perf(tokenizer): skip character-name lookups where no name can start

The greedy name pre-pass asked the Yomitan backend at every Japanese
position, because a character name can begin mid-token. With the character
dictionary enabled that roughly doubled the round trips per line (measured
10 -> 21 on a 23-char line).

SubMiner generates the character dictionary, so the cached snapshots already
list every form a character entry can be matched by (term and reading). Those
forms are installed into the scan runtime once per media and the pre-pass now
probes only positions where one of them starts, compared after kana
normalization so a katakana name still matches a hiragana reading form. The
overhead drops to zero (21 -> 10, the same as with the dictionary disabled).

Fail-safe: with no candidate list (no media id, no cached snapshot, failed
install) the pre-pass keeps its exhaustive behavior, so stale character data
costs speed rather than a missing name. Halfwidth katakana positions bypass
the filter since kana normalization does not fold them.

The candidate lookup is consulted per subtitle line, so it caches its snapshot
directory signature for 5s; dictionary writes still call invalidate().

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 +1,142 @@
+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];
+}
+
+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: `${signature ?? ''}:${normalizedMediaId}`, forms };
+    },
+    invalidate(): void {
+      signature = null;
+      lastSignatureCheckAtMs = 0;
+    },
+  };
+}

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),