remainder = source.substring(stemLength);
-          const last = segments[segments.length - 1];
-          if (last && last.reading.length === 0) { last.text += remainder; }
-          else { segments.push(createFuriganaSegment(remainder, '')); }
-        }
-        return segments;
-      }
-      function parsePositiveFrequencyNumber(value) {
-        if (typeof value === 'number' && Number.isFinite(value) && value > 0) {
-          return Math.max(1, Math.floor(value));
-        }
-        if (typeof value === 'string') {
-          const numericMatch = value.trim().match(/[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?/)?.[0];
-          if (!numericMatch) { return null; }
-          const parsed = Number.parseFloat(numericMatch);
-          if (!Number.isFinite(parsed) || parsed <= 0) { return null; }
-          return Math.max(1, Math.floor(parsed));
-        }
-        if (Array.isArray(value)) {
-          for (const item of value) {
-            const parsed = parsePositiveFrequencyNumber(item);
-            if (parsed !== null) { return parsed; }
-          }
-        }
-        return null;
-      }
-      function parseDisplayFrequencyNumber(value) {
-        if (typeof value === 'string') {
-          const leadingDigits = value.trim().match(/^\d+/)?.[0];
-          if (!leadingDigits) { return null; }
-          const parsed = Number.parseInt(leadingDigits, 10);
-          return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
-        }
-        return parsePositiveFrequencyNumber(value);
-      }
-      function getFrequencyDictionaryName(frequency) {
-        const candidates = [
-          frequency?.dictionary,
-          frequency?.dictionaryName,
-          frequency?.name,
-          frequency?.title,
-          frequency?.dictionaryTitle,
-          frequency?.dictionaryAlias
-        ];
-        for (const candidate of candidates) {
-          if (typeof candidate === 'string' && candidate.trim().length > 0) {
-            return candidate.trim();
-          }
-        }
-        return null;
-      }
-      function getBestFrequencyRank(dictionaryEntry, headwordIndex, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
-        let best = null;
-        const headwordCount = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords.length : 0;
-        for (const frequency of dictionaryEntry?.frequencies || []) {
-          if (!frequency || typeof frequency !== 'object') { continue; }
-          const frequencyHeadwordIndex = frequency.headwordIndex;
-          if (typeof frequencyHeadwordIndex === 'number') {
-            if (frequencyHeadwordIndex !== headwordIndex) { continue; }
-          } else if (headwordCount > 1) {
-            continue;
-          }
-          const dictionary = getFrequencyDictionaryName(frequency);
-          if (!dictionary) { continue; }
-          if (dictionaryFrequencyModeByName[dictionary] === 'occurrence-based') { continue; }
-          const rank =
-            parseDisplayFrequencyNumber(frequency.displayValue) ??
-            parsePositiveFrequencyNumber(frequency.frequency);
-          if (rank === null) { continue; }
-          const priorityRaw = dictionaryPriorityByName[dictionary];
-          const fallbackPriority =
-            typeof frequency.dictionaryIndex === 'number' && Number.isFinite(frequency.dictionaryIndex)
-              ? Math.max(0, Math.floor(frequency.dictionaryIndex))
-              : Number.MAX_SAFE_INTEGER;
-          const priority =
-            typeof priorityRaw === 'number' && Number.isFinite(priorityRaw)
-              ? Math.max(0, Math.floor(priorityRaw))
-              : fallbackPriority;
-          if (best === null || priority < best.priority || (priority === best.priority && rank < best.rank)) {
-            best = { priority, rank };
-          }
-        }
-        return best?.rank ?? null;
-      }
-      function hasExactSource(headword, token, requirePrimary) {
-        for (const src of headword.sources || []) {
-          if (src.originalText !== token) { continue; }
-          if (requirePrimary && !src.isPrimary) { continue; }
-          if (src.matchType !== 'exact') { continue; }
-          return true;
-        }
-        return false;
-      }
-      function collectExactHeadwordMatches(dictionaryEntries, token, requirePrimary) {
-        const matches = [];
-        for (const dictionaryEntry of dictionaryEntries || []) {
-          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
-          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
-            const headword = headwords[headwordIndex];
-            if (!hasExactSource(headword, token, requirePrimary)) { continue; }
-            matches.push({ dictionaryEntry, headword, headwordIndex });
-          }
-        }
-        return matches;
-      }
-      function sameHeadword(match, preferredMatch) {
-        if (!match || !preferredMatch) {
-          return false;
-        }
-        if (match.headword?.term !== preferredMatch.headword?.term) {
-          return false;
-        }
-        const matchReading = typeof match.headword?.reading === 'string' ? match.headword.reading : '';
-        const preferredReading =
-          typeof preferredMatch.headword?.reading === 'string' ? preferredMatch.headword.reading : '';
-        if (!matchReading || !preferredReading) {
-          return true;
-        }
-        return matchReading === preferredReading;
-      }
-      function getBestFrequencyRankForMatches(matches, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
-        let best = null;
-        for (const match of matches) {
-          const rank = getBestFrequencyRank(
-            match.dictionaryEntry,
-            match.headwordIndex,
-            dictionaryPriorityByName,
-            dictionaryFrequencyModeByName
-          );
-          if (rank === null) { continue; }
-          if (best === null || rank < best) {
-            best = rank;
-          }
-        }
-        return best;
-      }
-      function normalizeWordClasses(headword) {
-          if (!Array.isArray(headword?.wordClasses)) { return undefined; }
-          const classes = headword.wordClasses.filter((wordClass) => typeof wordClass === "string" && wordClass.trim().length > 0);
-          return classes.length > 0 ? classes : undefined;
-        }
-        function appendDictionaryNames(target, value) {
-          if (!value || typeof value !== 'object') {
-            return;
-          }
-          const candidates = [
-            value.dictionary,
-            value.dictionaryName,
-            value.name,
-            value.title,
-            value.dictionaryTitle,
-            value.dictionaryAlias
-          ];
-          for (const candidate of candidates) {
-            if (typeof candidate === 'string' && candidate.trim().length > 0) {
-              target.push(candidate.trim());
-            }
-          }
-        }
-        function getDictionaryEntryNames(entry) {
-          const names = [];
-          appendDictionaryNames(names, entry);
-          for (const definition of entry?.definitions || []) {
-            appendDictionaryNames(names, definition);
-          }
-          for (const frequency of entry?.frequencies || []) {
-            appendDictionaryNames(names, frequency);
-          }
-          for (const pronunciation of entry?.pronunciations || []) {
-            appendDictionaryNames(names, pronunciation);
-          }
-          return names;
-        }
-        function isNameDictionaryEntry(entry) {
-          if (!includeNameMatchMetadata || !entry || typeof entry !== 'object') {
-            return false;
-          }
-          return getDictionaryEntryNames(entry).some((name) => name.startsWith(${JSON.stringify(CHARACTER_DICTIONARY_TITLE_PREFIX)}));
-        }
-        function parseSubMinerMediaIdFromString(value) {
-          const imageMatch = value.match(/\bimg\/m(\d+)-/i);
-          if (imageMatch) {
-            const parsed = Number.parseInt(imageMatch[1], 10);
-            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
-          }
-          const titleMatch = value.match(/${CHARACTER_DICTIONARY_TITLE_PREFIX}[^\d]*(?:AniList\s*)?(\d+)/i);
-          if (titleMatch) {
-            const parsed = Number.parseInt(titleMatch[1], 10);
-            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
-          }
-          return null;
-        }
-        function parseSubMinerMediaIdCandidate(value) {
-          if (typeof value === 'number' && Number.isSafeInteger(value) && value > 0) {
-            return value;
-          }
-          if (typeof value === 'string' && /^\d+$/.test(value.trim())) {
-            const parsed = Number.parseInt(value.trim(), 10);
-            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
-          }
-          return null;
-        }
-        function collectSubMinerMediaIds(value, target) {
-          if (typeof value === 'string') {
-            const parsed = parseSubMinerMediaIdFromString(value);
-            if (parsed !== null) { target.add(parsed); }
-            return;
-          }
-          if (!value || typeof value !== 'object') {
-            return;
-          }
-          if (Array.isArray(value)) {
-            for (const item of value) { collectSubMinerMediaIds(item, target); }
-            return;
-          }
-          const mediaIdCandidates = [
-            value.subminerMediaId,
-            value.subMinerMediaId,
-            value.characterDictionaryMediaId,
-            value.data?.subminerMediaId,
-            value.data?.subMinerMediaId,
-            value.data?.characterDictionaryMediaId
-          ];
-          for (const candidate of mediaIdCandidates) {
-            const parsed = parseSubMinerMediaIdCandidate(candidate);
-            if (parsed !== null) { target.add(parsed); }
-          }
-          for (const child of Object.values(value)) {
-            collectSubMinerMediaIds(child, target);
-          }
-        }
-        function getSubMinerMediaIds(entry) {
-          const mediaIds = new Set();
-          collectSubMinerMediaIds(entry, mediaIds);
-          return mediaIds;
-        }
-        function isCurrentMediaNameDictionaryEntry(entry) {
-          if (!isNameDictionaryEntry(entry)) {
-            return false;
-          }
-          if (currentCharacterDictionaryMediaId === null) {
-            return true;
-          }
-          const mediaIds = getSubMinerMediaIds(entry);
-          return mediaIds.size === 0 || mediaIds.has(currentCharacterDictionaryMediaId);
-      }
-      function findLongestNameMatch(dictionaryEntries, textWindow) {
-        let best = null;
-        for (const dictionaryEntry of dictionaryEntries || []) {
-          if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
-          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
-          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
-            const headword = headwords[headwordIndex];
-            for (const src of headword?.sources || []) {
-              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
-              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
-              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
-              if (best === null || originalText.length > best.sourceLength) {
-                best = { dictionaryEntry, headword, headwordIndex, sourceLength: originalText.length };
-              }
-            }
-          }
-        }
-        return best;
-      }
-      function findLongestGenericMatchLength(dictionaryEntries, textWindow) {
-        let best = 0;
-        for (const dictionaryEntry of dictionaryEntries || []) {
-          if (isNameDictionaryEntry(dictionaryEntry)) { continue; }
-          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
-          for (const headword of headwords) {
-            for (const src of headword?.sources || []) {
-              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
-              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
-              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
-              if (originalText.length > best) { best = originalText.length; }
-            }
-          }
-        }
-        return best;
-      }
-      function getPreferredHeadword(dictionaryEntries, token, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
-        const currentMediaDictionaryEntries =
-          currentCharacterDictionaryMediaId === null
-            ? (dictionaryEntries || [])
-            : (dictionaryEntries || []).filter((entry) => {
-                if (!isNameDictionaryEntry(entry)) { return true; }
-                return isCurrentMediaNameDictionaryEntry(entry);
-              });
-        const exactPrimaryMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, true);
-        let matchedNameDictionary = false;
-        if (includeNameMatchMetadata) {
-          for (const dictionaryEntry of currentMediaDictionaryEntries || []) {
-            if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
-            for (const match of exactPrimaryMatches) {
-              if (match.dictionaryEntry !== dictionaryEntry) { continue; }
-              matchedNameDictionary = true;
-              break;
-            }
-            if (matchedNameDictionary) { break; }
-          }
-        }
-        const preferredMatch = exactPrimaryMatches[0];
-        if (preferredMatch) {
-          const exactFrequencyMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, false)
-            .filter((match) => sameHeadword(match, preferredMatch));
-          return {
-            term: preferredMatch.headword.term,
-            reading: preferredMatch.headword.reading,
-            wordClasses: normalizeWordClasses(preferredMatch.headword),
-            isNameMatch:
-              matchedNameDictionary || isCurrentMediaNameDictionaryEntry(preferredMatch.dictionaryEntry),
-            frequencyRank: getBestFrequencyRankForMatches(
-              exactFrequencyMatches.length > 0 ? exactFrequencyMatches : exactPrimaryMatches,
-              dictionaryPriorityByName,
-              dictionaryFrequencyModeByName
-            )
-          };
-        }
-        return null;
-      }
-`;
 
-function buildYomitanScanningScript(
+// Key of the character-name candidate list currently installed in each parser
-  text: string,
+// window, so an unchanged list costs nothing per line.
-  profileIndex: number,
+const yomitanScanNameCandidateKeyByWindow = new WeakMap<BrowserWindow, string>();
-  scanLength: number,
+
-  includeNameMatchMetadata: boolean,
+async function ensureYomitanScanNameCandidates(
-  greedyNameScanEnabled: boolean,
+  parserWindow: BrowserWindow,
-  currentCharacterDictionaryMediaId: number | null,
+  nameCandidates: { key: string; forms: string[] } | null,
-  dictionaryPriorityByName: Record<string, number>,
+  logger: LoggerLike,
-  dictionaryFrequencyModeByName: Partial<Record<string, YomitanFrequencyMode>>,
+): Promise<void> {
-): string {
+  const installedKey = yomitanScanNameCandidateKeyByWindow.get(parserWindow);
-  return `
+  const nextKey = nameCandidates?.key ?? '';
-    (async () => {
+  if (installedKey === nextKey) {
-      const invoke = (action, params) =>
-        new Promise((resolve, reject) => {
-          chrome.runtime.sendMessage({ action, params }, (response) => {
-            if (chrome.runtime.lastError) {
-              reject(new Error(chrome.runtime.lastError.message));
     return;
   }
-            if (!response || typeof response !== "object") {
+
-              reject(new Error("Invalid response from Yomitan backend"));
+  try {
-              return;
+    await parserWindow.webContents.executeJavaScript(
-            }
+      buildYomitanScanNameCandidatesScript(nameCandidates),
-            if (response.error) {
+      true,
-              reject(new Error(response.error.message || "Yomitan backend error"));
-              return;
-            }
-            resolve(response.result);
-          });
-        });
-${YOMITAN_SCANNING_HELPERS}
-      const includeNameMatchMetadata = ${includeNameMatchMetadata ? 'true' : 'false'};
-      const greedyNameScanEnabled = ${greedyNameScanEnabled ? 'true' : 'false'};
-      const currentCharacterDictionaryMediaId = ${
-        currentCharacterDictionaryMediaId !== null
-          ? String(currentCharacterDictionaryMediaId)
-          : 'null'
-      };
-      const dictionaryPriorityByName = ${JSON.stringify(dictionaryPriorityByName)};
-      const dictionaryFrequencyModeByName = ${JSON.stringify(dictionaryFrequencyModeByName)};
-      const text = ${JSON.stringify(text)};
-      const details = {matchType: "exact", deinflect: true};
-      const tokens = [];
-      const termsFindCache = new Map();
-      async function termsFindAt(position, windowLength) {
-        const cacheKey = position + ":" + windowLength;
-        const cached = termsFindCache.get(cacheKey);
-        if (cached) { return cached; }
-        const substring = text.substring(position, position + windowLength);
-        const result = await invoke("termsFind", { text: substring, details, optionsContext: { index: ${profileIndex} } });
-        termsFindCache.set(cacheKey, result);
-        return result;
-      }
-      function buildScanToken(position, source, preferredHeadword) {
-        const reading = typeof preferredHeadword.reading === "string" ? preferredHeadword.reading : "";
-        const segments = distributeFuriganaInflected(preferredHeadword.term, reading, source);
-        const tokenPayload = {
-          surface: segments.map((segment) => segment.text).join("") || source,
-          reading: segments.map(getSegmentReadingContribution).join(""),
-          headword: preferredHeadword.term,
-          headwordReading: reading || undefined,
-          startPos: position,
-          endPos: position + source.length,
-          isNameMatch: includeNameMatchMetadata && preferredHeadword.isNameMatch === true,
-          frequencyRank:
-            typeof preferredHeadword.frequencyRank === "number" && Number.isFinite(preferredHeadword.frequencyRank)
-              ? Math.max(1, Math.floor(preferredHeadword.frequencyRank))
-              : undefined,
-        };
-        if (Array.isArray(preferredHeadword.wordClasses) && preferredHeadword.wordClasses.length > 0) {
-          tokenPayload.wordClasses = preferredHeadword.wordClasses;
-        }
-        return tokenPayload;
-      }
-      async function findTokenAt(position, windowLength) {
-        const codePoint = text.codePointAt(position);
-        const character = String.fromCodePoint(codePoint);
-        const result = await termsFindAt(position, windowLength);
-        const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
-        const originalTextLength = typeof result?.originalTextLength === "number" ? result.originalTextLength : 0;
-        if (dictionaryEntries.length === 0 || originalTextLength <= 0 || (originalTextLength === character.length && !isCodePointJapanese(codePoint))) {
-          return { token: null, matchedLength: 0 };
-        }
-        const source = text.substring(position, position + originalTextLength);
-        const preferredHeadword = getPreferredHeadword(
-          dictionaryEntries,
-          source,
-          dictionaryPriorityByName,
-          dictionaryFrequencyModeByName
     );
-        if (!preferredHeadword || typeof preferredHeadword.term !== "string") {
+    yomitanScanNameCandidateKeyByWindow.set(parserWindow, nextKey);
-          return { token: null, matchedLength: originalTextLength };
+  } catch (err) {
+    // The scan falls back to checking every position when the list is absent,
+    // so a failed install costs speed, never a missed name.
+    logger.warn?.(
+      'Failed to install Yomitan character-name scan candidates:',
+      (err as Error).message,
+    );
+    yomitanScanNameCandidateKeyByWindow.delete(parserWindow);
   }
-        return { token: buildScanToken(position, source, preferredHeadword), matchedLength: originalTextLength };
-      }
-      // Greedy name pre-pass: character-name matches claim their spans before
-      // the left-to-right walk, so a longer generic match starting earlier
-      // (e.g. とヨー → 渡洋) cannot swallow the start of a name (ヨータ).
-      const nameTokens = [];
-      if (greedyNameScanEnabled) {
-        let namePos = 0;
-        while (namePos < text.length) {
-          const codePoint = text.codePointAt(namePos);
-          if (!isCodePointJapanese(codePoint)) {
-            namePos += String.fromCodePoint(codePoint).length;
-            continue;
-          }
-          const result = await termsFindAt(namePos, ${scanLength});
-          const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
-          const textWindow = text.substring(namePos, namePos + ${scanLength});
-          const nameMatch = findLongestNameMatch(dictionaryEntries, textWindow);
-          // A name only claims its span when no strictly longer generic word
-          // starts at the same position (a character named 空 must not split
-          // 空気). Ties go to the name. Generic matches that start earlier and
-          // overlap the name are still blocked by the reservation.
-          if (
-            !nameMatch ||
-            findLongestGenericMatchLength(dictionaryEntries, textWindow) > nameMatch.sourceLength
-          ) {
-            namePos += String.fromCodePoint(codePoint).length;
-            continue;
-          }
-          const source = text.substring(namePos, namePos + nameMatch.sourceLength);
-          nameTokens.push(buildScanToken(namePos, source, {
-            term: nameMatch.headword.term,
-            reading: nameMatch.headword.reading,
-            wordClasses: normalizeWordClasses(nameMatch.headword),
-            isNameMatch: true,
-            frequencyRank: getBestFrequencyRank(
-              nameMatch.dictionaryEntry,
-              nameMatch.headwordIndex,
-              dictionaryPriorityByName,
-              dictionaryFrequencyModeByName
-            )
-          }));
-          namePos += nameMatch.sourceLength;
-        }
-      }
-      let i = 0;
-      let nameIndex = 0;
-      while (i < text.length) {
-        while (nameIndex < nameTokens.length && nameTokens[nameIndex].startPos < i) { nameIndex += 1; }
-        const nextNameToken = nameIndex < nameTokens.length ? nameTokens[nameIndex] : null;
-        if (nextNameToken && nextNameToken.startPos === i) {
-          tokens.push(nextNameToken);
-          i = nextNameToken.endPos;
-          nameIndex += 1;
-          continue;
-        }
-        // Cap the window at the next reserved name span so a generic match
-        // cannot consume into it.
-        const windowLength = nextNameToken ? Math.min(${scanLength}, nextNameToken.startPos - i) : ${scanLength};
-        let attempt = await findTokenAt(i, windowLength);
-        // Yomitan text normalization can consume characters (whitespace,
-        // punctuation) beyond the matched term, leaving no headword whose
-        // source equals the consumed text. Retry with shorter windows so a
-        // valid prefix term (e.g. a character name before a paren) still
-        // tokenizes instead of the position being skipped.
-        let retryLength = Math.min(attempt.matchedLength, windowLength) - 1;
-        while (!attempt.token && retryLength >= 1) {
-          const retry = await findTokenAt(i, retryLength);
-          if (retry.token) {
-            attempt = retry;
-            break;
-          }
-          retryLength = Math.min(retryLength - 1, retry.matchedLength - 1);
-        }
-        if (attempt.token) {
-          tokens.push(attempt.token);
-          i += attempt.matchedLength;
-          continue;
-        }
-        i += String.fromCodePoint(text.codePointAt(i)).length;
-      }
-      return tokens;
-    })();
-  `;
 }
 
 export async function requestYomitanParseResults(
@@ -1635,6 +963,20 @@ export async function requestYomitanParseResults(
   }
 }
 
+// parseText fallback for when the in-window scanner cannot run (script eval
+// failure, unexpected payload). The scanner walk is the primary tokenizer and
+// emits its own filler runs, so this extra full parse only happens on errors.
+async function requestYomitanParseFallbackTokens(
+  text: string,
+  deps: YomitanParserRuntimeDeps,
+  logger: LoggerLike,
+): Promise<YomitanScanToken[] | null> {
+  const parseResults = await requestYomitanParseResults(text, deps, logger);
+  const selectedTokens = selectYomitanParseTokens(parseResults, () => false, 'headword');
+  const parseScanTokens = selectedTokens?.map(toYomitanScanToken) ?? null;
+  return parseScanTokens && parseScanTokens.length > 0 ? parseScanTokens : null;
+}
+
 export async function requestYomitanScanTokens(
   text: string,
   deps: YomitanParserRuntimeDeps,
@@ -1642,6 +984,7 @@ export async function requestYomitanScanTokens(
   options?: {
     includeNameMatchMetadata?: boolean;
     currentCharacterDictionaryMediaId?: number | null;
+    nameCandidates?: { key: string; forms: string[] } | null;
   },
 ): Promise<YomitanScanToken[] | null> {
   const yomitanExt = deps.getYomitanExt();
@@ -1655,10 +998,6 @@ export async function requestYomitanScanTokens(
     return null;
   }
 
-  const parseResults = await requestYomitanParseResults(text, deps, logger);
-  const selectedParseTokens = selectYomitanParseTokens(parseResults, () => false, 'headword');
-  const parseScanTokens = selectedParseTokens?.map(toYomitanScanToken) ?? null;
-
   const metadata = await requestYomitanProfileMetadata(parserWindow, logger);
   const profileIndex = metadata?.profileIndex ?? 0;
   const scanLength = metadata?.scanLength ?? DEFAULT_YOMITAN_SCAN_LENGTH;
@@ -1669,44 +1008,63 @@ export async function requestYomitanScanTokens(
       name.startsWith(CHARACTER_DICTIONARY_TITLE_PREFIX),
     );
 
-  try {
+  // Candidate name forms let the in-page pre-pass skip positions where no
-    const rawResult = await parserWindow.webContents.executeJavaScript(
+  // character name can start. Installed only when it changes (per media), so
-      buildYomitanScanningScript(
+  // the per-line call stays a single tiny script.
+  const nameCandidates = greedyNameScanEnabled ? (options?.nameCandidates ?? null) : null;
+  await ensureYomitanScanNameCandidates(parserWindow, nameCandidates, logger);
+
+  const callScript = buildYomitanScanCallScript({
     text,
     profileIndex,
     scanLength,
     includeNameMatchMetadata,
     greedyNameScanEnabled,
+    currentCharacterDictionaryMediaId:
       typeof options?.currentCharacterDictionaryMediaId === 'number' &&
       Number.isFinite(options.currentCharacterDictionaryMediaId) &&
       options.currentCharacterDictionaryMediaId > 0
         ? Math.floor(options.currentCharacterDictionaryMediaId)
         : null,
-        metadata?.dictionaryPriorityByName ?? {},
+    dictionaryPriorityByName: metadata?.dictionaryPriorityByName ?? {},
-        metadata?.dictionaryFrequencyModeByName ?? {},
+    dictionaryFrequencyModeByName: metadata?.dictionaryFrequencyModeByName ?? {},
-      ),
+    cacheEpoch: getYomitanScanCacheEpoch(parserWindow),
-      true,
+    nameCandidateKey: nameCandidates?.key ?? null,
-    );
+  });
+
+  try {
+    let rawResult = await parserWindow.webContents.executeJavaScript(callScript, true);
+    if (rawResult === YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL) {
+      // First request for this window, or the page reloaded and dropped the
+      // installed runtime: install and retry once. The candidate list lives in
+      // the same page state, so it has to be reinstalled alongside it.
+      await installYomitanScanRuntime(parserWindow);
+      await ensureYomitanScanNameCandidates(parserWindow, nameCandidates, logger);
+      rawResult = await parserWindow.webContents.executeJavaScript(callScript, true);
+    }
+    // The scanner reports a line where a position ran out of shrinking-window
+    // retries: it stopped short of windows an uncapped ladder would have tried,
+    // so a real term may be sitting in an unparsed run. One parseText for the
+    // line is the bounded way to get the exhaustive answer back (this is the
+    // parse the scanner replaced, and it only runs for these rare lines).
+    if (isObject(rawResult) && rawResult.retryBudgetExhausted === true) {
+      logger.info?.('Yomitan scanner exhausted its retry budget; parsing the line as a fallback.');
+      const fallbackTokens = await requestYomitanParseFallbackTokens(text, deps, logger);
+      if (fallbackTokens) {
+        return fallbackTokens;
+      }
+      rawResult = rawResult.tokens;
+    }
     if (isScanTokenArray(rawResult)) {
-      if (parseScanTokens && parseScanTokens.length > 0) {
+      // Filler-only results carry no dictionary match; keep the historical
-        return mergeScannerTokensIntoParseTokens(parseScanTokens, rawResult);
+      // contract of returning null so callers fall back to raw text.
+      return rawResult.some((token) => token.isUnparsedRun !== true) ? rawResult : null;
     }
-      return rawResult;
+    logger.error('Yomitan scanner returned an unexpected payload; using parseText fallback.');
-    }
+    return await requestYomitanParseFallbackTokens(text, deps, logger);
-    if (Array.isArray(rawResult)) {
-      const selectedTokens = selectYomitanParseTokens(rawResult, () => false, 'headword');
-      return selectedTokens?.map(toYomitanScanToken) ?? null;
-    }
-    if (parseScanTokens && parseScanTokens.length > 0) {
-      return parseScanTokens;
-    }
-    return null;
   } catch (err) {
-    if (parseScanTokens && parseScanTokens.length > 0) {
-      return parseScanTokens;
-    }
     logger.error('Yomitan scanner request failed:', (err as Error).message);
-    return null;
+    return await requestYomitanParseFallbackTokens(text, deps, logger);
   }
 }
 

src/core/services/tokenizer/yomitan-scan-runtime-script.ts

@@ -0,0 +1,513 @@
+// In-page Yomitan scan runtime: the scan walk that gets installed once per
+// parser window as globalThis.__subminerYomitanScan, plus the tiny per-line
+// call script. Kept separate from the host runtime module so the injected
+// script text (which is data, not executed here) does not dominate that file;
+// the helper bundle it embeds lives in yomitan-scanning-helpers-script.ts.
+import { YOMITAN_SCANNING_HELPERS } from './yomitan-scanning-helpers-script';
+
+export { CHARACTER_DICTIONARY_TITLE_PREFIX } from './yomitan-scanning-helpers-script';
+
+export type YomitanFrequencyMode = 'occurrence-based' | 'rank-based';
+
+// Bump whenever the install script below changes so already-loaded parser
+// windows re-install the new scan runtime instead of running the stale one.
+export const YOMITAN_SCAN_RUNTIME_VERSION = 6;
+export const YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL = '__subminer-yomitan-scan-runtime-missing__';
+
+export interface YomitanScanRequestParams {
+  text: string;
+  profileIndex: number;
+  scanLength: number;
+  includeNameMatchMetadata: boolean;
+  greedyNameScanEnabled: boolean;
+  currentCharacterDictionaryMediaId: number | null;
+  dictionaryPriorityByName: Record<string, number>;
+  dictionaryFrequencyModeByName: Partial<Record<string, YomitanFrequencyMode>>;
+  cacheEpoch: number;
+  /**
+   * Key of the character-name candidate list installed for the current media,
+   * or null to scan every Japanese position (see the pre-pass prefilter).
+   */
+  nameCandidateKey: string | null;
+}
+
+// Installed once per parser window (and re-installed after in-page reloads):
+// keeps V8 from re-parsing the helper bundle on every subtitle line, and hosts
+// the cross-line termsFind cache. Each subtitle line then only evaluates a tiny
+// call into globalThis.__subminerYomitanScan.
+export const YOMITAN_SCAN_RUNTIME_INSTALL_SCRIPT = String.raw`
+  (() => {
+    if (globalThis.__subminerYomitanScanVersion === ${YOMITAN_SCAN_RUNTIME_VERSION}) {
+      return true;
+    }
+    const invoke = (action, params) =>
+      new Promise((resolve, reject) => {
+        chrome.runtime.sendMessage({ action, params }, (response) => {
+          if (chrome.runtime.lastError) {
+            reject(new Error(chrome.runtime.lastError.message));
+            return;
+          }
+          if (!response || typeof response !== "object") {
+            reject(new Error("Invalid response from Yomitan backend"));
+            return;
+          }
+          if (response.error) {
+            reject(new Error(response.error.message || "Yomitan backend error"));
+            return;
+          }
+          resolve(response.result);
+        });
+      });
+    // Cross-line termsFind LRU keyed by profile + substring: subtitle lines
+    // repeat particles and inflections constantly, so most lookups hit here.
+    // Entries hold in-flight promises so concurrent identical lookups dedupe.
+    const termsFindCache = new Map();
+    // Two bounds. The key count keeps the map itself small; the accumulated
+    // dictionary-entry count stands in for retained bytes, because a single
+    // lookup over a common prefix can hold hundreds of entries with their full
+    // glossaries and a key-count cap alone would not bound that.
+    const TERMS_FIND_CACHE_LIMIT = 2000;
+    const TERMS_FIND_CACHE_DICTIONARY_ENTRY_LIMIT = 20000;
+    let termsFindCacheDictionaryEntries = 0;
+    let termsFindCacheEpoch = -1;
+    function dropCachedTermsFind(cacheKey, entry) {
+      if (termsFindCache.get(cacheKey) !== entry) { return; }
+      termsFindCache.delete(cacheKey);
+      termsFindCacheDictionaryEntries -= entry.dictionaryEntryCount;
+    }
+    // Runs on insert and again once a lookup resolves: an entry is only worth
+    // its estimated weight of 1 until then, so a single oversized response
+    // would otherwise sit in the cache forever, over the limit and reused.
+    function evictOverflowingTermsFindEntries() {
+      while (
+        termsFindCache.size > TERMS_FIND_CACHE_LIMIT ||
+        termsFindCacheDictionaryEntries > TERMS_FIND_CACHE_DICTIONARY_ENTRY_LIMIT
+      ) {
+        const oldest = termsFindCache.entries().next().value;
+        if (oldest === undefined) { break; }
+        dropCachedTermsFind(oldest[0], oldest[1]);
+      }
+    }
+    // Only blind ladder steps are capped (see the retry loop): those are the
+    // ones that would otherwise degrade into O(scanLength) lookups at a single
+    // position. Steps the backend guides by reporting a shorter consumed length
+    // stay uncapped, so a valid prefix term is still found on lines where
+    // normalization eats a long tail.
+    const MAX_BLIND_SHRINKING_WINDOW_RETRIES = 4;
+    // Character-name candidate forms for the current media, installed
+    // separately from the per-line scan call so the per-line script stays tiny.
+    // Stored raw here; the normalized lookup index is built inside the scan,
+    // where the kana-normalization helper is in scope, and reused by key.
+    let rawNameCandidates = null;
+    let nameCandidateIndex = null;
+    globalThis.__subminerYomitanScanSetNameCandidates = (key, forms) => {
+      if (!key || !Array.isArray(forms) || forms.length === 0) {
+        rawNameCandidates = null;
+        nameCandidateIndex = null;
+        return false;
+      }
+      rawNameCandidates = { key, forms };
+      nameCandidateIndex = null;
+      return true;
+    };
+    globalThis.__subminerYomitanScanVersion = ${YOMITAN_SCAN_RUNTIME_VERSION};
+    globalThis.__subminerYomitanScan = async (scanParams) => {
+      const {
+        text,
+        profileIndex,
+        scanLength,
+        includeNameMatchMetadata,
+        greedyNameScanEnabled,
+        currentCharacterDictionaryMediaId,
+        dictionaryPriorityByName,
+        dictionaryFrequencyModeByName,
+        cacheEpoch,
+        nameCandidateKey
+      } = scanParams;
+      if (cacheEpoch !== termsFindCacheEpoch) {
+        termsFindCache.clear();
+        termsFindCacheDictionaryEntries = 0;
+        termsFindCacheEpoch = cacheEpoch;
+      }
+${YOMITAN_SCANNING_HELPERS}
+      const CAPTION_OPENING_BRACKETS = new Set(["(", "（", "[", "［", "{", "｛", "「", "『", "【", "〈", "《", "≪", "＜", "<"]);
+      function shouldEmitUnparsedRunAsToken(runText) {
+        if (!/[\p{L}\p{N}]/u.test(runText)) { return false; }
+        const firstChar = Array.from(runText.trim())[0];
+        return firstChar !== undefined && !CAPTION_OPENING_BRACKETS.has(firstChar);
+      }
+      function isLookupWorthyCodePoint(codePoint) {
+        if (isCodePointJapanese(codePoint)) { return true; }
+        return /[\p{L}\p{N}]/u.test(String.fromCodePoint(codePoint));
+      }
+      function isKanaOnlyRunText(runText) {
+        const chars = Array.from(runText);
+        return chars.length > 0 && chars.every((char) => isCodePointKana(char.codePointAt(0)));
+      }
+      const details = {matchType: "exact", deinflect: true};
+      const tokens = [];
+      async function termsFindAt(position, windowLength) {
+        const substring = text.substring(position, position + windowLength);
+        const cacheKey = profileIndex + "\u0000" + substring;
+        const cached = termsFindCache.get(cacheKey);
+        if (cached !== undefined) {
+          termsFindCache.delete(cacheKey);
+          termsFindCache.set(cacheKey, cached);
+          return await cached.promise;
+        }
+        // An in-flight lookup counts as one entry until it resolves; the real
+        // weight replaces that estimate once the result is known.
+        const entry = { promise: null, dictionaryEntryCount: 1 };
+        entry.promise = invoke("termsFind", { text: substring, details, optionsContext: { index: profileIndex } })
+          .then((result) => {
+            const resolvedCount =
+              1 + (Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries.length : 0);
+            const isCached = termsFindCache.get(cacheKey) === entry;
+            if (isCached) {
+              termsFindCacheDictionaryEntries += resolvedCount - entry.dictionaryEntryCount;
+            }
+            entry.dictionaryEntryCount = resolvedCount;
+            // The real weight can push the cache over its budget, and a single
+            // response can exceed it on its own, so re-check here.
+            if (isCached) { evictOverflowingTermsFindEntries(); }
+            return result;
+          });
+        termsFindCache.set(cacheKey, entry);
+        termsFindCacheDictionaryEntries += entry.dictionaryEntryCount;
+        evictOverflowingTermsFindEntries();
+        try {
+          return await entry.promise;
+        } catch (error) {
+          dropCachedTermsFind(cacheKey, entry);
+          throw error;
+        }
+      }
+      // Text the walk skips accumulates into unparsed runs, mirroring the
+      // filler chunks the parseText segmentation used to provide: runs stay
+      // hoverable (flagged isUnparsedRun) unless they are punctuation-only or
+      // caption-style asides, and kana continuations of a longer headword
+      // extend the previous token instead.
+      function flushUnparsedRun(runStart, runEnd) {
+        if (runStart === null || runEnd <= runStart) { return; }
+        const runText = text.substring(runStart, runEnd);
+        const previousToken = tokens[tokens.length - 1];
+        if (
+          previousToken &&
+          previousToken.endPos === runStart &&
+          isKanaOnlyRunText(runText) &&
+          typeof previousToken.headword === "string" &&
+          previousToken.headword.length > previousToken.surface.length &&
+          previousToken.headword.startsWith(previousToken.surface + runText)
+        ) {
+          previousToken.surface += runText;
+          // The run is kana-only, so its reading is itself: append it or the
+          // reading stops covering the surface, which disables the known-word
+          // reading fallback (isCompleteReadingForSurface) downstream.
+          previousToken.reading += runText;
+          // The run is kana-only, so its reading is itself: append it or the
+          // reading stops covering the surface, which disables the known-word
+          // reading fallback (isCompleteReadingForSurface) downstream.
+          previousToken.endPos = runEnd;
+          return;
+        }
+        if (!shouldEmitUnparsedRunAsToken(runText)) { return; }
+        tokens.push({
+          surface: runText,
+          reading: "",
+          headword: runText,
+          startPos: runStart,
+          endPos: runEnd,
+          isUnparsedRun: true
+        });
+      }
+      function buildScanToken(position, source, preferredHeadword) {
+        const reading = typeof preferredHeadword.reading === "string" ? preferredHeadword.reading : "";
+        const segments = distributeFuriganaInflected(preferredHeadword.term, reading, source);
+        const tokenPayload = {
+          surface: segments.map((segment) => segment.text).join("") || source,
+          reading: segments.map(getSegmentReadingContribution).join(""),
+          headword: preferredHeadword.term,
+          headwordReading: reading || undefined,
+          startPos: position,
+          endPos: position + source.length,
+          isNameMatch: includeNameMatchMetadata && preferredHeadword.isNameMatch === true,
+          frequencyRank:
+            typeof preferredHeadword.frequencyRank === "number" && Number.isFinite(preferredHeadword.frequencyRank)
+              ? Math.max(1, Math.floor(preferredHeadword.frequencyRank))
+              : undefined,
+        };
+        if (Array.isArray(preferredHeadword.wordClasses) && preferredHeadword.wordClasses.length > 0) {
+          tokenPayload.wordClasses = preferredHeadword.wordClasses;
+        }
+        return tokenPayload;
+      }
+      // findTokenAt plus the shrinking-window ladder below it: Yomitan text
+      // normalization can consume characters (whitespace, punctuation) beyond
+      // the matched term, leaving no headword whose source equals the consumed
+      // text. Retry with shorter windows so a valid prefix term (e.g. a
+      // character name before a paren) still tokenizes instead of the position
+      // being skipped.
+      // Every window at or above the consumed length repeats the same result,
+      // so the next informative window sits just below it. A lookup that
+      // consumed its whole window reports nothing to aim at, and the step down
+      // from it is a blind guess: only those are budgeted.
+      // The window can run past the end of the line, so blindness is judged
+      // against the text the lookup actually saw.
+      // Set when a position stopped short of windows an uncapped ladder would
+      // still have tried; the line then escalates to parseText at the end.
+      let blindRetryBudgetExhausted = false;
+      async function resolveTokenAt(position, windowLength) {
+        let attempt = await findTokenAt(position, windowLength);
+        const scannedLength = Math.min(windowLength, text.length - position);
+        let retryLength = Math.min(attempt.matchedLength, scannedLength) - 1;
+        let stepIsBlind = attempt.matchedLength >= scannedLength;
+        let blindRetriesRemaining = MAX_BLIND_SHRINKING_WINDOW_RETRIES;
+        while (!attempt.token && retryLength >= 1) {
+          if (stepIsBlind) {
+            if (blindRetriesRemaining <= 0) {
+              blindRetryBudgetExhausted = true;
+              break;
+            }
+            blindRetriesRemaining -= 1;
+          }
+          const retry = await findTokenAt(position, retryLength);
+          if (retry.token) { return retry; }
+          const guidedLength = retry.matchedLength - 1;
+          stepIsBlind = guidedLength >= retryLength - 1;
+          retryLength = Math.min(retryLength - 1, guidedLength);
+        }
+        return attempt;
+      }
+      async function findTokenAt(position, windowLength) {
+        const codePoint = text.codePointAt(position);
+        const character = String.fromCodePoint(codePoint);
+        const result = await termsFindAt(position, windowLength);
+        const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
+        const originalTextLength = typeof result?.originalTextLength === "number" ? result.originalTextLength : 0;
+        if (dictionaryEntries.length === 0 || originalTextLength <= 0 || (originalTextLength === character.length && !isCodePointJapanese(codePoint))) {
+          return { token: null, matchedLength: 0 };
+        }
+        const source = text.substring(position, position + originalTextLength);
+        const preferredHeadword = getPreferredHeadword(
+          dictionaryEntries,
+          source,
+          dictionaryPriorityByName,
+          dictionaryFrequencyModeByName
+        );
+        if (!preferredHeadword || typeof preferredHeadword.term !== "string") {
+          return { token: null, matchedLength: originalTextLength };
+        }
+        return { token: buildScanToken(position, source, preferredHeadword), matchedLength: originalTextLength };
+      }
+      // Halfwidth katakana survives kana normalization unchanged, so a name
+      // written that way would not prefix-match a candidate form. Those
+      // positions bypass the prefilter rather than risk a missed name.
+      function isHalfwidthKatakanaCodePoint(codePoint) {
+        return codePoint >= 0xff66 && codePoint <= 0xff9f;
+      }
+      // Build (once per candidate list) a first-character bucket index of the
+      // normalized name forms, so the pre-pass can reject a position with a
+      // single map hit instead of a backend round trip.
+      if (rawNameCandidates && nameCandidateIndex?.key !== rawNameCandidates.key) {
+        const byFirstChar = new Map();
+        for (const form of rawNameCandidates.forms) {
+          const normalized = typeof form === "string" ? convertKatakanaToHiragana(form.trim()) : "";
+          if (!normalized) { continue; }
+          const bucket = byFirstChar.get(normalized[0]);
+          if (bucket) { bucket.push(normalized); } else { byFirstChar.set(normalized[0], [normalized]); }
+        }
+        nameCandidateIndex = byFirstChar.size > 0 ? { key: rawNameCandidates.key, byFirstChar } : null;
+      } else if (!rawNameCandidates) {
+        nameCandidateIndex = null;
+      }
+      // Only meaningful when the installed list matches the media this scan is
+      // for; otherwise fall back to scanning every position.
+      const activeNameCandidateIndex =
+        nameCandidateKey !== null && nameCandidateIndex?.key === nameCandidateKey
+          ? nameCandidateIndex
+          : null;
+      const normalizedText = activeNameCandidateIndex ? convertKatakanaToHiragana(text) : "";
+      // Yomitan collapses emphatic sequences before matching (すっっごーーい →
+      // すごい), so a stretched name still resolves to its entry. Skipping these
+      // characters keeps such spellings candidates; the filter only ever grows
+      // the probe set, so a false positive costs one lookup, never a name.
+      const EMPHATIC_SKIP_CHARS = new Set(["ぁ", "ぃ", "ぅ", "ぇ", "ぉ", "っ", "ゃ", "ゅ", "ょ", "ー"]);
+      function matchesCandidateFormAt(form, position) {
+        let textIndex = position;
+        for (let formIndex = 0; formIndex < form.length; formIndex += 1) {
+          while (
+            textIndex < normalizedText.length &&
+            normalizedText[textIndex] !== form[formIndex] &&
+            EMPHATIC_SKIP_CHARS.has(normalizedText[textIndex])
+          ) {
+            textIndex += 1;
+          }
+          if (normalizedText[textIndex] !== form[formIndex]) { return false; }
+          textIndex += 1;
+        }
+        return true;
+      }
+      function couldNameStartAt(position, codePoint) {
+        if (!activeNameCandidateIndex) { return true; }
+        if (isHalfwidthKatakanaCodePoint(codePoint)) { return true; }
+        const bucket = activeNameCandidateIndex.byFirstChar.get(normalizedText[position]);
+        if (!bucket) { return false; }
+        for (const form of bucket) {
+          if (matchesCandidateFormAt(form, position)) { return true; }
+        }
+        return false;
+      }
+      // Greedy name pre-pass: character-name matches claim their spans before
+      // the left-to-right walk, so a longer generic match starting earlier
+      // (e.g. とヨー → 渡洋) cannot swallow the start of a name (ヨータ).
+      const nameTokens = [];
+      if (greedyNameScanEnabled) {
+        let namePos = 0;
+        while (namePos < text.length) {
+          const codePoint = text.codePointAt(namePos);
+          if (!isCodePointJapanese(codePoint) || !couldNameStartAt(namePos, codePoint)) {
+            namePos += String.fromCodePoint(codePoint).length;
+            continue;
+          }
+          const result = await termsFindAt(namePos, scanLength);
+          const dictionaryEntries = Array.isArray(result?.dictionaryEntries) ? result.dictionaryEntries : [];
+          const textWindow = text.substring(namePos, namePos + scanLength);
+          const nameMatch = findLongestNameMatch(dictionaryEntries, textWindow);
+          // A name only claims its span when no strictly longer generic word
+          // starts at the same position (a character named 空 must not split
+          // 空気). Ties go to the name. Generic matches that start earlier and
+          // overlap the name are still blocked by the reservation.
+          if (
+            !nameMatch ||
+            findLongestGenericMatchLength(dictionaryEntries, textWindow) > nameMatch.sourceLength
+          ) {
+            namePos += String.fromCodePoint(codePoint).length;
+            continue;
+          }
+          const source = text.substring(namePos, namePos + nameMatch.sourceLength);
+          nameTokens.push(buildScanToken(namePos, source, {
+            term: nameMatch.headword.term,
+            reading: nameMatch.headword.reading,
+            wordClasses: normalizeWordClasses(nameMatch.headword),
+            isNameMatch: true,
+            frequencyRank: getBestFrequencyRank(
+              nameMatch.dictionaryEntry,
+              nameMatch.headwordIndex,
+              dictionaryPriorityByName,
+              dictionaryFrequencyModeByName
+            )
+          }));
+          namePos += nameMatch.sourceLength;
+        }
+      }
+      // First reserved name span that a match ending at endPos would leave
+      // half-consumed. Spans the match covers entirely are not returned: those
+      // lose to the longer word instead of splitting it.
+      function findSplitNameToken(startIndex, endPos) {
+        for (let index = startIndex; index < nameTokens.length; index += 1) {
+          const nameToken = nameTokens[index];
+          if (nameToken.startPos >= endPos) { return null; }
+          if (nameToken.endPos > endPos) { return nameToken; }
+        }
+        return null;
+      }
+      let i = 0;
+      let nameIndex = 0;
+      let unparsedRunStart = null;
+      while (i < text.length) {
+        while (nameIndex < nameTokens.length && nameTokens[nameIndex].startPos < i) { nameIndex += 1; }
+        const nextNameToken = nameIndex < nameTokens.length ? nameTokens[nameIndex] : null;
+        if (nextNameToken && nextNameToken.startPos === i) {
+          flushUnparsedRun(unparsedRunStart, i);
+          unparsedRunStart = null;
+          tokens.push(nextNameToken);
+          i = nextNameToken.endPos;
+          nameIndex += 1;
+          continue;
+        }
+        const codePoint = text.codePointAt(i);
+        // Punctuation and whitespace can never start a token: skip the backend
+        // round trip entirely. Latin letters and digits stay lookup-worthy
+        // (terms like Tシャツ start on an ASCII letter).
+        if (!isLookupWorthyCodePoint(codePoint)) {
+          if (unparsedRunStart === null) { unparsedRunStart = i; }
+          i += String.fromCodePoint(codePoint).length;
+          continue;
+        }
+        // A reservation only outranks generic matches that would cut into it.
+        // Look the position up unrestricted first: a generic word that starts
+        // earlier and covers the whole name span (写真 over a character named
+        // 真) is the better reading, so the reservation yields rather than
+        // splitting the word. Only a match that ends inside a name span gets
+        // re-run against a window capped at that span.
+        let attempt = await resolveTokenAt(i, scanLength);
+        if (attempt.token) {
+          const splitNameToken = findSplitNameToken(nameIndex, attempt.token.endPos);
+          if (splitNameToken) {
+            attempt = await resolveTokenAt(i, splitNameToken.startPos - i);
+          }
+        }
+        if (attempt.token) {
+          flushUnparsedRun(unparsedRunStart, i);
+          unparsedRunStart = null;
+          tokens.push(attempt.token);
+          i += attempt.matchedLength;
+          continue;
+        }
+        if (unparsedRunStart === null) { unparsedRunStart = i; }
+        i += String.fromCodePoint(text.codePointAt(i)).length;
+      }
+      flushUnparsedRun(unparsedRunStart, text.length);
+      if (blindRetryBudgetExhausted) {
+        // A position gave up with shorter windows still worth trying. The walk
+        // is the only tokenizer now, so stopping there would leave a real term
+        // as an unparsed run; report it so the host can spend one parseText on
+        // the line instead of letting the ladder run to O(scanLength) lookups.
+        return { tokens, retryBudgetExhausted: true };
+      }
+      return tokens;
+    };
+    return true;
+  })();
+`;
+
+// Installs (or clears) the character-name candidate forms for the current
+// media. Runs only when the list changes, not per line. Passing null restores
+// the exhaustive every-position pre-pass.
+export function buildYomitanScanNameCandidatesScript(
+  nameCandidates: { key: string; forms: string[] } | null,
+): string {
+  if (!nameCandidates) {
+    return `
+    (() => {
+      if (typeof globalThis.__subminerYomitanScanSetNameCandidates !== "function") {
+        return false;
+      }
+      return globalThis.__subminerYomitanScanSetNameCandidates(null, null);
+    })();
+  `;
+  }
+
+  return `
+    (() => {
+      if (typeof globalThis.__subminerYomitanScanSetNameCandidates !== "function") {
+        return false;
+      }
+      return globalThis.__subminerYomitanScanSetNameCandidates(
+        ${JSON.stringify(nameCandidates.key)},
+        ${JSON.stringify(nameCandidates.forms)}
+      );
+    })();
+  `;
+}
+
+export function buildYomitanScanCallScript(params: YomitanScanRequestParams): string {
+  return `
+    (async () => {
+      if (typeof globalThis.__subminerYomitanScan !== "function") {
+        return ${JSON.stringify(YOMITAN_SCAN_RUNTIME_MISSING_SENTINEL)};
+      }
+      return await globalThis.__subminerYomitanScan(${JSON.stringify(params)});
+    })();
+  `;
+}

src/core/services/tokenizer/yomitan-scanning-helpers-script.ts

@@ -0,0 +1,496 @@
+// Helper bundle for the in-page Yomitan scan runtime: kana/furigana handling,
+// headword preference, and frequency-rank resolution. Injected as text into the
+// parser window by yomitan-scan-runtime-script.ts, so it is data here, not code
+// this process runs.
+import { HAN_CODE_POINT_RANGES } from '../../text/han-code-points';
+
+export const CHARACTER_DICTIONARY_TITLE_PREFIX = 'SubMiner Character Dictionary';
+
+export const YOMITAN_SCANNING_HELPERS = String.raw`
+      const HIRAGANA_CONVERSION_RANGE = [0x3041, 0x3096];
+      const KATAKANA_CONVERSION_RANGE = [0x30a1, 0x30f6];
+      const KANA_PROLONGED_SOUND_MARK_CODE_POINT = 0x30fc;
+      const KATAKANA_SMALL_KA_CODE_POINT = 0x30f5;
+      const KATAKANA_SMALL_KE_CODE_POINT = 0x30f6;
+      const KANA_RANGES = [[0x3040, 0x309f], [0x30a0, 0x30ff]];
+      // Han ranges come from the shared table so the scan walk and the character
+      // dictionary agree on what a kanji is (supplementary planes included).
+      // Halfwidth katakana counts as Japanese text: a name written that way has
+      // to reach the greedy pre-pass, which has its own handling for it.
+      const JAPANESE_RANGES = [[0x3040, 0x30ff], [0xff66, 0xff9f], ...${JSON.stringify(HAN_CODE_POINT_RANGES)}];
+      function isCodePointInRange(codePoint, range) { return codePoint >= range[0] && codePoint <= range[1]; }
+      function isCodePointInRanges(codePoint, ranges) { return ranges.some((range) => isCodePointInRange(codePoint, range)); }
+      function isCodePointKana(codePoint) { return isCodePointInRanges(codePoint, KANA_RANGES); }
+      function isCodePointJapanese(codePoint) { return isCodePointInRanges(codePoint, JAPANESE_RANGES); }
+      function createFuriganaSegment(text, reading) { return {text, reading}; }
+      function getSegmentReadingContribution(segment) {
+        if (typeof segment.reading === "string" && segment.reading.length > 0) { return segment.reading; }
+        const segmentText = typeof segment.text === "string" ? segment.text : "";
+        const isKanaOnly = segmentText.length > 0 && [...segmentText].every((char) => isCodePointKana(char.codePointAt(0)));
+        return isKanaOnly ? segmentText : "";
+      }
+      function getProlongedHiragana(previousCharacter) {
+        switch (previousCharacter) {
+          case "あ": case "か": case "が": case "さ": case "ざ": case "た": case "だ": case "な": case "は": case "ば": case "ぱ": case "ま": case "や": case "ら": case "わ": case "ぁ": case "ゃ": case "ゎ": return "あ";
+          case "い": case "き": case "ぎ": case "し": case "じ": case "ち": case "ぢ": case "に": case "ひ": case "び": case "ぴ": case "み": case "り": case "ぃ": return "い";
+          case "う": case "く": case "ぐ": case "す": case "ず": case "つ": case "づ": case "ぬ": case "ふ": case "ぶ": case "ぷ": case "む": case "ゆ": case "る": case "ぅ": case "ゅ": return "う";
+          case "え": case "け": case "げ": case "せ": case "ぜ": case "て": case "で": case "ね": case "へ": case "べ": case "ぺ": case "め": case "れ": case "ぇ": return "え";
+          case "お": case "こ": case "ご": case "そ": case "ぞ": case "と": case "ど": case "の": case "ほ": case "ぼ": case "ぽ": case "も": case "よ": case "ろ": case "を": case "ぉ": case "ょ": return "う";
+          default: return null;
+        }
+      }
+      function getFuriganaKanaSegments(text, reading) {
+        const newSegments = [];
+        let start = 0;
+        let state = (reading[0] === text[0]);
+        for (let i = 1; i < text.length; ++i) {
+          const newState = (reading[i] === text[i]);
+          if (state === newState) { continue; }
+          newSegments.push(createFuriganaSegment(text.substring(start, i), state ? '' : reading.substring(start, i)));
+          state = newState;
+          start = i;
+        }
+        newSegments.push(createFuriganaSegment(text.substring(start), state ? '' : reading.substring(start)));
+        return newSegments;
+      }
+      function convertKatakanaToHiragana(text, keepProlongedSoundMarks = false) {
+        let result = '';
+        const offset = (HIRAGANA_CONVERSION_RANGE[0] - KATAKANA_CONVERSION_RANGE[0]);
+        for (let char of text) {
+          const codePoint = char.codePointAt(0);
+          switch (codePoint) {
+            case KATAKANA_SMALL_KA_CODE_POINT:
+            case KATAKANA_SMALL_KE_CODE_POINT:
+              break;
+            case KANA_PROLONGED_SOUND_MARK_CODE_POINT:
+              if (!keepProlongedSoundMarks && result.length > 0) {
+                const char2 = getProlongedHiragana(result[result.length - 1]);
+                if (char2 !== null) { char = char2; }
+              }
+              break;
+            default:
+              if (isCodePointInRange(codePoint, KATAKANA_CONVERSION_RANGE)) {
+                char = String.fromCodePoint(codePoint + offset);
+              }
+              break;
+          }
+          result += char;
+        }
+        return result;
+      }
+      function segmentizeFurigana(reading, readingNormalized, groups, groupsStart) {
+        const groupCount = groups.length - groupsStart;
+        if (groupCount <= 0) { return reading.length === 0 ? [] : null; }
+        const group = groups[groupsStart];
+        const {isKana, text} = group;
+        if (isKana) {
+          if (group.textNormalized !== null && readingNormalized.startsWith(group.textNormalized)) {
+            const segments = segmentizeFurigana(reading.substring(text.length), readingNormalized.substring(text.length), groups, groupsStart + 1);
+            if (segments !== null) {
+              if (reading.startsWith(text)) { segments.unshift(createFuriganaSegment(text, '')); }
+              else { segments.unshift(...getFuriganaKanaSegments(text, reading)); }
+              return segments;
+            }
+          }
+          return null;
+        }
+        let result = null;
+        for (let i = reading.length; i >= text.length; --i) {
+          const segments = segmentizeFurigana(reading.substring(i), readingNormalized.substring(i), groups, groupsStart + 1);
+          if (segments !== null) {
+            if (result !== null) { return null; }
+            segments.unshift(createFuriganaSegment(text, reading.substring(0, i)));
+            result = segments;
+          }
+          if (groupCount === 1) { break; }
+        }
+        return result;
+      }
+      function distributeFurigana(term, reading) {
+        if (reading === term) { return [createFuriganaSegment(term, '')]; }
+        const groups = [];
+        let groupPre = null;
+        let isKanaPre = null;
+        for (const c of term) {
+          const isKana = isCodePointKana(c.codePointAt(0));
+          if (isKana === isKanaPre) { groupPre.text += c; }
+          else {
+            groupPre = {isKana, text: c, textNormalized: null};
+            groups.push(groupPre);
+            isKanaPre = isKana;
+          }
+        }
+        for (const group of groups) {
+          if (group.isKana) { group.textNormalized = convertKatakanaToHiragana(group.text); }
+        }
+        const segments = segmentizeFurigana(reading, convertKatakanaToHiragana(reading), groups, 0);
+        return segments !== null ? segments : [createFuriganaSegment(term, reading)];
+      }
+      function getStemLength(text1, text2) {
+        const minLength = Math.min(text1.length, text2.length);
+        if (minLength === 0) { return 0; }
+        let i = 0;
+        while (true) {
+          const char1 = text1.codePointAt(i);
+          const char2 = text2.codePointAt(i);
+          if (char1 !== char2) { break; }
+          const charLength = String.fromCodePoint(char1).length;
+          i += charLength;
+          if (i >= minLength) {
+            if (i > minLength) { i -= charLength; }
+            break;
+          }
+        }
+        return i;
+      }
+      function distributeFuriganaInflected(term, reading, source) {
+        const termNormalized = convertKatakanaToHiragana(term);
+        const readingNormalized = convertKatakanaToHiragana(reading);
+        const sourceNormalized = convertKatakanaToHiragana(source);
+        let mainText = term;
+        let stemLength = getStemLength(termNormalized, sourceNormalized);
+        const readingStemLength = getStemLength(readingNormalized, sourceNormalized);
+        if (readingStemLength > 0 && readingStemLength >= stemLength) {
+          mainText = reading;
+          stemLength = readingStemLength;
+          reading = source.substring(0, stemLength) + reading.substring(stemLength);
+        }
+        const segments = [];
+        if (stemLength > 0) {
+          mainText = source.substring(0, stemLength) + mainText.substring(stemLength);
+          const segments2 = distributeFurigana(mainText, reading);
+          let consumed = 0;
+          for (const segment of segments2) {
+            const start = consumed;
+            consumed += segment.text.length;
+            if (consumed < stemLength) { segments.push(segment); }
+            else if (consumed === stemLength) { segments.push(segment); break; }
+            else {
+              if (start < stemLength) { segments.push(createFuriganaSegment(mainText.substring(start, stemLength), '')); }
+              break;
+            }
+          }
+        }
+        if (stemLength < source.length) {
+          const remainder = source.substring(stemLength);
+          const last = segments[segments.length - 1];
+          if (last && last.reading.length === 0) { last.text += remainder; }
+          else { segments.push(createFuriganaSegment(remainder, '')); }
+        }
+        return segments;
+      }
+      function parsePositiveFrequencyNumber(value) {
+        if (typeof value === 'number' && Number.isFinite(value) && value > 0) {
+          return Math.max(1, Math.floor(value));
+        }
+        if (typeof value === 'string') {
+          const numericMatch = value.trim().match(/[+-]?(\d+(\.\d*)?|\.\d+)([eE][+-]?\d+)?/)?.[0];
+          if (!numericMatch) { return null; }
+          const parsed = Number.parseFloat(numericMatch);
+          if (!Number.isFinite(parsed) || parsed <= 0) { return null; }
+          return Math.max(1, Math.floor(parsed));
+        }
+        if (Array.isArray(value)) {
+          for (const item of value) {
+            const parsed = parsePositiveFrequencyNumber(item);
+            if (parsed !== null) { return parsed; }
+          }
+        }
+        return null;
+      }
+      function parseDisplayFrequencyNumber(value) {
+        if (typeof value === 'string') {
+          const leadingDigits = value.trim().match(/^\d+/)?.[0];
+          if (!leadingDigits) { return null; }
+          const parsed = Number.parseInt(leadingDigits, 10);
+          return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
+        }
+        return parsePositiveFrequencyNumber(value);
+      }
+      function getFrequencyDictionaryName(frequency) {
+        const candidates = [
+          frequency?.dictionary,
+          frequency?.dictionaryName,
+          frequency?.name,
+          frequency?.title,
+          frequency?.dictionaryTitle,
+          frequency?.dictionaryAlias
+        ];
+        for (const candidate of candidates) {
+          if (typeof candidate === 'string' && candidate.trim().length > 0) {
+            return candidate.trim();
+          }
+        }
+        return null;
+      }
+      function getBestFrequencyRank(dictionaryEntry, headwordIndex, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
+        let best = null;
+        const headwordCount = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords.length : 0;
+        for (const frequency of dictionaryEntry?.frequencies || []) {
+          if (!frequency || typeof frequency !== 'object') { continue; }
+          const frequencyHeadwordIndex = frequency.headwordIndex;
+          if (typeof frequencyHeadwordIndex === 'number') {
+            if (frequencyHeadwordIndex !== headwordIndex) { continue; }
+          } else if (headwordCount > 1) {
+            continue;
+          }
+          const dictionary = getFrequencyDictionaryName(frequency);
+          if (!dictionary) { continue; }
+          if (dictionaryFrequencyModeByName[dictionary] === 'occurrence-based') { continue; }
+          const rank =
+            parseDisplayFrequencyNumber(frequency.displayValue) ??
+            parsePositiveFrequencyNumber(frequency.frequency);
+          if (rank === null) { continue; }
+          const priorityRaw = dictionaryPriorityByName[dictionary];
+          const fallbackPriority =
+            typeof frequency.dictionaryIndex === 'number' && Number.isFinite(frequency.dictionaryIndex)
+              ? Math.max(0, Math.floor(frequency.dictionaryIndex))
+              : Number.MAX_SAFE_INTEGER;
+          const priority =
+            typeof priorityRaw === 'number' && Number.isFinite(priorityRaw)
+              ? Math.max(0, Math.floor(priorityRaw))
+              : fallbackPriority;
+          if (best === null || priority < best.priority || (priority === best.priority && rank < best.rank)) {
+            best = { priority, rank };
+          }
+        }
+        return best?.rank ?? null;
+      }
+      function hasExactSource(headword, token, requirePrimary) {
+        for (const src of headword.sources || []) {
+          if (src.originalText !== token) { continue; }
+          if (requirePrimary && !src.isPrimary) { continue; }
+          if (src.matchType !== 'exact') { continue; }
+          return true;
+        }
+        return false;
+      }
+      function collectExactHeadwordMatches(dictionaryEntries, token, requirePrimary) {
+        const matches = [];
+        for (const dictionaryEntry of dictionaryEntries || []) {
+          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
+          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
+            const headword = headwords[headwordIndex];
+            if (!hasExactSource(headword, token, requirePrimary)) { continue; }
+            matches.push({ dictionaryEntry, headword, headwordIndex });
+          }
+        }
+        return matches;
+      }
+      function sameHeadword(match, preferredMatch) {
+        if (!match || !preferredMatch) {
+          return false;
+        }
+        if (match.headword?.term !== preferredMatch.headword?.term) {
+          return false;
+        }
+        const matchReading = typeof match.headword?.reading === 'string' ? match.headword.reading : '';
+        const preferredReading =
+          typeof preferredMatch.headword?.reading === 'string' ? preferredMatch.headword.reading : '';
+        if (!matchReading || !preferredReading) {
+          return true;
+        }
+        return matchReading === preferredReading;
+      }
+      function getBestFrequencyRankForMatches(matches, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
+        let best = null;
+        for (const match of matches) {
+          const rank = getBestFrequencyRank(
+            match.dictionaryEntry,
+            match.headwordIndex,
+            dictionaryPriorityByName,
+            dictionaryFrequencyModeByName
+          );
+          if (rank === null) { continue; }
+          if (best === null || rank < best) {
+            best = rank;
+          }
+        }
+        return best;
+      }
+      function normalizeWordClasses(headword) {
+          if (!Array.isArray(headword?.wordClasses)) { return undefined; }
+          const classes = headword.wordClasses.filter((wordClass) => typeof wordClass === "string" && wordClass.trim().length > 0);
+          return classes.length > 0 ? classes : undefined;
+        }
+        function appendDictionaryNames(target, value) {
+          if (!value || typeof value !== 'object') {
+            return;
+          }
+          const candidates = [
+            value.dictionary,
+            value.dictionaryName,
+            value.name,
+            value.title,
+            value.dictionaryTitle,
+            value.dictionaryAlias
+          ];
+          for (const candidate of candidates) {
+            if (typeof candidate === 'string' && candidate.trim().length > 0) {
+              target.push(candidate.trim());
+            }
+          }
+        }
+        function getDictionaryEntryNames(entry) {
+          const names = [];
+          appendDictionaryNames(names, entry);
+          for (const definition of entry?.definitions || []) {
+            appendDictionaryNames(names, definition);
+          }
+          for (const frequency of entry?.frequencies || []) {
+            appendDictionaryNames(names, frequency);
+          }
+          for (const pronunciation of entry?.pronunciations || []) {
+            appendDictionaryNames(names, pronunciation);
+          }
+          return names;
+        }
+        function isNameDictionaryEntry(entry) {
+          if (!includeNameMatchMetadata || !entry || typeof entry !== 'object') {
+            return false;
+          }
+          return getDictionaryEntryNames(entry).some((name) => name.startsWith(${JSON.stringify(CHARACTER_DICTIONARY_TITLE_PREFIX)}));
+        }
+        function parseSubMinerMediaIdFromString(value) {
+          const imageMatch = value.match(/\bimg\/m(\d+)-/i);
+          if (imageMatch) {
+            const parsed = Number.parseInt(imageMatch[1], 10);
+            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
+          }
+          const titleMatch = value.match(/${CHARACTER_DICTIONARY_TITLE_PREFIX}[^\d]*(?:AniList\s*)?(\d+)/i);
+          if (titleMatch) {
+            const parsed = Number.parseInt(titleMatch[1], 10);
+            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
+          }
+          return null;
+        }
+        function parseSubMinerMediaIdCandidate(value) {
+          if (typeof value === 'number' && Number.isSafeInteger(value) && value > 0) {
+            return value;
+          }
+          if (typeof value === 'string' && /^\d+$/.test(value.trim())) {
+            const parsed = Number.parseInt(value.trim(), 10);
+            if (Number.isSafeInteger(parsed) && parsed > 0) { return parsed; }
+          }
+          return null;
+        }
+        function collectSubMinerMediaIds(value, target) {
+          if (typeof value === 'string') {
+            const parsed = parseSubMinerMediaIdFromString(value);
+            if (parsed !== null) { target.add(parsed); }
+            return;
+          }
+          if (!value || typeof value !== 'object') {
+            return;
+          }
+          if (Array.isArray(value)) {
+            for (const item of value) { collectSubMinerMediaIds(item, target); }
+            return;
+          }
+          const mediaIdCandidates = [
+            value.subminerMediaId,
+            value.subMinerMediaId,
+            value.characterDictionaryMediaId,
+            value.data?.subminerMediaId,
+            value.data?.subMinerMediaId,
+            value.data?.characterDictionaryMediaId
+          ];
+          for (const candidate of mediaIdCandidates) {
+            const parsed = parseSubMinerMediaIdCandidate(candidate);
+            if (parsed !== null) { target.add(parsed); }
+          }
+          for (const child of Object.values(value)) {
+            collectSubMinerMediaIds(child, target);
+          }
+        }
+        function getSubMinerMediaIds(entry) {
+          const mediaIds = new Set();
+          collectSubMinerMediaIds(entry, mediaIds);
+          return mediaIds;
+        }
+        function isCurrentMediaNameDictionaryEntry(entry) {
+          if (!isNameDictionaryEntry(entry)) {
+            return false;
+          }
+          if (currentCharacterDictionaryMediaId === null) {
+            return true;
+          }
+          const mediaIds = getSubMinerMediaIds(entry);
+          return mediaIds.size === 0 || mediaIds.has(currentCharacterDictionaryMediaId);
+      }
+      function findLongestNameMatch(dictionaryEntries, textWindow) {
+        let best = null;
+        for (const dictionaryEntry of dictionaryEntries || []) {
+          if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
+          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
+          for (let headwordIndex = 0; headwordIndex < headwords.length; headwordIndex += 1) {
+            const headword = headwords[headwordIndex];
+            for (const src of headword?.sources || []) {
+              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
+              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
+              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
+              if (best === null || originalText.length > best.sourceLength) {
+                best = { dictionaryEntry, headword, headwordIndex, sourceLength: originalText.length };
+              }
+            }
+          }
+        }
+        return best;
+      }
+      function findLongestGenericMatchLength(dictionaryEntries, textWindow) {
+        let best = 0;
+        for (const dictionaryEntry of dictionaryEntries || []) {
+          if (isNameDictionaryEntry(dictionaryEntry)) { continue; }
+          const headwords = Array.isArray(dictionaryEntry?.headwords) ? dictionaryEntry.headwords : [];
+          for (const headword of headwords) {
+            for (const src of headword?.sources || []) {
+              if (src.matchType !== 'exact' || src.isPrimary !== true) { continue; }
+              const originalText = typeof src.originalText === 'string' ? src.originalText : '';
+              if (!originalText || !textWindow.startsWith(originalText)) { continue; }
+              if (originalText.length > best) { best = originalText.length; }
+            }
+          }
+        }
+        return best;
+      }
+      function getPreferredHeadword(dictionaryEntries, token, dictionaryPriorityByName, dictionaryFrequencyModeByName) {
+        const currentMediaDictionaryEntries =
+          currentCharacterDictionaryMediaId === null
+            ? (dictionaryEntries || [])
+            : (dictionaryEntries || []).filter((entry) => {
+                if (!isNameDictionaryEntry(entry)) { return true; }
+                return isCurrentMediaNameDictionaryEntry(entry);
+              });
+        const exactPrimaryMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, true);
+        let matchedNameDictionary = false;
+        if (includeNameMatchMetadata) {
+          for (const dictionaryEntry of currentMediaDictionaryEntries || []) {
+            if (!isCurrentMediaNameDictionaryEntry(dictionaryEntry)) { continue; }
+            for (const match of exactPrimaryMatches) {
+              if (match.dictionaryEntry !== dictionaryEntry) { continue; }
+              matchedNameDictionary = true;
+              break;
+            }
+            if (matchedNameDictionary) { break; }
+          }
+        }
+        const preferredMatch = exactPrimaryMatches[0];
+        if (preferredMatch) {
+          const exactFrequencyMatches = collectExactHeadwordMatches(currentMediaDictionaryEntries, token, false)
+            .filter((match) => sameHeadword(match, preferredMatch));
+          return {
+            term: preferredMatch.headword.term,
+            reading: preferredMatch.headword.reading,
+            wordClasses: normalizeWordClasses(preferredMatch.headword),
+            isNameMatch:
+              matchedNameDictionary || isCurrentMediaNameDictionaryEntry(preferredMatch.dictionaryEntry),
+            frequencyRank: getBestFrequencyRankForMatches(
+              exactFrequencyMatches.length > 0 ? exactFrequencyMatches : exactPrimaryMatches,
+              dictionaryPriorityByName,
+              dictionaryFrequencyModeByName
+            )
+          };
+        }
+        return null;
+      }
+`;

src/core/text/han-code-points.test.ts

@@ -0,0 +1,54 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { HAN_CODE_POINT_RANGES, HAN_REGEXP_CLASS_BODY, isHanCodePoint } from './han-code-points';
+
+test('every range boundary is inside the table', () => {
+  for (const [start, end] of HAN_CODE_POINT_RANGES) {
+    for (const codePoint of [start, end]) {
+      assert.ok(isHanCodePoint(codePoint), `expected U+${codePoint.toString(16)} to be Han`);
+    }
+  }
+
+  // Extension J (Unicode 17) and the Compatibility blocks are the ones a
+  // BMP-only table used to miss.
+  assert.ok(isHanCodePoint(0x323b0));
+  assert.ok(isHanCodePoint(0x33479));
+  assert.ok(isHanCodePoint(0xf900));
+  assert.ok(isHanCodePoint(0x2f800));
+});
+
+test('no unified ideograph the runtime knows about falls outside the table', () => {
+  // One direction only: a runtime with older Unicode data simply checks fewer
+  // code points, where asserting the reverse would fail on Extension J.
+  const unifiedIdeograph = /\p{Unified_Ideograph}/u;
+
+  for (let codePoint = 0x3000; codePoint <= 0x40000; codePoint += 1) {
+    if (unifiedIdeograph.test(String.fromCodePoint(codePoint))) {
+      assert.ok(
+        isHanCodePoint(codePoint),
+        `expected unified ideograph U+${codePoint.toString(16)} to be in the table`,
+      );
+    }
+  }
+});
+
+test('code points just outside the table are rejected', () => {
+  for (const codePoint of [0x33ff, 0x4dc0, 0xa000, 0x1f000, 0x3347a]) {
+    assert.equal(
+      isHanCodePoint(codePoint),
+      false,
+      `expected U+${codePoint.toString(16)} not to be Han`,
+    );
+  }
+});
+
+test('the regexp class body matches the same code points as the predicate', () => {
+  const classRegExp = new RegExp(`^[${HAN_REGEXP_CLASS_BODY}]$`, 'u');
+
+  for (const codePoint of [0x3400, 0x4e00, 0x9fff, 0xf900, 0x20000, 0x323b0, 0x33479]) {
+    assert.match(String.fromCodePoint(codePoint), classRegExp);
+  }
+  for (const codePoint of [0x3040, 0x30ff, 0x33fa, 0x3347a]) {
+    assert.doesNotMatch(String.fromCodePoint(codePoint), classRegExp);
+  }
+});

src/core/text/han-code-points.ts

@@ -0,0 +1,31 @@
+// Single source of truth for "this code point is a Han character", shared by
+// the main-process character dictionary and the in-page Yomitan scan runtime.
+// The two used to carry separate range lists, and they drifted: a name written
+// with a supplementary-plane kanji could enter the generated dictionary while
+// the scanner's greedy name pre-pass refused to probe the position.
+//
+// Ranges rather than \p{Script=Han}: the scan walk tests one code point per
+// character of every subtitle line, where an integer compare beats building a
+// string for a regex, and the script is injected as text into a page where a
+// shared helper cannot be imported.
+export const HAN_CODE_POINT_RANGES: ReadonlyArray<readonly [number, number]> = [
+  [0x3400, 0x4dbf], // Extension A
+  [0x4e00, 0x9fff], // CJK Unified Ideographs
+  [0xf900, 0xfaff], // Compatibility Ideographs
+  [0x20000, 0x2a6df], // Extension B
+  [0x2a700, 0x2ebef], // Extensions C-F
+  [0x2ebf0, 0x2ee5f], // Extension I
+  [0x2f800, 0x2fa1f], // Compatibility Ideographs Supplement
+  [0x30000, 0x3134f], // Extension G
+  [0x31350, 0x323af], // Extension H
+  [0x323b0, 0x33479], // Extension J (Unicode 17)
+];
+
+export function isHanCodePoint(codePoint: number): boolean {
+  return HAN_CODE_POINT_RANGES.some(([start, end]) => codePoint >= start && codePoint <= end);
+}
+
+/** The same ranges as a regular expression character class body (needs the `u` flag). */
+export const HAN_REGEXP_CLASS_BODY = HAN_CODE_POINT_RANGES.map(
+  ([start, end]) => `\\u{${start.toString(16)}}-\\u{${end.toString(16)}}`,
+).join('');

src/main.ts

@@ -295,6 +295,7 @@ import {
   importYomitanDictionaryFromZip,
   initializeOverlayAnkiIntegration as initializeOverlayAnkiIntegrationCore,
   initializeOverlayRuntime as initializeOverlayRuntimeCore,
+  isSubtitleAnnotationUpgrade,
   isOverlayWindowContentReady,
   jellyfinTicksToSecondsRuntime,
   listJellyfinItemsRuntime,
@@ -486,6 +487,7 @@ import { createOverlayVisibilityRuntimeService } from './main/overlay-visibility
 import { createDiscordPresenceRuntime } from './main/runtime/discord-presence-runtime';
 import { createCharacterDictionaryRuntimeService } from './main/character-dictionary-runtime';
 import { createCharacterDictionaryImageLookup } from './main/character-dictionary-runtime/image-lookup';
+import { createCharacterNameCandidateLookup } from './main/character-dictionary-runtime/name-candidates';
 import {
   createCharacterDictionaryAutoSyncRuntimeService,
   getCharacterDictionaryManagerSnapshot,
@@ -1815,8 +1817,10 @@ function withCurrentSubtitleTiming(payload: SubtitleData): SubtitleData {
     endTime: appState.mpvClient?.currentSubEnd ?? null,
   };
 }
-function emitSubtitlePayload(payload: SubtitleData): void {
+function emitSubtitlePayload(payload: SubtitleData, options?: { resumePrefetch?: boolean }): void {
   const timedPayload = withCurrentSubtitleTiming(payload);
+  const currentSubtitleData = appState.currentSubtitleData;
+  const isAnnotationUpgrade = isSubtitleAnnotationUpgrade(currentSubtitleData, timedPayload);
   const frequencyDictionary = configService.getConfig().subtitleStyle.frequencyDictionary;
   const frequencyOptions = {
     enabled: frequencyDictionary.enabled,
@@ -1825,10 +1829,18 @@ function emitSubtitlePayload(payload: SubtitleData): void {
   };
   appState.currentSubtitleData = timedPayload;
   overlayManager.broadcastToOverlayWindows('subtitle:set', timedPayload);
+  if (!isAnnotationUpgrade) {
     subtitleWsService.broadcast(timedPayload, frequencyOptions);
+  }
   annotationSubtitleWsService.broadcast(timedPayload, frequencyOptions);
   autoplayReadyGate.maybeSignalPluginAutoplayReady(timedPayload, { forceWhilePaused: true });
+  // resumePrefetch: false marks an emit that is not the end of the work for
+  // this line; prefetch stays paused until the subtitle processing controller
+  // settles so it does not compete with the on-screen line for the single
+  // Yomitan parser window.
+  if (options?.resumePrefetch !== false) {
     subtitlePrefetchService?.resume();
+  }
 }
 function getCurrentAutoplaySubtitlePayload(): SubtitleData | null {
   const payload = appState.currentSubtitleData;
@@ -1884,7 +1896,17 @@ const buildSubtitleProcessingControllerMainDepsHandler =
   createBuildSubtitleProcessingControllerMainDepsHandler({
     tokenizeSubtitle: async (text: string) =>
       tokenizeSubtitleDeferred ? await tokenizeSubtitleDeferred(text) : { text, tokens: null },
-    emitSubtitle: (payload) => emitSubtitlePayload(payload),
+    // Controller emits never release the prefetch pause: the first emit for an
+    // uncached line is the provisional plain payload, sent before tokenization
+    // starts, so resuming on it would put prefetch back in contention with the
+    // on-screen line for the single parser window.
+    emitSubtitle: (payload) => emitSubtitlePayload(payload, { resumePrefetch: false }),
+    // The pause is released once the controller has no work left, which covers
+    // the runs that end without an emit (suppressed duplicate, failed
+    // tokenization) as well as the ones that deliver a payload.
+    onProcessingSettled: () => {
+      subtitlePrefetchService?.resume();
+    },
     logDebug: (message) => {
       logger.debug(`[subtitle-processing] ${message}`);
     },
@@ -1921,7 +1943,7 @@ const autoplaySubtitlePrimingRuntime = createAutoplaySubtitlePrimingRuntime({
     appState.activeParsedSubtitleMediaPath = mediaPath;
   },
   subtitleProcessingController,
-  emitSubtitlePayload: (payload) => emitSubtitlePayload(payload),
+  emitSubtitlePayload: (payload, options) => emitSubtitlePayload(payload, options),
   getSubtitlePrefetchService: () => subtitlePrefetchService,
   getLastObservedTimePos: () => lastObservedTimePos,
   getVisibleOverlayVisible: () => overlayManager.getVisibleOverlayVisible(),
@@ -1955,7 +1977,6 @@ 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) => {
@@ -1982,15 +2003,22 @@ const resolveActiveSubtitleSidebarSourceHandler = createResolveActiveSubtitleSid
   getFfmpegPath: () => configService.getConfig().subsync.ffmpeg_path.trim() || 'ffmpeg',
   extractInternalSubtitleTrack: (ffmpegPath, videoPath, track) =>
     extractInternalSubtitleTrackToTempFile(ffmpegPath, videoPath, track),
+  logDebug: (message) => logger.debug(message),
 });
 
 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),
+    logWarn: (message) => logger.warn(message),
   });
 
 const subtitlePrefetchRuntime = {
@@ -2521,6 +2549,10 @@ const characterDictionaryAutoSyncRuntime = createCharacterDictionaryAutoSyncRunt
       },
       {
         hasParserWindow: () => Boolean(appState.yomitanParserWindow),
+        invalidateCharacterDictionaryLookups: () => {
+          characterDictionaryImageLookup.invalidate();
+          characterNameCandidateLookup.invalidate();
+        },
         clearParserCaches: () => {
           if (appState.yomitanParserWindow) {
             clearYomitanParserCachesForWindow(appState.yomitanParserWindow);
@@ -2546,6 +2578,13 @@ const characterDictionaryImageLookup = createCharacterDictionaryImageLookup({
   getCurrentMediaId: () => characterDictionaryAutoSyncRuntime.getCurrentMediaId(),
 });
 
+// Lets the Yomitan scan runtime skip name lookups at positions where no
+// character name can start; absent candidates just mean the exhaustive scan.
+const characterNameCandidateLookup = createCharacterNameCandidateLookup({
+  userDataPath: USER_DATA_PATH,
+  getCurrentMediaId: () => characterDictionaryAutoSyncRuntime.getCurrentMediaId(),
+});
+
 const overlayVisibilityRuntime = createOverlayVisibilityRuntimeService(
   createBuildOverlayVisibilityRuntimeMainDepsHandler({
     getMainWindow: () => overlayManager.getMainWindow(),
@@ -2994,6 +3033,8 @@ const {
         streamIndex,
         delaySeconds,
       }),
+    initSubtitlePrefetch: (sourcePath) =>
+      subtitlePrefetchRuntime.refreshSubtitleSidebarFromSource(sourcePath),
     logDebug: (message, error) => {
       logger.debug(message, error);
     },
@@ -3958,7 +3999,10 @@ const refreshCurrentSubtitleAfterKnownWordUpdate = (): void => {
   }
   subtitleProcessingController.invalidateTokenizationCache();
   subtitlePrefetchService?.onSeek(lastObservedTimePos);
-  subtitleProcessingController.refreshCurrentSubtitle(appState.currentSubText);
+  if (!subtitleProcessingController.refreshCurrentSubtitle(appState.currentSubText)) {
+    // Idle controller: no settle is coming to release the pause above.
+    subtitlePrefetchService?.resume();
+  }
 };
 let hasAttemptedImmersionTrackerStartup = false;
 const ensureImmersionTrackerStarted = (): void => {
@@ -4359,13 +4403,20 @@ const {
       emitSubtitlePayload(payload);
     },
     onSubtitleChange: (text) => {
+      // Pause only; restarting the prefetch run here would discard in-flight
+      // tokenization work on every line. Real seeks restart via onTimePosUpdate.
       subtitlePrefetchService?.pause();
-      subtitlePrefetchService?.onSeek(lastObservedTimePos);
+      if (!subtitleProcessingController.onSubtitleChange(text)) {
-      subtitleProcessingController.onSubtitleChange(text);
+        // Repeat of the current text: the controller is idle, so no settle is
+        // coming to release the pause. Resume now instead of idling prefetch
+        // for the rest of the cue.
+        subtitlePrefetchService?.resume();
+      }
     },
     refreshDiscordPresence: () => {
       discordPresenceRuntime.publishDiscordPresence();
     },
+    logSubtitleProcessingDebug: (message: string) => logger.debug(message),
     ensureImmersionTrackerInitialized: () => {
       ensureImmersionTrackerStarted();
     },
@@ -4604,6 +4655,7 @@ const {
       getCharacterNameImage: (term) => characterDictionaryImageLookup.get(term),
       getCurrentCharacterDictionaryMediaId: () =>
         characterDictionaryAutoSyncRuntime.getCurrentMediaId(),
+      getCharacterNameCandidates: () => characterNameCandidateLookup.get(),
       getFrequencyDictionaryEnabled: () =>
         getRuntimeBooleanOption(
           'subtitle.annotation.frequency',
@@ -5658,7 +5710,6 @@ const { registerIpcRuntimeHandlers } = composeIpcRuntimeHandlers({
           if (result.ok && result.rebuildRequired) {
             try {
               await characterDictionaryAutoSyncRuntime.runSyncNow();
-              characterDictionaryImageLookup.invalidate();
             } catch (error) {
               logger.warn('Failed to rebuild character dictionary after manager override:', error);
             }
@@ -5689,7 +5740,6 @@ const { registerIpcRuntimeHandlers } = composeIpcRuntimeHandlers({
         if (result.ok && result.rebuildRequired) {
           try {
             await characterDictionaryAutoSyncRuntime.runSyncNow();
-            characterDictionaryImageLookup.invalidate();
           } catch (error) {
             logger.warn('Failed to rebuild character dictionary after manager removal:', error);
           }
@@ -5706,7 +5756,6 @@ const { registerIpcRuntimeHandlers } = composeIpcRuntimeHandlers({
         if (result.ok && result.rebuildRequired) {
           try {
             await characterDictionaryAutoSyncRuntime.runSyncNow();
-            characterDictionaryImageLookup.invalidate();
           } catch (error) {
             logger.warn('Failed to rebuild character dictionary after manager reorder:', error);
           }

src/main/character-dictionary-runtime/constants.ts

@@ -1,7 +1,7 @@
 export const ANILIST_GRAPHQL_URL = 'https://graphql.anilist.co';
 export const ANILIST_REQUEST_DELAY_MS = 2000;
 export const CHARACTER_IMAGE_DOWNLOAD_DELAY_MS = 250;
-export const CHARACTER_DICTIONARY_FORMAT_VERSION = 19;
+export const CHARACTER_DICTIONARY_FORMAT_VERSION = 20;
 export const CHARACTER_DICTIONARY_MERGED_TITLE = 'SubMiner Character Dictionary';
 
 export const HONORIFIC_SUFFIXES = [

src/main/character-dictionary-runtime/name-candidates.test.ts

@@ -0,0 +1,163 @@
+import assert from 'node:assert/strict';
+import * as fs from 'fs';
+import * as os from 'os';
+import * as path from 'path';
+import test from 'node:test';
+import { CHARACTER_DICTIONARY_FORMAT_VERSION } from './constants';
+import { createCharacterNameCandidateLookup } from './name-candidates';
+
+function writeSnapshot(outputDir: string, mediaId: number, entries: Array<[string, string]>): void {
+  const snapshotsDir = path.join(outputDir, 'snapshots');
+  fs.mkdirSync(snapshotsDir, { recursive: true });
+  fs.writeFileSync(
+    path.join(snapshotsDir, `anilist-${mediaId}.json`),
+    JSON.stringify({
+      formatVersion: CHARACTER_DICTIONARY_FORMAT_VERSION,
+      mediaId,
+      mediaTitle: `title-${mediaId}`,
+      entryCount: entries.length,
+      updatedAt: 1,
+      termEntries: entries.map(([term, reading]) => [
+        term,
+        reading,
+        'name main',
+        '',
+        100,
+        [],
+        0,
+        '',
+      ]),
+      images: [],
+    }),
+  );
+}
+
+function withTempDir<T>(run: (dir: string) => T): T {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-name-candidates-'));
+  try {
+    return run(dir);
+  } finally {
+    fs.rmSync(dir, { recursive: true, force: true });
+  }
+}
+
+test('collects terms and readings for the current media', () => {
+  withTempDir((dir) => {
+    writeSnapshot(dir, 1, [
+      ['ミナト', 'みなと'],
+      ['湊', 'みなと'],
+    ]);
+    writeSnapshot(dir, 2, [['カズマ', 'かずま']]);
+
+    const lookup = createCharacterNameCandidateLookup({
+      outputDir: dir,
+      getCurrentMediaId: () => 1,
+    });
+    const candidates = lookup.get();
+
+    assert.ok(candidates);
+    assert.deepEqual([...candidates.forms].sort(), ['みなと', 'ミナト', '湊'].sort());
+    // Deduplicated: both entries share the みなと reading.
+    assert.equal(candidates.forms.length, 3);
+  });
+});
+
+test('returns null without a media scope so the scanner stays exhaustive', () => {
+  withTempDir((dir) => {
+    writeSnapshot(dir, 1, [['ミナト', 'みなと']]);
+
+    const lookup = createCharacterNameCandidateLookup({
+      outputDir: dir,
+      getCurrentMediaId: () => null,
+    });
+
+    assert.equal(lookup.get(), null);
+  });
+});
+
+test('returns null for a media with no cached snapshot', () => {
+  withTempDir((dir) => {
+    writeSnapshot(dir, 1, [['ミナト', 'みなと']]);
+
+    const lookup = createCharacterNameCandidateLookup({
+      outputDir: dir,
+      getCurrentMediaId: () => 999,
+    });
+
+    assert.equal(lookup.get(), null);
+  });
+});
+
+test('key changes when the snapshot content changes', () => {
+  withTempDir((dir) => {
+    writeSnapshot(dir, 1, [['ミナト', 'みなと']]);
+    const lookup = createCharacterNameCandidateLookup({
+      outputDir: dir,
+      getCurrentMediaId: () => 1,
+    });
+    const first = lookup.get();
+
+    writeSnapshot(dir, 1, [
+      ['ミナト', 'みなと'],
+      ['アクア', 'あくあ'],
+    ]);
+    lookup.invalidate();
+    const second = lookup.get();
+
+    assert.ok(first && second);
+    assert.notEqual(first.key, second.key);
+    assert.equal(second.forms.length, 4);
+  });
+});
+
+// The lookup runs once per subtitle line, so it must not stat the snapshot
+// directory every call. Asserted behaviorally: an unannounced on-disk change is
+// invisible until the recheck interval elapses, which can only be true if the
+// filesystem is not consulted per lookup.
+test('does not re-read the snapshot directory on every lookup', () => {
+  withTempDir((dir) => {
+    writeSnapshot(dir, 1, [['ミナト', 'みなと']]);
+    let nowMs = 1_000_000;
+    const lookup = createCharacterNameCandidateLookup({
+      outputDir: dir,
+      getCurrentMediaId: () => 1,
+      now: () => nowMs,
+    });
+
+    assert.equal(lookup.get()?.forms.length, 2);
+
+    writeSnapshot(dir, 1, [
+      ['ミナト', 'みなと'],
+      ['アクア', 'あくあ'],
+    ]);
+
+    nowMs += 1000;
+    assert.equal(lookup.get()?.forms.length, 2, 'expected the cached list within the interval');
+
+    nowMs += 10_000;
+    assert.equal(lookup.get()?.forms.length, 4, 'expected a refresh past the interval');
+  });
+});
+
+test('invalidate picks up a snapshot change immediately', () => {
+  withTempDir((dir) => {
+    writeSnapshot(dir, 1, [['ミナト', 'みなと']]);
+    let nowMs = 1_000_000;
+    const lookup = createCharacterNameCandidateLookup({
+      outputDir: dir,
+      getCurrentMediaId: () => 1,
+      now: () => nowMs,
+    });
+
+    assert.equal(lookup.get()?.forms.length, 2);
+
+    writeSnapshot(dir, 1, [
+      ['ミナト', 'みなと'],
+      ['アクア', 'あくあ'],
+    ]);
+    nowMs += 1;
+    lookup.invalidate();
+
+    assert.equal(lookup.get()?.forms.length, 4);
+  });
+});

src/main/character-dictionary-runtime/name-candidates.ts

@@ -0,0 +1,159 @@
+import * as fs from 'fs';
+import * as path from 'path';
+import { readCachedSnapshots } from './cache';
+import type { CharacterDictionarySnapshot } from './types';
+
+// Candidate name forms for the greedy name pre-pass in the Yomitan scan
+// runtime. The scanner otherwise has to ask the backend at every Japanese
+// position, because a character name can start mid-token; knowing which forms
+// exist lets it look up only where a name can actually begin.
+//
+// A form is any string Yomitan could match a character entry by: the term and
+// its reading. Both come from the dictionary SubMiner generated, so the pair is
+// the complete matchable set for an entry. Callers treat a missing list as
+// "scan every position", so a stale or absent snapshot costs speed, never a
+// missed name.
+
+function getSnapshotsDir(outputDir: string): string {
+  return path.join(outputDir, 'snapshots');
+}
+
+function collectSnapshotNameForms(snapshot: CharacterDictionarySnapshot): string[] {
+  const forms = new Set<string>();
+  for (const entry of snapshot.termEntries) {
+    const term = typeof entry[0] === 'string' ? entry[0].trim() : '';
+    if (term) {
+      forms.add(term);
+    }
+    const reading = typeof entry[1] === 'string' ? entry[1].trim() : '';
+    if (reading) {
+      forms.add(reading);
+    }
+  }
+  return [...forms];
+}
+
+// The signature grows with the size of the dictionary library, and it rides
+// along in every per-line scan call, so it is folded into a fixed-width digest
+// first. Collisions only matter against the immediately previous signature (the
+// runtime compares keys for equality), and FNV-1a over the file list is far
+// beyond what that needs.
+function digestSnapshotDirectorySignature(signature: string): string {
+  let hash = 0x811c9dc5;
+  for (let index = 0; index < signature.length; index += 1) {
+    hash ^= signature.charCodeAt(index);
+    hash = Math.imul(hash, 0x01000193);
+  }
+  return (hash >>> 0).toString(36);
+}
+
+function getSnapshotDirectorySignature(outputDir: string): string {
+  let entries: fs.Dirent[] = [];
+  try {
+    entries = fs.readdirSync(getSnapshotsDir(outputDir), { withFileTypes: true });
+  } catch {
+    return '';
+  }
+
+  const parts: string[] = [];
+  for (const entry of entries) {
+    if (!entry.isFile() || !/^anilist-\d+\.json$/.test(entry.name)) {
+      continue;
+    }
+    try {
+      const stat = fs.statSync(path.join(getSnapshotsDir(outputDir), entry.name));
+      parts.push(`${entry.name}:${stat.mtimeMs}:${stat.size}`);
+    } catch {
+      // Ignore files that disappear during a refresh; the next lookup rebuilds.
+    }
+  }
+  return parts.sort().join('|');
+}
+
+export interface CharacterNameCandidateSet {
+  /** Identifies this exact form list, so the scan runtime can cache it. */
+  key: string;
+  forms: string[];
+}
+
+// This lookup is consulted once per subtitle line, so it must not stat the
+// snapshot directory every time. Dictionary writes are rare and always call
+// invalidate(), which forces the next lookup to re-read; the interval only
+// bounds staleness from changes made behind our back.
+const SNAPSHOT_SIGNATURE_RECHECK_INTERVAL_MS = 5000;
+
+export function createCharacterNameCandidateLookup(deps: {
+  userDataPath?: string;
+  outputDir?: string;
+  getCurrentMediaId?: () => number | null | undefined;
+  now?: () => number;
+}): {
+  get: (mediaId?: number | null) => CharacterNameCandidateSet | null;
+  invalidate: () => void;
+} {
+  const outputDir =
+    deps.outputDir ??
+    (deps.userDataPath ? path.join(deps.userDataPath, 'character-dictionaries') : '');
+  const now = deps.now ?? (() => Date.now());
+  let signature: string | null = null;
+  let lastSignatureCheckAtMs = 0;
+  let formsByMediaId = new Map<number, string[]>();
+
+  function refreshIfNeeded(): void {
+    if (!outputDir) {
+      formsByMediaId = new Map<number, string[]>();
+      signature = '';
+      return;
+    }
+    const nowMs = now();
+    if (
+      signature !== null &&
+      nowMs - lastSignatureCheckAtMs < SNAPSHOT_SIGNATURE_RECHECK_INTERVAL_MS
+    ) {
+      return;
+    }
+    lastSignatureCheckAtMs = nowMs;
+    const nextSignature = getSnapshotDirectorySignature(outputDir);
+    if (nextSignature === signature) {
+      return;
+    }
+    signature = nextSignature;
+    formsByMediaId = new Map<number, string[]>();
+    for (const snapshot of readCachedSnapshots(outputDir)) {
+      const forms = collectSnapshotNameForms(snapshot);
+      if (forms.length > 0) {
+        formsByMediaId.set(snapshot.mediaId, forms);
+      }
+    }
+  }
+
+  return {
+    get(mediaId?: number | null): CharacterNameCandidateSet | null {
+      refreshIfNeeded();
+      const rawMediaId = mediaId ?? deps.getCurrentMediaId?.() ?? null;
+      const normalizedMediaId =
+        typeof rawMediaId === 'number' && Number.isFinite(rawMediaId) && rawMediaId > 0
+          ? Math.floor(rawMediaId)
+          : null;
+
+      // Without a media scope the pre-pass would need every character of every
+      // cached title, which is both slow to match and pointless: report no
+      // candidates so the scanner keeps its exhaustive behavior.
+      if (normalizedMediaId === null) {
+        return null;
+      }
+      const forms = formsByMediaId.get(normalizedMediaId);
+      if (!forms || forms.length === 0) {
+        return null;
+      }
+      return {
+        key: `${digestSnapshotDirectorySignature(signature ?? '')}:${normalizedMediaId}`,
+        forms,
+      };
+    },
+    invalidate(): void {
+      signature = null;
+      lastSignatureCheckAtMs = 0;
+    },
+  };
+}

src/main/character-dictionary-runtime/name-reading.ts

@@ -1,3 +1,4 @@
+import { isHanCodePoint } from '../../core/text/han-code-points';
 import { HONORIFIC_SUFFIXES } from './constants';
 import type { JapaneseNameParts, NameReadings, ResolvedNameSplits } from './types';
 
@@ -26,10 +27,12 @@ export function buildReading(term: string): string {
   return katakanaToHiragana(compact);
 }
 
+// Code points, not code units: a supplementary-plane kanji (𠮷, U+20BB7) is a
+// surrogate pair, and reading only the high surrogate would classify a real
+// single-character name as non-kanji and drop it.
 export function containsKanji(value: string): boolean {
   for (const char of value) {
-    const code = char.charCodeAt(0);
+    if (isHanCodePoint(char.codePointAt(0) ?? 0)) {
-    if ((code >= 0x4e00 && code <= 0x9fff) || (code >= 0x3400 && code <= 0x4dbf)) {
       return true;
     }
   }

src/main/character-dictionary-runtime/term-building.test.ts

@@ -36,3 +36,75 @@ test('buildNameTerms adds surname honorifics from Japanese localized aliases', (
   assert.ok(terms.includes('馬渕さん'));
   assert.ok(!terms.includes('송치'));
 });
+
+test('buildNameTerms drops the disambiguator letter of a mob character name', () => {
+  const terms = buildNameTerms(
+    characterRecord({
+      firstNameHint: '',
+      lastNameHint: '',
+      fullName: 'Joshi A',
+      nativeName: '女子A',
+    }),
+  );
+
+  // ア would match every あ〜 in the subtitles; the letter is a disambiguator
+  // (Girl A / Girl B), not a name.
+  assert.ok(!terms.includes('ア'));
+  assert.ok(!terms.includes('アさん'));
+  assert.ok(terms.includes('女子A'));
+  assert.ok(terms.includes('ジョシア'));
+});
+
+test('buildNameTerms keeps a character whose whole name is one kana', () => {
+  const terms = buildNameTerms(
+    characterRecord({
+      firstNameHint: '',
+      lastNameHint: '',
+      fullName: 'A',
+      nativeName: 'あ',
+    }),
+  );
+
+  // The mob-disambiguator filter targets letters split off a longer name; an
+  // explicit one-character name is the character's actual name.
+  assert.ok(terms.includes('あ'));
+  assert.ok(terms.includes('あさん'));
+  // The romanized "A" is still a label, so it contributes neither itself nor
+  // its single-kana alias.
+  assert.ok(!terms.includes('A'));
+  assert.ok(!terms.includes('ア'));
+});
+
+test('buildNameTerms keeps a single-kanji name part', () => {
+  // The name is an alias, not the native name, so the parts come from the
+  // space split rather than from the native-name split.
+  const terms = buildNameTerms(
+    characterRecord({
+      firstNameHint: 'Sora',
+      lastNameHint: 'Yamada',
+      fullName: 'Sora Yamada',
+      nativeName: '',
+      alternativeNames: ['山田 空'],
+    }),
+  );
+
+  assert.ok(terms.includes('山田'));
+  assert.ok(terms.includes('空'));
+});
+
+test('buildNameTerms keeps a single supplementary-plane kanji name part', () => {
+  // 𠮷 (U+20BB7) is a surrogate pair: a code-unit kanji check reads only the
+  // high surrogate and drops the part as if it were a mob disambiguator.
+  const terms = buildNameTerms(
+    characterRecord({
+      firstNameHint: 'Tsukasa',
+      lastNameHint: 'Yoshi',
+      fullName: 'Tsukasa Yoshi',
+      nativeName: '',
+      alternativeNames: ['𠮷 司'],
+    }),
+  );
+
+  assert.ok(terms.includes('𠮷'));
+  assert.ok(terms.includes('司'));
+});

src/main/character-dictionary-runtime/term-building.ts

@@ -1,3 +1,4 @@
+import { HAN_REGEXP_CLASS_BODY } from '../../core/text/han-code-points';
 import { HONORIFIC_SUFFIXES } from './constants';
 import {
   addRomanizedKanaAliases,
@@ -42,11 +43,34 @@ export function expandRawNameVariants(rawName: string): string[] {
   return [...variants];
 }
 
+// Kana, halfwidth included: one of these can stand alone as a name, where a
+// latin letter or a digit cannot.
+const SINGLE_KANA_CHARACTER = /^[\u3040-\u30ff\u31f0-\u31ff\uff66-\uff9f]$/u;
+
+// AniList disambiguates unnamed mob characters with a trailing letter (女子A /
+// "Joshi A"), and a lone letter romanizes into a single-kana alias (A → ア)
+// that collides with interjections (あ〜 matching ア). That letter is a label,
+// not a name, so it is dropped where a name splits into it and before it can
+// become a kana alias. A name that is genuinely one character, a character
+// actually called あ or a single kanji, is a real lookup target and is kept.
+function isNameDisambiguatorLetter(name: string): boolean {
+  return [...name].length === 1 && !containsKanji(name) && !SINGLE_KANA_CHARACTER.test(name);
+}
+
+function isUsableNameTerm(name: string): boolean {
+  return !isNameDisambiguatorLetter(name);
+}
+
+// Kana, Han (shared ranges), and the marks that only ever appear inside a
+// Japanese name: iteration marks and the small ka/ke used in place names.
+const JAPANESE_NAME_CHARACTERS = new RegExp(
+  `^[\\u3040-\\u30ff${HAN_REGEXP_CLASS_BODY}\u3005\u3006\u30f5\u30f6\u30fc]+$`,
+  'u',
+);
+
 export function isJapaneseNameSplitCandidate(name: string): boolean {
   const compact = name.replace(/[\s\u3000・･·•]/g, '');
-  return (
+  return containsKanji(compact) && JAPANESE_NAME_CHARACTERS.test(compact);
-    containsKanji(compact) && /^[\u3040-\u30ff\u3400-\u4dbf\u4e00-\u9fff々〆ヵヶー]+$/.test(compact)
-  );
 }
 
 function addJapaneseNameParts(
@@ -97,8 +121,11 @@ export function buildNameTerms(
 
       const split = name.split(/[\s\u3000]+/).filter((part) => part.trim().length > 0);
       if (split.length === 2) {
-        target.add(split[0]!);
+        for (const part of split) {
-        target.add(split[1]!);
+          if (isUsableNameTerm(part)) {
+            target.add(part);
+          }
+        }
       }
 
       const splitByMiddleDot = name
@@ -107,9 +134,11 @@ export function buildNameTerms(
         .filter((part) => part.length > 0);
       if (splitByMiddleDot.length >= 2) {
         for (const part of splitByMiddleDot) {
+          if (isUsableNameTerm(part)) {
             target.add(part);
           }
         }
+      }
 
       if (target === base) {
         addJapaneseNameParts(character, name, base, resolvedSplits);
@@ -117,7 +146,10 @@ export function buildNameTerms(
     }
   }
 
-  for (const alias of addRomanizedKanaAliases(romanizedBase)) {
+  // Romanized forms that are a bare letter would become a single-kana alias.
+  for (const alias of addRomanizedKanaAliases(
+    [...romanizedBase].filter((entry) => !isNameDisambiguatorLetter(entry)),
+  )) {
     base.add(alias);
   }
 
@@ -136,6 +168,9 @@ export function buildNameTerms(
 
   const withHonorifics = new Set<string>();
   for (const entry of base) {
+    // Only labels split off a longer name are filtered (see above); an explicit
+    // one-character name reaches this point intact.
+    if (isNameDisambiguatorLetter(entry)) continue;
     withHonorifics.add(entry);
     for (const suffix of HONORIFIC_SUFFIXES) {
       withHonorifics.add(`${entry}${suffix.term}`);

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');
@@ -200,7 +223,7 @@ test('update overlay notification action triggers install flow', () => {
   assert.match(runtimeSource, /fallbackClient\.openNoteInBrowser\(noteId\)/);
 });
 
-test('subtitle change re-prioritizes prefetch around live playback before tokenizing current line', () => {
+test('subtitle change pauses prefetch without restarting its run before tokenizing current line', () => {
   const source = readMainSource();
   const actionBlock = source.match(
     /onSubtitleChange:\s*\(text\)\s*=>\s*\{(?<body>[\s\S]*?)\n    \},\n    refreshDiscordPresence:/,
@@ -208,15 +231,19 @@ test('subtitle change re-prioritizes prefetch around live playback before tokeni
 
   assert.ok(actionBlock);
   assert.match(actionBlock, /subtitlePrefetchService\?\.pause\(\);/);
-  assert.match(actionBlock, /subtitlePrefetchService\?\.onSeek\(lastObservedTimePos\);/);
+  // Restarting the run per line (onSeek) discards in-flight prefetch work;
-  assert.match(actionBlock, /subtitleProcessingController\.onSubtitleChange\(text\);/);
+  // only real seeks restart via onTimePosUpdate.
+  assert.doesNotMatch(actionBlock, /subtitlePrefetchService\?\.onSeek\(/);
+  assert.match(actionBlock, /subtitleProcessingController\.onSubtitleChange\(text\)/);
   assert.ok(
     actionBlock.indexOf('subtitlePrefetchService?.pause();') <
-      actionBlock.indexOf('subtitlePrefetchService?.onSeek(lastObservedTimePos);'),
+      actionBlock.indexOf('subtitleProcessingController.onSubtitleChange(text)'),
   );
-  assert.ok(
+  // A repeated subtitle emits nothing, so the pause has to be released here or
-    actionBlock.indexOf('subtitlePrefetchService?.onSeek(lastObservedTimePos);') <
+  // prefetching idles until the next distinct line.
-      actionBlock.indexOf('subtitleProcessingController.onSubtitleChange(text);'),
+  assert.match(
+    actionBlock,
+    /if \(!subtitleProcessingController\.onSubtitleChange\(text\)\) \{[\s\S]*?subtitlePrefetchService\?\.resume\(\);/,
   );
 });
 
@@ -466,16 +493,35 @@ test('known-word updates invalidate prefetched tokenizations before refreshing c
   assert.match(actionBlock, /subtitlePrefetchService\?\.onSeek\(lastObservedTimePos\);/);
   assert.match(
     actionBlock,
-    /subtitleProcessingController\.refreshCurrentSubtitle\(appState\.currentSubText\);/,
+    /if \(!subtitleProcessingController\.refreshCurrentSubtitle\(appState\.currentSubText\)\) \{[\s\S]*?subtitlePrefetchService\?\.resume\(\);/,
   );
   assert.ok(
     actionBlock.indexOf('subtitleProcessingController.invalidateTokenizationCache();') <
       actionBlock.indexOf(
-        'subtitleProcessingController.refreshCurrentSubtitle(appState.currentSubText);',
+        'subtitleProcessingController.refreshCurrentSubtitle(appState.currentSubText)',
       ),
   );
 });
 
+test('subtitle processing controller resumes prefetch on settle, not on its emits', () => {
+  const source = readMainSource();
+  const depsBlock = source.match(
+    /createBuildSubtitleProcessingControllerMainDepsHandler\(\{(?<body>[\s\S]*?)\n  \}\);/,
+  )?.groups?.body;
+
+  assert.ok(depsBlock);
+  // A controller emit can be the provisional plain payload sent before the
+  // scan runs, so it must not release the prefetch pause.
+  assert.match(
+    depsBlock,
+    /emitSubtitle: \(payload\) => emitSubtitlePayload\(payload, \{ resumePrefetch: false \}\),/,
+  );
+  assert.match(
+    depsBlock,
+    /onProcessingSettled: \(\) => \{\s+subtitlePrefetchService\?\.resume\(\);/,
+  );
+});
+
 test('manual visible overlay changes notify mpv plugin visibility state', () => {
   const source = readMainSource();
   const setBlock = source.match(
@@ -570,7 +616,7 @@ test('YouTube media cache lifecycle routes through configured status notificatio
 test('subtitle broadcasts share one frequency options snapshot per emitted payload', () => {
   const source = readMainSource();
   const emitBlock = source.match(
-    /function emitSubtitlePayload\(payload: SubtitleData\): void \{(?<body>[\s\S]*?)\n\}/,
+    /function emitSubtitlePayload\([\s\S]*?\): void \{(?<body>[\s\S]*?)\n\}/,
   )?.groups?.body;
   const frequencyOptionsSnapshot = emitBlock?.match(
     /const frequencyDictionary = configService\.getConfig\(\)\.subtitleStyle\.frequencyDictionary;(?<body>[\s\S]*?)\n  \};/,
@@ -590,6 +636,35 @@ test('subtitle broadcasts share one frequency options snapshot per emitted paylo
   );
 });
 
+test('annotation upgrades skip the duplicate basic websocket event', () => {
+  const source = readMainSource();
+  const emitBlock = source.match(
+    /function emitSubtitlePayload\([\s\S]*?\): void \{(?<body>[\s\S]*?)\n\}/,
+  )?.groups?.body;
+
+  assert.ok(emitBlock);
+  assert.match(
+    emitBlock,
+    /const isAnnotationUpgrade = isSubtitleAnnotationUpgrade\(currentSubtitleData, timedPayload\);/,
+  );
+  assert.match(
+    emitBlock,
+    /if \(!isAnnotationUpgrade\) \{\s+subtitleWsService\.broadcast\(timedPayload, frequencyOptions\);\s+\}/,
+  );
+  assert.equal(
+    (emitBlock.match(/overlayManager\.broadcastToOverlayWindows\('subtitle:set'/g) ?? []).length,
+    1,
+  );
+  assert.equal(
+    (
+      emitBlock.match(
+        /annotationSubtitleWsService\.broadcast\(timedPayload, frequencyOptions\)/g,
+      ) ?? []
+    ).length,
+    1,
+  );
+});
+
 test('websocket frequency options callbacks each read one configuration snapshot', () => {
   const source = readMainSource();
   const subtitleBlock = source.match(

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

@@ -1,5 +1,7 @@
 import assert from 'node:assert/strict';
 import test from 'node:test';
+import { createSubtitleProcessingController } from '../../core/services/subtitle-processing-controller';
+import type { SubtitleData } from '../../types';
 import {
   createAutoplaySubtitlePrimingRuntime,
   setMpvCurrentSecondarySubText,
@@ -42,8 +44,8 @@ test('scheduleSubtitlePrefetchRefresh logs refresh failures from timer callback'
     setActiveParsedSubtitleMediaPath: () => {},
     subtitleProcessingController: {
       consumeCachedSubtitle: () => null,
-      onSubtitleChange: () => {},
+      onSubtitleChange: () => true,
-      refreshCurrentSubtitle: () => {},
+      refreshCurrentSubtitle: () => true,
     },
     emitSubtitlePayload: () => {},
     getSubtitlePrefetchService: () => null,
@@ -93,13 +95,24 @@ test('primeCurrentSubtitleForAutoplay refreshes active subtitle cues when mpv su
     setActiveParsedSubtitleMediaPath: () => {},
     subtitleProcessingController: {
       consumeCachedSubtitle: () => null,
-      onSubtitleChange: (text) => calls.push(`change:${text}`),
+      onSubtitleChange: (text) => {
-      refreshCurrentSubtitle: (text) => calls.push(`refresh:${text ?? ''}`),
+        calls.push(`change:${text}`);
+        return true;
       },
-    emitSubtitlePayload: (payload) => calls.push(`emit:${payload.text}`),
+      refreshCurrentSubtitle: (text) => {
+        calls.push(`refresh:${text ?? ''}`);
+        return true;
+      },
+    },
+    emitSubtitlePayload: (payload, options) =>
+      calls.push(`emit:${payload.text}:resume=${options?.resumePrefetch !== false}`),
     getSubtitlePrefetchService: () => ({
-      pause: () => calls.push('prefetch:pause'),
+      pause: () => {
-      onSeek: (timePos) => calls.push(`prefetch:seek:${timePos}`),
+        calls.push('prefetch:pause');
+      },
+      resume: () => {
+        calls.push('prefetch:resume');
+      },
     }),
     getLastObservedTimePos: () => 12,
     getVisibleOverlayVisible: () => true,
@@ -120,8 +133,10 @@ test('primeCurrentSubtitleForAutoplay refreshes active subtitle cues when mpv su
     'request:time-pos',
     'set:起動字幕',
     'prefetch:pause',
-    'emit:起動字幕',
+    'emit:起動字幕:resume=false',
-    'change:起動字幕',
+    // Uncached priming refreshes rather than announcing a change, so an
+    // invalidated-but-unchanged line is still re-tokenized.
+    'refresh:起動字幕',
   ]);
 });
 
@@ -151,13 +166,24 @@ test('primeCurrentSubtitleForAutoplay emits raw first paint on cache miss before
     setActiveParsedSubtitleMediaPath: () => {},
     subtitleProcessingController: {
       consumeCachedSubtitle: () => null,
-      onSubtitleChange: (text) => calls.push(`change:${text}`),
+      onSubtitleChange: (text) => {
-      refreshCurrentSubtitle: (text) => calls.push(`refresh:${text ?? ''}`),
+        calls.push(`change:${text}`);
+        return true;
       },
-    emitSubtitlePayload: (payload) => calls.push(`emit:${payload.text}`),
+      refreshCurrentSubtitle: (text) => {
+        calls.push(`refresh:${text ?? ''}`);
+        return true;
+      },
+    },
+    emitSubtitlePayload: (payload, options) =>
+      calls.push(`emit:${payload.text}:resume=${options?.resumePrefetch !== false}`),
     getSubtitlePrefetchService: () => ({
-      pause: () => calls.push('prefetch:pause'),
+      pause: () => {
-      onSeek: (timePos) => calls.push(`prefetch:seek:${timePos}`),
+        calls.push('prefetch:pause');
+      },
+      resume: () => {
+        calls.push('prefetch:resume');
+      },
     }),
     getLastObservedTimePos: () => 12,
     getVisibleOverlayVisible: () => true,
@@ -175,7 +201,222 @@ test('primeCurrentSubtitleForAutoplay emits raw first paint on cache miss before
     'request:sub-text',
     'set:起動字幕',
     'prefetch:pause',
-    'emit:起動字幕',
+    'emit:起動字幕:resume=false',
-    'change:起動字幕',
+    // Uncached priming refreshes rather than announcing a change, so an
+    // invalidated-but-unchanged line is still re-tokenized.
+    'refresh:起動字幕',
+  ]);
+});
+
+// Driven by the real processing controller rather than a stub: the failure this
+// covers is a disagreement between the priming path and the controller's own
+// staleness rules, which a hand-written stub cannot reproduce.
+function createPrimingRuntimeWithRealController(options: {
+  text: string;
+  calls: string[];
+  onTokenize: () => void;
+  tokenize?: (text: string) => SubtitleData | null | Promise<SubtitleData | null>;
+  cacheLimit?: number;
+}) {
+  const { text, calls } = options;
+  let currentSubText = '';
+  let currentSubtitleData: SubtitleData | null = null;
+  const mediaPath = '/media/video.mkv';
+
+  const prefetchService = {
+    pause: () => calls.push('prefetch:pause'),
+    resume: () => calls.push('prefetch:resume'),
+  };
+  // Mirrors main.ts emitSubtitlePayload: an emit resumes prefetching unless it
+  // is explicitly marked as not the end of the work for the line, and every
+  // controller emit is so marked.
+  const emitSubtitlePayload = (
+    payload: SubtitleData,
+    emitOptions?: { resumePrefetch?: boolean },
+  ): void => {
+    currentSubtitleData = payload;
+    calls.push(
+      emitOptions?.resumePrefetch === false
+        ? `emit-raw:${payload.text}`
+        : `emit-direct:${payload.text}`,
+    );
+    if (emitOptions?.resumePrefetch !== false) {
+      prefetchService.resume();
+    }
+  };
+
+  const subtitleProcessingController = createSubtitleProcessingController({
+    tokenizeSubtitle: async (subtitleText) => {
+      options.onTokenize();
+      return options.tokenize ? options.tokenize(subtitleText) : { text: subtitleText, tokens: [] };
+    },
+    // main.ts routes controller emits through emitSubtitlePayload with
+    // resumePrefetch: false, so they never release the pause on their own.
+    emitSubtitle: (payload) => {
+      currentSubtitleData = payload;
+      calls.push(`emit:${payload.text}:tokens=${payload.tokens === null ? 'none' : 'yes'}`);
+    },
+    onProcessingSettled: () => {
+      prefetchService.resume();
+    },
+    ...(options.cacheLimit === undefined ? {} : { cacheLimit: options.cacheLimit }),
+  });
+
+  const runtime = createAutoplaySubtitlePrimingRuntime({
+    getCurrentMediaPath: () => mediaPath,
+    getMpvClient: () => ({
+      connected: true,
+      currentVideoPath: mediaPath,
+      requestProperty: async (name) => (name === 'sub-text' ? text : null),
+    }),
+    setCurrentSubText: (value) => {
+      currentSubText = value;
+    },
+    getCurrentSubText: () => currentSubText,
+    getCurrentSubtitleData: () => currentSubtitleData,
+    getActiveParsedSubtitleCues: () => [],
+    setActiveParsedSubtitleMediaPath: () => {},
+    subtitleProcessingController,
+    emitSubtitlePayload,
+    getSubtitlePrefetchService: () => prefetchService,
+    getLastObservedTimePos: () => 12,
+    getVisibleOverlayVisible: () => true,
+    emitSecondarySubtitle: () => {},
+    initSubtitlePrefetch: async () => {},
+    refreshSubtitlePrefetchFromActiveTrack: async () => {},
+    logDebug: () => {},
+  });
+
+  return { runtime, subtitleProcessingController, mediaPath };
+}
+
+test('primeCurrentSubtitleForAutoplay re-tokenizes text whose cached annotation was invalidated', async () => {
+  const calls: string[] = [];
+  let tokenizations = 0;
+  const text = '起動字幕';
+  const { runtime, subtitleProcessingController, mediaPath } =
+    createPrimingRuntimeWithRealController({
+      text,
+      calls,
+      onTokenize: () => {
+        tokenizations += 1;
+      },
+    });
+
+  // The line was already tokenized and cached during normal playback.
+  subtitleProcessingController.onSubtitleChange(text);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+  const tokenizationsBeforeInvalidation = tokenizations;
+
+  // Mining a card drops every cached tokenization.
+  subtitleProcessingController.invalidateTokenizationCache();
+  calls.length = 0;
+
+  await runtime.primeCurrentSubtitleForAutoplay(mediaPath);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+
+  // The cache miss must schedule fresh work, or the line stays unannotated for
+  // as long as it is on screen.
+  assert.equal(
+    tokenizations,
+    tokenizationsBeforeInvalidation + 1,
+    'expected the invalidated subtitle to be tokenized again',
+  );
+  assert.ok(
+    calls.includes(`emit:${text}:tokens=yes`),
+    `expected an annotated emit, saw ${JSON.stringify(calls)}`,
+  );
+});
+
+test('primeCurrentSubtitleForAutoplay releases the prefetch pause when nothing is scheduled', async () => {
+  const calls: string[] = [];
+  const text = '起動字幕';
+  const { runtime, subtitleProcessingController, mediaPath } =
+    createPrimingRuntimeWithRealController({
+      text,
+      calls,
+      onTokenize: () => {},
+      cacheLimit: 1,
+    });
+
+  // Emitted at the current cache generation, then evicted from the one-entry
+  // cache: priming misses the cache but the controller has nothing to redo, so
+  // no emit is coming and the pause must be released here.
+  subtitleProcessingController.onSubtitleChange(text);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+  subtitleProcessingController.preCacheTokenization('別の字幕', {
+    text: '別の字幕',
+    tokens: [],
+  });
+  calls.length = 0;
+
+  await runtime.primeCurrentSubtitleForAutoplay(mediaPath);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+
+  assert.deepEqual(calls, ['prefetch:pause', `emit-raw:${text}`, 'prefetch:resume']);
+});
+
+test('primeCurrentSubtitleForAutoplay releases the prefetch pause when tokenization emits nothing', async () => {
+  const calls: string[] = [];
+  const text = '起動字幕';
+  const { runtime, subtitleProcessingController, mediaPath } =
+    createPrimingRuntimeWithRealController({
+      text,
+      calls,
+      onTokenize: () => {},
+      // Transient tokenizer failure: the controller falls back to plain text it
+      // has already shown, so it suppresses the emit entirely.
+      tokenize: () => null,
+    });
+
+  subtitleProcessingController.onSubtitleChange(text);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+  subtitleProcessingController.invalidateTokenizationCache();
+  calls.length = 0;
+
+  await runtime.primeCurrentSubtitleForAutoplay(mediaPath);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+
+  assert.ok(
+    !calls.some((call) => call.startsWith('emit:')),
+    `expected no controller emit, saw ${JSON.stringify(calls)}`,
+  );
+  assert.equal(
+    calls.filter((call) => call === 'prefetch:resume').length,
+    1,
+    `expected the prefetch pause to be released, saw ${JSON.stringify(calls)}`,
+  );
+});
+
+test('prefetch stays paused until tokenization of an uncached line completes', async () => {
+  const calls: string[] = [];
+  const text = '起動字幕';
+  let finishTokenization = (): void => {};
+  const tokenizationGate = new Promise<void>((resolve) => {
+    finishTokenization = resolve;
+  });
+  const { subtitleProcessingController } = createPrimingRuntimeWithRealController({
+    text,
+    calls,
+    onTokenize: () => {},
+    tokenize: async (subtitleText) => {
+      await tokenizationGate;
+      return { text: subtitleText, tokens: [] };
+    },
+  });
+
+  subtitleProcessingController.onSubtitleChange(text);
+  await new Promise((resolve) => setTimeout(resolve, 0));
+
+  // The provisional plain emit must not release the pause: the expensive scan
+  // is still ahead of it and would compete with prefetching for the parser.
+  assert.deepEqual(calls, [`emit:${text}:tokens=none`]);
+
+  finishTokenization();
+  await new Promise((resolve) => setTimeout(resolve, 0));
+  assert.deepEqual(calls, [
+    `emit:${text}:tokens=none`,
+    `emit:${text}:tokens=yes`,
+    'prefetch:resume',
   ]);
 });

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

@@ -17,7 +17,7 @@ type AutoplaySubtitlePrimingMpvClient = {
 
 type AutoplaySubtitlePrimingPrefetchService = {
   pause: () => void;
-  onSeek: (timePos: number) => void;
+  resume: () => void;
 };
 
 export interface AutoplaySubtitlePrimingRuntimeDeps {
@@ -30,10 +30,11 @@ export interface AutoplaySubtitlePrimingRuntimeDeps {
   setActiveParsedSubtitleMediaPath: (mediaPath: string | null) => void;
   subtitleProcessingController: {
     consumeCachedSubtitle: (text: string) => SubtitleData | null;
-    onSubtitleChange: (text: string) => void;
+    // Both report whether processing is pending; see pausePrefetchUntilProcessed.
-    refreshCurrentSubtitle: (text: string) => void;
+    onSubtitleChange: (text: string) => boolean;
+    refreshCurrentSubtitle: (text: string) => boolean;
   };
-  emitSubtitlePayload: (payload: SubtitleData) => void;
+  emitSubtitlePayload: (payload: SubtitleData, options?: { resumePrefetch?: boolean }) => void;
   getSubtitlePrefetchService: () => AutoplaySubtitlePrimingPrefetchService | null;
   getLastObservedTimePos: () => number;
   getVisibleOverlayVisible: () => boolean;
@@ -64,6 +65,19 @@ export function setMpvCurrentSecondarySubText(
 export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimingRuntimeDeps) {
   const { subtitleProcessingController, emitSubtitlePayload } = deps;
 
+  // Prefetching is paused so the on-screen line gets the parser to itself; the
+  // resume rides on the controller settling (see onProcessingSettled), not on
+  // an emit, which a suppressed duplicate or a failed tokenization never sends.
+  // When the controller reports it has nothing scheduled, no settle is coming
+  // either, so release the pause here or prefetching idles indefinitely.
+  function pausePrefetchUntilProcessed(scheduleTokenization: () => boolean): void {
+    const prefetch = deps.getSubtitlePrefetchService();
+    prefetch?.pause();
+    if (!scheduleTokenization()) {
+      prefetch?.resume();
+    }
+  }
+
   let subtitlePrefetchRefreshTimer: ReturnType<typeof setTimeout> | null = null;
   let autoplaySubtitlePrimedMediaPath: string | null = null;
   let visibleOverlaySubtitleRefreshAfterFirstPaintTimer: ReturnType<typeof setTimeout> | null =
@@ -104,12 +118,23 @@ export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimi
     const cachedPayload = subtitleProcessingController.consumeCachedSubtitle(text);
     if (cachedPayload) {
       subtitleProcessingController.onSubtitleChange(text);
+      // This emit resumes prefetching, so no pause is left outstanding.
       emitSubtitlePayload(cachedPayload);
       return true;
     }
 
-    emitSubtitlePayload({ text, tokens: null });
+    // Provisional raw emit: keep prefetch paused until the processing
-    subtitleProcessingController.onSubtitleChange(text);
+    // controller is done with this line.
+    emitSubtitlePayload({ text, tokens: null }, { resumePrefetch: false });
+    // refreshCurrentSubtitle, not onSubtitleChange: the cache miss above can be
+    // an invalidation (mining a card) on text the controller still holds, and
+    // onSubtitleChange treats unchanged text as nothing to do, which would
+    // leave this line permanently unannotated. refreshCurrentSubtitle also
+    // re-tokenizes for a new cache generation.
+    if (!subtitleProcessingController.refreshCurrentSubtitle(text)) {
+      // Nothing scheduled, so no settle is coming to release the pause.
+      deps.getSubtitlePrefetchService()?.resume();
+    }
     return true;
   }
 
@@ -153,14 +178,12 @@ export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimi
       getCurrentSubtitleData: () => deps.getCurrentSubtitleData(),
       consumeCachedSubtitle: (text) => subtitleProcessingController.consumeCachedSubtitle(text),
       onSubtitleChange: (text) => {
-        deps.getSubtitlePrefetchService()?.pause();
+        pausePrefetchUntilProcessed(() => subtitleProcessingController.onSubtitleChange(text));
-        deps.getSubtitlePrefetchService()?.onSeek(deps.getLastObservedTimePos());
-        subtitleProcessingController.onSubtitleChange(text);
       },
       refreshCurrentSubtitle: (text) => {
-        deps.getSubtitlePrefetchService()?.pause();
+        pausePrefetchUntilProcessed(() =>
-        deps.getSubtitlePrefetchService()?.onSeek(deps.getLastObservedTimePos());
+          subtitleProcessingController.refreshCurrentSubtitle(text),
-        subtitleProcessingController.refreshCurrentSubtitle(text);
+        );
       },
       deferUncachedRefresh: true,
       emitSubtitle: (payload) => emitSubtitlePayload(payload),
@@ -204,9 +227,7 @@ export function createAutoplaySubtitlePrimingRuntime(deps: AutoplaySubtitlePrimi
       if (!text.trim()) {
         return;
       }
-      deps.getSubtitlePrefetchService()?.pause();
+      pausePrefetchUntilProcessed(() => subtitleProcessingController.refreshCurrentSubtitle(text));
-      deps.getSubtitlePrefetchService()?.onSeek(deps.getLastObservedTimePos());
-      subtitleProcessingController.refreshCurrentSubtitle(text);
     }, VISIBLE_OVERLAY_SUBTITLE_REFRESH_AFTER_FIRST_PAINT_DELAY_MS);
     visibleOverlaySubtitleRefreshAfterFirstPaintTimer.unref?.();
   }

src/main/runtime/character-dictionary-auto-sync-completion.test.ts

@@ -53,3 +53,52 @@ test('character dictionary sync completion refreshes subtitle state when diction
     'log:[dictionary:auto-sync] refreshed current subtitle after sync (AniList 1, changed=yes, title=Frieren)',
   ]);
 });
+
+test('character dictionary sync completion drops cached dictionary reads before refreshing', () => {
+  const calls: string[] = [];
+
+  handleCharacterDictionaryAutoSyncComplete(
+    {
+      mediaId: 1,
+      mediaTitle: 'Frieren',
+      changed: true,
+    },
+    {
+      hasParserWindow: () => true,
+      invalidateCharacterDictionaryLookups: () => calls.push('invalidate-dictionary-lookups'),
+      clearParserCaches: () => calls.push('clear-parser'),
+      invalidateTokenizationCache: () => calls.push('invalidate'),
+      refreshSubtitlePrefetch: () => calls.push('prefetch'),
+      refreshCurrentSubtitle: () => calls.push('refresh-subtitle'),
+      logInfo: () => {},
+    },
+  );
+
+  // Must run before the refreshes, or they re-tokenize against the character
+  // names and images from the previous dictionary build.
+  assert.equal(calls[0], 'invalidate-dictionary-lookups');
+  assert.ok(calls.indexOf('invalidate-dictionary-lookups') < calls.indexOf('refresh-subtitle'));
+});
+
+test('character dictionary sync completion leaves cached dictionary reads alone when unchanged', () => {
+  const calls: string[] = [];
+
+  handleCharacterDictionaryAutoSyncComplete(
+    {
+      mediaId: 1,
+      mediaTitle: 'Frieren',
+      changed: false,
+    },
+    {
+      hasParserWindow: () => true,
+      invalidateCharacterDictionaryLookups: () => calls.push('invalidate-dictionary-lookups'),
+      clearParserCaches: () => calls.push('clear-parser'),
+      invalidateTokenizationCache: () => calls.push('invalidate'),
+      refreshSubtitlePrefetch: () => calls.push('prefetch'),
+      refreshCurrentSubtitle: () => calls.push('refresh-subtitle'),
+      logInfo: () => {},
+    },
+  );
+
+  assert.deepEqual(calls, []);
+});

src/main/runtime/character-dictionary-auto-sync-completion.ts

@@ -7,6 +7,12 @@ export function handleCharacterDictionaryAutoSyncComplete(
   deps: {
     hasParserWindow: () => boolean;
     clearParserCaches: () => void;
+    /**
+     * Drops cached reads of the generated dictionary (character images, and the
+     * name candidates the scanner uses to skip lookups). Runs before the
+     * refreshes below so they re-tokenize against the new dictionary content.
+     */
+    invalidateCharacterDictionaryLookups?: () => void;
     invalidateTokenizationCache: () => void;
     refreshSubtitlePrefetch: () => void;
     refreshCurrentSubtitle: () => void;
@@ -14,6 +20,7 @@ export function handleCharacterDictionaryAutoSyncComplete(
   },
 ): void {
   if (completion.changed) {
+    deps.invalidateCharacterDictionaryLookups?.();
     if (deps.hasParserWindow()) {
       deps.clearParserCaches();
     }

src/main/runtime/composers/mpv-runtime-composer.ts

@@ -225,24 +225,35 @@ export function composeMpvRuntimeHandlers<
     }
     return tokenizationWarmupInFlight;
   };
-  const tokenizeSubtitle = async (text: string): Promise<TTokenizedSubtitle> => {
+  // Built once and reused for every tokenization: per-call rebuilds create
-    if (!tokenizationWarmupCompleted) void startTokenizationWarmups();
+  // fresh closures, which defeats identity-keyed caches downstream (the JLPT
-    await ensureTokenizationPrerequisites();
+  // lookup cache keys on the getJlptLevel function, and the mecab availability
+  // WeakSet keys on the runtime deps instance).
+  let cachedTokenizerRuntimeDeps: TTokenizerRuntimeDeps | null = null;
+  const getTokenizerRuntimeDeps = (): TTokenizerRuntimeDeps => {
+    if (cachedTokenizerRuntimeDeps) {
+      return cachedTokenizerRuntimeDeps;
+    }
     const tokenizerMainDeps = buildTokenizerDepsHandler();
-    if (shouldWarmupAnnotationDictionaries()) {
+    const baseOnTokenizationReady = tokenizerMainDeps.onTokenizationReady;
-      const onTokenizationReady = tokenizerMainDeps.onTokenizationReady;
     tokenizerMainDeps.onTokenizationReady = (tokenizedText: string): void => {
+      if (!shouldWarmupAnnotationDictionaries()) {
+        baseOnTokenizationReady?.(tokenizedText);
+        return;
+      }
       markTokenizationPlaybackReady();
-        onTokenizationReady?.(tokenizedText);
+      baseOnTokenizationReady?.(tokenizedText);
       if (!tokenizationWarmupCompleted) {
         void prewarmSubtitleDictionaries({ showLoadingOsd: true }).catch(() => {});
       }
     };
-    }
+    cachedTokenizerRuntimeDeps = options.tokenizer.createTokenizerRuntimeDeps(tokenizerMainDeps);
-    return options.tokenizer.tokenizeSubtitle(
+    return cachedTokenizerRuntimeDeps;
-      text,
+  };
-      options.tokenizer.createTokenizerRuntimeDeps(tokenizerMainDeps),
+  const tokenizeSubtitle = async (text: string): Promise<TTokenizedSubtitle> => {
-    );
+    if (!tokenizationWarmupCompleted) void startTokenizationWarmups();
+    await ensureTokenizationPrerequisites();
+    return options.tokenizer.tokenizeSubtitle(text, getTokenizerRuntimeDeps());
   };
 
   const launchBackgroundWarmupTask = createLaunchBackgroundWarmupTaskFromStartup(

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

@@ -70,6 +70,24 @@ test('subtitle change handler broadcasts cached annotated payload immediately wh
   ]);
 });
 
+test('subtitle change handler logs debug when a cached payload is emitted immediately', () => {
+  const debugs: string[] = [];
+  const handler = createHandleMpvSubtitleChangeHandler({
+    setCurrentSubText: () => {},
+    getImmediateSubtitlePayload: (text) => (text ? { text, tokens: [] } : null),
+    broadcastSubtitle: () => {},
+    onSubtitleChange: () => {},
+    refreshDiscordPresence: () => {},
+    logDebug: (message) => debugs.push(message),
+  });
+
+  handler({ text: 'キャッシュ済みの行' });
+  handler({ text: '' });
+
+  assert.equal(debugs.length, 1);
+  assert.match(debugs[0]!, /cached subtitle/);
+});
+
 test('subtitle change handler emits cached annotation after forwarding the subtitle change', () => {
   const calls: string[] = [];
   const handler = createHandleMpvSubtitleChangeHandler({

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

@@ -20,11 +20,15 @@ export function createHandleMpvSubtitleChangeHandler(deps: {
   broadcastSubtitle: (payload: SubtitleData) => void;
   onSubtitleChange: (text: string) => void;
   refreshDiscordPresence: () => void;
+  logDebug?: (message: string) => void;
 }) {
   return ({ text }: { text: string }): void => {
     deps.setCurrentSubText(text);
     const immediatePayload = deps.getImmediateSubtitlePayload?.(text) ?? null;
     if (immediatePayload) {
+      deps.logDebug?.(
+        `[subtitle-processing] emitted cached subtitle immediately (${text.length} chars)`,
+      );
       deps.onSubtitleChange(text);
       (deps.emitImmediateSubtitle ?? deps.broadcastSubtitle)(immediatePayload);
     } else {

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

@@ -47,6 +47,7 @@ export function createBindMpvMainEventHandlersHandler(deps: {
   emitImmediateSubtitle?: (payload: SubtitleData) => void;
   broadcastSubtitle: (payload: SubtitleData) => void;
   onSubtitleChange: (text: string) => void;
+  logSubtitleProcessingDebug?: (message: string) => void;
   refreshDiscordPresence: () => void;
 
   setCurrentSubAssText: (text: string) => void;
@@ -123,6 +124,9 @@ export function createBindMpvMainEventHandlersHandler(deps: {
         : undefined,
       broadcastSubtitle: (payload) => deps.broadcastSubtitle(payload),
       onSubtitleChange: (text) => deps.onSubtitleChange(text),
+      logDebug: deps.logSubtitleProcessingDebug
+        ? (message) => deps.logSubtitleProcessingDebug?.(message)
+        : undefined,
       refreshDiscordPresence: () => deps.refreshDiscordPresence(),
     });
     const handleMpvSubtitleAssChange = createHandleMpvSubtitleAssChangeHandler({

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

@@ -54,6 +54,7 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
   getImmediateSubtitlePayload?: (text: string) => SubtitleData | null;
   emitImmediateSubtitle?: (payload: SubtitleData) => void;
   onSubtitleChange: (text: string) => void;
+  logSubtitleProcessingDebug?: (message: string) => void;
   onSubtitleTrackChange?: (sid: number | null) => void;
   onSubtitleTrackListChange?: (trackList: unknown[] | null) => void;
   updateCurrentMediaPath: (path: string) => void;
@@ -155,6 +156,9 @@ export function createBuildBindMpvMainEventHandlersMainDepsHandler(deps: {
     broadcastSubtitle: (payload: SubtitleData) =>
       deps.broadcastToOverlayWindows('subtitle:set', payload),
     onSubtitleChange: (text: string) => deps.onSubtitleChange(text),
+    logSubtitleProcessingDebug: deps.logSubtitleProcessingDebug
+      ? (message: string) => deps.logSubtitleProcessingDebug!(message)
+      : undefined,
     onSubtitleTrackChange: deps.onSubtitleTrackChange
       ? (sid: number | null) => deps.onSubtitleTrackChange!(sid)
       : undefined,

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) => {
@@ -234,3 +229,25 @@ test('subtitle prefetch init clears parsed cues when initialization fails', asyn
 
   assert.deepEqual(cueUpdates, [null]);
 });
+
+test('subtitle prefetch init logs a warning when the source parses to zero cues', async () => {
+  const warnings: string[] = [];
+  const controller = createSubtitlePrefetchInitController({
+    getCurrentService: () => null,
+    setCurrentService: () => {},
+    loadSubtitleSourceText: async () => 'not really subtitles',
+    parseSubtitleCues: (): SubtitleCue[] => [],
+    createSubtitlePrefetchService: () => {
+      throw new Error('should not create a service without cues');
+    },
+    tokenizeSubtitle: async () => null,
+    preCacheTokenization: () => {},
+    logInfo: () => {},
+    logWarn: (message) => warnings.push(message),
+  });
+
+  await controller.initSubtitlePrefetch('/tmp/broken.ass', 0);
+
+  assert.equal(warnings.length, 1);
+  assert.match(warnings[0]!, /\[subtitle-prefetch\].*0 cues.*\/tmp\/broken\.ass/);
+});

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;
@@ -59,6 +58,9 @@ export function createSubtitlePrefetchInitController(
       const cues = deps.parseSubtitleCues(content, sourcePath);
       if (revision !== initRevision || cues.length === 0) {
         if (revision === initRevision) {
+          deps.logWarn(
+            `[subtitle-prefetch] parsed 0 cues from ${sourcePath}; prefetch disabled for this source`,
+          );
           deps.onParsedSubtitleCuesChanged?.(null, null);
         }
         return;
@@ -69,7 +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) {

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

@@ -130,3 +130,121 @@ test('subtitle prefetch runtime does not extract internal subtitle tracks from r
   assert.equal(resolved, null);
   assert.equal(extracted, false);
 });
+
+test('subtitle prefetch refresh logs a warning when source resolution throws', async () => {
+  const warnings: string[] = [];
+  const refresh = createRefreshSubtitlePrefetchFromActiveTrackHandler({
+    getMpvClient: () => ({
+      connected: true,
+      requestProperty: async (name) => (name === 'path' ? '/media/video.mkv' : null),
+    }),
+    getLastObservedTimePos: () => 0,
+    subtitlePrefetchInitController: {
+      cancelPendingInit: () => {},
+      initSubtitlePrefetch: async () => {},
+    },
+    resolveActiveSubtitleSidebarSource: async () => {
+      throw new Error('ffmpeg ENOENT');
+    },
+    logWarn: (message) => warnings.push(message),
+  });
+
+  await refresh();
+
+  assert.equal(warnings.length, 1);
+  assert.match(warnings[0]!, /\[subtitle-prefetch\].*ffmpeg ENOENT/);
+});
+
+test('subtitle prefetch refresh logs debug when mpv client is not connected', async () => {
+  const debugs: string[] = [];
+  const refresh = createRefreshSubtitlePrefetchFromActiveTrackHandler({
+    getMpvClient: () => null,
+    getLastObservedTimePos: () => 0,
+    subtitlePrefetchInitController: {
+      cancelPendingInit: () => {},
+      initSubtitlePrefetch: async () => {},
+    },
+    resolveActiveSubtitleSidebarSource: async () => null,
+    logDebug: (message) => debugs.push(message),
+  });
+
+  await refresh();
+
+  assert.equal(debugs.length, 1);
+  assert.match(debugs[0]!, /\[subtitle-prefetch\].*not connected/);
+});
+
+test('subtitle prefetch refresh logs debug when no subtitle source resolves', async () => {
+  const debugs: string[] = [];
+  const cancels: number[] = [];
+  const refresh = createRefreshSubtitlePrefetchFromActiveTrackHandler({
+    getMpvClient: () => ({
+      connected: true,
+      requestProperty: async (name) => (name === 'path' ? '/media/video.mkv' : null),
+    }),
+    getLastObservedTimePos: () => 0,
+    subtitlePrefetchInitController: {
+      cancelPendingInit: () => {
+        cancels.push(1);
+      },
+      initSubtitlePrefetch: async () => {},
+    },
+    resolveActiveSubtitleSidebarSource: async () => null,
+    logDebug: (message) => debugs.push(message),
+  });
+
+  await refresh();
+
+  assert.deepEqual(cancels, [1]);
+  assert.equal(debugs.length, 1);
+  assert.match(debugs[0]!, /\[subtitle-prefetch\].*no active subtitle source/);
+});
+
+test('subtitle source resolver logs debug when internal track extraction is unavailable', async () => {
+  const debugs: string[] = [];
+  const resolveSource = createResolveActiveSubtitleSidebarSourceHandler({
+    getFfmpegPath: () => 'ffmpeg',
+    extractInternalSubtitleTrack: async () => null,
+    logDebug: (message) => debugs.push(message),
+  });
+
+  const resolved = await resolveSource({
+    currentExternalFilenameRaw: null,
+    currentTrackRaw: {
+      type: 'sub',
+      id: 3,
+      'ff-index': 7,
+      codec: 'hdmv_pgs_subtitle',
+    },
+    trackListRaw: [],
+    sidRaw: 3,
+    videoPath: '/media/video.mkv',
+  });
+
+  assert.equal(resolved, null);
+  assert.equal(debugs.length, 1);
+  assert.match(debugs[0]!, /\[subtitle-prefetch\].*extraction.*hdmv_pgs_subtitle/);
+});
+
+test('subtitle source resolver logs debug when no active subtitle track is selected', async () => {
+  const debugs: string[] = [];
+  const resolveSource = createResolveActiveSubtitleSidebarSourceHandler({
+    getFfmpegPath: () => 'ffmpeg',
+    extractInternalSubtitleTrack: async () => {
+      throw new Error('should not extract without a track');
+    },
+    logDebug: (message) => debugs.push(message),
+  });
+
+  const resolved = await resolveSource({
+    currentExternalFilenameRaw: null,
+    currentTrackRaw: null,
+    trackListRaw: [],
+    sidRaw: null,
+    videoPath: '/media/video.mkv',
+  });
+
+  assert.equal(resolved, null);
+  assert.equal(debugs.length, 1);
+  assert.match(debugs[0]!, /\[subtitle-prefetch\].*no active subtitle track/);
+});

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

@@ -86,6 +86,7 @@ export function createResolveActiveSubtitleSidebarSourceHandler(deps: {
     videoPath: string,
     track: MpvSubtitleTrackLike,
   ) => Promise<{ path: string; cleanup: () => Promise<void> } | null>;
+  logDebug?: (message: string) => void;
 }) {
   return async (input: {
     currentExternalFilenameRaw: unknown;
@@ -104,6 +105,7 @@ export function createResolveActiveSubtitleSidebarSourceHandler(deps: {
 
     const track = getActiveSubtitleTrack(input.currentTrackRaw, input.trackListRaw, input.sidRaw);
     if (!track) {
+      deps.logDebug?.('[subtitle-prefetch] no active subtitle track selected yet');
       return null;
     }
 
@@ -114,6 +116,7 @@ export function createResolveActiveSubtitleSidebarSourceHandler(deps: {
     }
 
     if (isRemoteMediaPath(input.videoPath)) {
+      deps.logDebug?.('[subtitle-prefetch] skipping internal subtitle extraction for remote media');
       return null;
     }
 
@@ -123,6 +126,9 @@ export function createResolveActiveSubtitleSidebarSourceHandler(deps: {
       track,
     );
     if (!extracted) {
+      deps.logDebug?.(
+        `[subtitle-prefetch] internal subtitle extraction unavailable (codec=${String(track.codec ?? 'unknown')}, ff-index=${String(track['ff-index'] ?? 'unknown')})`,
+      );
       return null;
     }
 
@@ -144,10 +150,13 @@ export function createRefreshSubtitlePrefetchFromActiveTrackHandler(deps: {
   resolveActiveSubtitleSidebarSource: (
     input: Parameters<ReturnType<typeof createResolveActiveSubtitleSidebarSourceHandler>>[0],
   ) => Promise<ActiveSubtitleSidebarSource | null>;
+  logDebug?: (message: string) => void;
+  logWarn?: (message: string) => void;
 }) {
   return async (): Promise<void> => {
     const client = deps.getMpvClient();
     if (!client?.connected) {
+      deps.logDebug?.('[subtitle-prefetch] skipped refresh: mpv client not connected');
       return;
     }
 
@@ -162,6 +171,7 @@ export function createRefreshSubtitlePrefetchFromActiveTrackHandler(deps: {
         ]);
       const videoPath = typeof videoPathRaw === 'string' ? videoPathRaw : '';
       if (!videoPath) {
+        deps.logDebug?.('[subtitle-prefetch] skipped refresh: no media path');
         deps.subtitlePrefetchInitController.cancelPendingInit();
         return;
       }
@@ -175,8 +185,14 @@ export function createRefreshSubtitlePrefetchFromActiveTrackHandler(deps: {
       });
       if (!resolvedSource) {
         if (deps.shouldKeepExistingCuesOnMissingSource?.(videoPath) === true) {
+          deps.logDebug?.(
+            '[subtitle-prefetch] no active subtitle source resolved; keeping existing cues',
+          );
           return;
         }
+        deps.logDebug?.(
+          '[subtitle-prefetch] no active subtitle source resolved; cancelling prefetch',
+        );
         deps.subtitlePrefetchInitController.cancelPendingInit();
         return;
       }
@@ -190,8 +206,12 @@ export function createRefreshSubtitlePrefetchFromActiveTrackHandler(deps: {
       } finally {
         await resolvedSource.cleanup?.();
       }
-    } catch {
+    } catch (error) {
-      // Skip refresh when the track query fails.
+      deps.logWarn?.(
+        `[subtitle-prefetch] failed to refresh from active track: ${
+          error instanceof Error ? error.message : String(error)
+        }`,
+      );
     }
   };
 }

src/main/runtime/subtitle-processing-main-deps.ts

@@ -6,6 +6,7 @@ export function createBuildSubtitleProcessingControllerMainDepsHandler(
   return (): SubtitleProcessingControllerDeps => ({
     tokenizeSubtitle: (text: string) => deps.tokenizeSubtitle(text),
     emitSubtitle: (payload) => deps.emitSubtitle(payload),
+    onProcessingSettled: () => deps.onProcessingSettled?.(),
     logDebug: deps.logDebug,
     now: deps.now,
   });

src/main/runtime/subtitle-tokenization-main-deps.ts

@@ -9,6 +9,9 @@ type TokenizerMainDeps = TokenizerDepsRuntimeOptions & {
   getCurrentCharacterDictionaryMediaId?: NonNullable<
     TokenizerDepsRuntimeOptions['getCurrentCharacterDictionaryMediaId']
   >;
+  getCharacterNameCandidates?: NonNullable<
+    TokenizerDepsRuntimeOptions['getCharacterNameCandidates']
+  >;
   getFrequencyDictionaryEnabled: NonNullable<
     TokenizerDepsRuntimeOptions['getFrequencyDictionaryEnabled']
   >;
@@ -84,6 +87,11 @@ export function createBuildTokenizerDepsMainHandler(deps: TokenizerMainDeps) {
           getCurrentCharacterDictionaryMediaId: () => deps.getCurrentCharacterDictionaryMediaId!(),
         }
       : {}),
+    ...(deps.getCharacterNameCandidates
+      ? {
+          getCharacterNameCandidates: () => deps.getCharacterNameCandidates!(),
+        }
+      : {}),
     getFrequencyDictionaryEnabled: () => deps.getFrequencyDictionaryEnabled(),
     getFrequencyDictionaryMatchMode: () => deps.getFrequencyDictionaryMatchMode(),
     getFrequencyRank: (text: string) => deps.getFrequencyRank(text),