feat(stats): add library entry deletion and app-wide delete progress

- Add "Delete Entry" to the anime detail view: removes every episode, session, subtitle line, rollup and cover art for a title plus the vocab counts derived from them, then drops it from the Library grid; refused while that title is currently playing.
- Show delete progress (session, session group, episode, entry) app-wide via a top progress bar + status toast that persist across tab switches, detail views, and the stats overlay window instead of going blank off-tab.
- Fix delete and Vocabulary-tab performance: store seen_ms on word/kanji occurrences so deletes subtract only removed rows via a covering index instead of rescanning each word's full history, and paginate vocab rows before aggregating anime counts.
- Migrate existing databases in place on first launch after upgrade.

src/core/services/__tests__/stats-server.test.ts

@@ -2873,6 +2873,39 @@ Aligned English subtitle
     assert.equal(deleteCalls, 0);
   });
 
+  it('DELETE /api/stats/anime/:animeId deletes the whole library entry', async () => {
+    let deletedAnimeId: number | null = null;
+    const app = createStatsApp(
+      createMockTracker({
+        deleteAnime: async (animeId: number) => {
+          deletedAnimeId = animeId;
+        },
+      } as Partial<ImmersionTrackerService>),
+    );
+
+    const res = await app.request('/api/stats/anime/7', { method: 'DELETE' });
+
+    assert.equal(res.status, 200);
+    assert.equal(deletedAnimeId, 7);
+    assert.deepEqual(await res.json(), { ok: true });
+  });
+
+  it('DELETE /api/stats/anime/:animeId rejects non-positive anime ids', async () => {
+    let deleteCalls = 0;
+    const app = createStatsApp(
+      createMockTracker({
+        deleteAnime: async () => {
+          deleteCalls += 1;
+        },
+      } as Partial<ImmersionTrackerService>),
+    );
+
+    const res = await app.request('/api/stats/anime/0', { method: 'DELETE' });
+
+    assert.equal(res.status, 400);
+    assert.equal(deleteCalls, 0);
+  });
+
   it('POST /api/stats/anki/browse returns 400 for missing noteId', async () => {
     const app = createStatsApp(createMockTracker());
     const res = await app.request('/api/stats/anki/browse', { method: 'POST' });

src/core/services/immersion-tracker-service.test.ts

@@ -1466,6 +1466,97 @@ test('deleteVideo ignores the currently active video and keeps new writes flusha
   }
 });
 
+test('deleteAnime removes every episode, session and library row for the title', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+
+  try {
+    const Ctor = await loadTrackerCtor();
+
+    for (const episode of ['S02E05', 'S02E06']) {
+      tracker = new Ctor({ dbPath });
+      tracker.handleMediaChange(`/tmp/Little Witch Academia ${episode}.mkv`, `Episode ${episode}`);
+      await waitForPendingAnimeMetadata(tracker);
+      tracker.recordSubtitleLine('今日は晴れです', 0, 1.2);
+      tracker.recordCardsMined(1);
+      tracker.destroy();
+      tracker = null;
+    }
+
+    tracker = new Ctor({ dbPath });
+    const privateApi = tracker as unknown as { db: DatabaseSync };
+    const animeId = (
+      privateApi.db.prepare('SELECT anime_id FROM imm_anime LIMIT 1').get() as {
+        anime_id: number;
+      } | null
+    )?.anime_id;
+    assert.ok(animeId);
+
+    const libraryBefore = await tracker.getAnimeLibrary();
+    assert.equal(libraryBefore.length, 1);
+    assert.equal(libraryBefore[0]?.episodeCount, 2);
+
+    await tracker.deleteAnime(animeId);
+
+    const libraryAfter = await tracker.getAnimeLibrary();
+    assert.equal(libraryAfter.length, 0);
+
+    const countOf = (sql: string): number =>
+      (privateApi.db.prepare(sql).get() as { total: number }).total;
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_anime'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_lifetime_anime'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_videos'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_sessions'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_subtitle_lines'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_daily_rollups'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_lifetime_media'), 0);
+    assert.equal(countOf('SELECT COUNT(*) AS total FROM imm_words'), 0);
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('deleteAnime ignores the title of the currently active session', async () => {
+  const dbPath = makeDbPath();
+  let tracker: ImmersionTrackerService | null = null;
+
+  try {
+    const Ctor = await loadTrackerCtor();
+    tracker = new Ctor({ dbPath });
+    tracker.handleMediaChange('/tmp/Little Witch Academia S02E05.mkv', 'Episode 5');
+    await waitForPendingAnimeMetadata(tracker);
+
+    const privateApi = tracker as unknown as {
+      db: DatabaseSync;
+      sessionState: { sessionId: number; videoId: number } | null;
+    };
+    const videoId = privateApi.sessionState?.videoId;
+    assert.ok(videoId);
+    const animeId = (
+      privateApi.db.prepare('SELECT anime_id FROM imm_videos WHERE video_id = ?').get(videoId) as {
+        anime_id: number | null;
+      } | null
+    )?.anime_id;
+    assert.ok(animeId);
+
+    await tracker.deleteAnime(animeId);
+
+    const animeCountRow = privateApi.db
+      .prepare('SELECT COUNT(*) AS total FROM imm_anime WHERE anime_id = ?')
+      .get(animeId) as { total: number };
+    const videoCountRow = privateApi.db
+      .prepare('SELECT COUNT(*) AS total FROM imm_videos WHERE video_id = ?')
+      .get(videoId) as { total: number };
+
+    assert.equal(animeCountRow.total, 1);
+    assert.equal(videoCountRow.total, 1);
+  } finally {
+    tracker?.destroy();
+    cleanupDbPath(dbPath);
+  }
+});
+
 test('handleMediaChange links parsed anime metadata on the active video row', async () => {
   const dbPath = makeDbPath();
   let tracker: ImmersionTrackerService | null = null;

src/core/services/immersion-tracker-service.ts

@@ -83,6 +83,7 @@ import {
 } from './immersion-tracker/query-library';
 import {
   cleanupVocabularyStats,
+  deleteAnime as deleteAnimeQuery,
   deleteSession as deleteSessionQuery,
   deleteSessions as deleteSessionsQuery,
   deleteVideo as deleteVideoQuery,
@@ -733,6 +734,20 @@ export class ImmersionTrackerService {
     deleteVideoQuery(this.db, videoId);
   }
 
+  async deleteAnime(animeId: number): Promise<void> {
+    const activeVideoId = this.sessionState?.videoId;
+    if (activeVideoId !== undefined) {
+      const activeAnime = this.db
+        .prepare('SELECT anime_id FROM imm_videos WHERE video_id = ?')
+        .get(activeVideoId) as { anime_id: number | null } | null;
+      if (activeAnime?.anime_id === animeId) {
+        this.logger.warn(`Ignoring delete request for active immersion anime ${animeId}`);
+        return;
+      }
+    }
+    deleteAnimeQuery(this.db, animeId);
+  }
+
   async reassignAnimeAnilist(
     animeId: number,
     info: {

src/core/services/immersion-tracker/__tests__/lexical-removal.test.ts

@@ -0,0 +1,331 @@
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import test from 'node:test';
+import { Database } from '../sqlite.js';
+import type { DatabaseSync } from '../sqlite.js';
+import { ensureSchema } from '../storage.js';
+import { deleteSession, deleteSessions, deleteVideo } from '../query-maintenance.js';
+
+const DAY_MS = 86_400_000;
+const BASE_MS = 1_700_000_000_000;
+
+function makeDbPath(): string {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-lexical-removal-test-'));
+  return path.join(dir, 'immersion.sqlite');
+}
+
+function cleanupDbPath(dbPath: string): void {
+  const dir = path.dirname(dbPath);
+  if (!fs.existsSync(dir)) return;
+  fs.rmSync(dir, { recursive: true, force: true });
+}
+
+/**
+ * Seed two episodes of one anime, each with one ended session.
+ *
+ * `lines` places a word occurrence on a specific day so tests can control which
+ * session holds a word's first/last occurrence.
+ */
+function seed(
+  db: DatabaseSync,
+  lines: Array<{ session: 1 | 2; wordId: number; dayOffset: number; count?: number }>,
+  options: { legacyRows?: boolean } = {},
+): void {
+  db.exec(`
+    INSERT INTO imm_anime(anime_id, normalized_title_key, canonical_title, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'show', 'Show', ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_videos(video_id, video_key, anime_id, canonical_title, source_type, watched, duration_ms, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'v1', 1, 'Ep 1', 1, 1, 1440000, ${BASE_MS}, ${BASE_MS}),
+             (2, 'v2', 1, 'Ep 2', 1, 1, 1440000, ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_sessions(session_id, session_uuid, video_id, started_at_ms, ended_at_ms, status, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 's1', 1, '${BASE_MS}', '${BASE_MS + 1000}', 2, ${BASE_MS}, ${BASE_MS}),
+             (2, 's2', 2, '${BASE_MS + DAY_MS}', '${BASE_MS + DAY_MS + 1000}', 2, ${BASE_MS}, ${BASE_MS});
+  `);
+
+  const insertLine = db.prepare(
+    `INSERT INTO imm_subtitle_lines(line_id, session_id, video_id, anime_id, line_index, text, CREATED_DATE, LAST_UPDATE_DATE)
+     VALUES (?, ?, ?, 1, ?, ?, ?, ?)`,
+  );
+  const insertWord = db.prepare(
+    `INSERT OR IGNORE INTO imm_words(id, headword, word, reading, part_of_speech, pos1, first_seen, last_seen, frequency)
+     VALUES (?, ?, ?, '', 'noun', '名詞', 0, 0, 0)`,
+  );
+  // `legacyRows` reproduces databases written before the seen_ms column existed,
+  // where the timestamp has to be read back off the subtitle line.
+  const insertOccurrence = options.legacyRows
+    ? db.prepare(
+        `INSERT INTO imm_word_line_occurrences(line_id, word_id, occurrence_count) VALUES (?, ?, ?)`,
+      )
+    : db.prepare(
+        `INSERT INTO imm_word_line_occurrences(line_id, word_id, occurrence_count, seen_ms) VALUES (?, ?, ?, ?)`,
+      );
+
+  let lineId = 0;
+  for (const line of lines) {
+    lineId += 1;
+    const seenMs = BASE_MS + line.dayOffset * DAY_MS;
+    insertLine.run(lineId, line.session, line.session, lineId, `line ${lineId}`, seenMs, seenMs);
+    insertWord.run(line.wordId, `語${line.wordId}`, `語${line.wordId}`);
+    if (options.legacyRows) {
+      insertOccurrence.run(lineId, line.wordId, line.count ?? 1);
+    } else {
+      insertOccurrence.run(lineId, line.wordId, line.count ?? 1, seenMs);
+    }
+  }
+
+  // Match what the tracker maintains: aggregates derived from the occurrences.
+  db.exec(`
+    UPDATE imm_words SET
+      frequency = (
+        SELECT COALESCE(SUM(o.occurrence_count), 0)
+        FROM imm_word_line_occurrences o WHERE o.word_id = imm_words.id
+      ),
+      first_seen = (
+        SELECT MIN(sl.CREATED_DATE) / 1000
+        FROM imm_word_line_occurrences o
+        JOIN imm_subtitle_lines sl ON sl.line_id = o.line_id
+        WHERE o.word_id = imm_words.id
+      ),
+      last_seen = (
+        SELECT MAX(sl.LAST_UPDATE_DATE) / 1000
+        FROM imm_word_line_occurrences o
+        JOIN imm_subtitle_lines sl ON sl.line_id = o.line_id
+        WHERE o.word_id = imm_words.id
+      )
+  `);
+}
+
+function createDb(
+  lines: Parameters<typeof seed>[1],
+  options: Parameters<typeof seed>[2] = {},
+): { db: DatabaseSync; dbPath: string } {
+  const dbPath = makeDbPath();
+  const db = new Database(dbPath);
+  ensureSchema(db);
+  seed(db, lines, options);
+  return { db, dbPath };
+}
+
+function readWord(
+  db: DatabaseSync,
+  wordId: number,
+): { frequency: number; firstSeen: number; lastSeen: number } | null {
+  return (
+    (db
+      .prepare(
+        'SELECT frequency, first_seen AS firstSeen, last_seen AS lastSeen FROM imm_words WHERE id = ?',
+      )
+      .get(wordId) as { frequency: number; firstSeen: number; lastSeen: number } | null) ?? null
+  );
+}
+
+test('deleting a session subtracts only the occurrences it removed', () => {
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 10, dayOffset: 0, count: 3 },
+    { session: 2, wordId: 10, dayOffset: 1, count: 4 },
+  ]);
+
+  try {
+    assert.equal(readWord(db, 10)?.frequency, 7);
+
+    deleteSession(db, 1);
+
+    assert.equal(readWord(db, 10)?.frequency, 4, 'only session 1 occurrences are subtracted');
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('deleting the session that held a word first drops the word entirely', () => {
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 11, dayOffset: 0 },
+    { session: 2, wordId: 12, dayOffset: 1 },
+  ]);
+
+  try {
+    deleteSession(db, 1);
+
+    assert.equal(readWord(db, 11), null, 'word seen only in the deleted session is removed');
+    assert.ok(readWord(db, 12), 'word seen elsewhere survives');
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('deleting the earliest session moves first_seen forward to the surviving line', () => {
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 13, dayOffset: 0 },
+    { session: 2, wordId: 13, dayOffset: 5 },
+  ]);
+
+  try {
+    assert.equal(readWord(db, 13)?.firstSeen, Math.floor(BASE_MS / 1000));
+
+    deleteSession(db, 1);
+
+    const word = readWord(db, 13);
+    assert.equal(word?.frequency, 1);
+    assert.equal(
+      word?.firstSeen,
+      Math.floor((BASE_MS + 5 * DAY_MS) / 1000),
+      'first_seen advances to the remaining occurrence',
+    );
+    assert.equal(word?.lastSeen, Math.floor((BASE_MS + 5 * DAY_MS) / 1000));
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('deleting the latest session moves last_seen back to the surviving line', () => {
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 14, dayOffset: 0 },
+    { session: 2, wordId: 14, dayOffset: 5 },
+  ]);
+
+  try {
+    deleteSession(db, 2);
+
+    const word = readWord(db, 14);
+    assert.equal(word?.frequency, 1);
+    assert.equal(word?.lastSeen, Math.floor(BASE_MS / 1000), 'last_seen falls back to session 1');
+    assert.equal(word?.firstSeen, Math.floor(BASE_MS / 1000));
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('deleting an interior occurrence leaves the surrounding extremes untouched', () => {
+  // Session 2 carries the middle occurrence; sessions bracket it in time.
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 15, dayOffset: 0 },
+    { session: 2, wordId: 15, dayOffset: 3 },
+    { session: 1, wordId: 15, dayOffset: 9 },
+  ]);
+
+  try {
+    deleteSessions(db, [2]);
+
+    const word = readWord(db, 15);
+    assert.equal(word?.frequency, 2);
+    assert.equal(word?.firstSeen, Math.floor(BASE_MS / 1000));
+    assert.equal(word?.lastSeen, Math.floor((BASE_MS + 9 * DAY_MS) / 1000));
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('a stored frequency that has drifted low is repaired instead of dropping a live word', () => {
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 16, dayOffset: 0, count: 5 },
+    { session: 2, wordId: 16, dayOffset: 4, count: 5 },
+  ]);
+
+  try {
+    // Simulate drift: the stored total is lower than the occurrences justify, so
+    // naive subtraction would take the word to zero while rows still reference it.
+    db.prepare('UPDATE imm_words SET frequency = 5 WHERE id = ?').run(16);
+
+    deleteSession(db, 1);
+
+    const word = readWord(db, 16);
+    assert.ok(word, 'word with surviving occurrences is not deleted');
+    assert.equal(word?.frequency, 5, 'frequency is recomputed from the surviving occurrences');
+    assert.equal(word?.firstSeen, Math.floor((BASE_MS + 4 * DAY_MS) / 1000));
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('deleting a video subtracts every occurrence carried by its lines', () => {
+  const { db, dbPath } = createDb([
+    { session: 1, wordId: 17, dayOffset: 0, count: 2 },
+    { session: 1, wordId: 17, dayOffset: 1, count: 3 },
+    { session: 2, wordId: 17, dayOffset: 2, count: 4 },
+    { session: 2, wordId: 18, dayOffset: 2, count: 1 },
+  ]);
+
+  try {
+    deleteVideo(db, 1);
+
+    assert.equal(readWord(db, 17)?.frequency, 4, 'both lines from video 1 are subtracted');
+    assert.equal(readWord(db, 18)?.frequency, 1, 'untouched video keeps its counts');
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('occurrence rows written before the seen_ms column still resolve their dates', () => {
+  const { db, dbPath } = createDb(
+    [
+      { session: 1, wordId: 20, dayOffset: 0 },
+      { session: 2, wordId: 20, dayOffset: 6 },
+    ],
+    { legacyRows: true },
+  );
+
+  try {
+    assert.equal(
+      (
+        db
+          .prepare('SELECT COUNT(*) AS total FROM imm_word_line_occurrences WHERE seen_ms IS NULL')
+          .get() as { total: number }
+      ).total,
+      2,
+      'precondition: rows carry no denormalised timestamp',
+    );
+
+    deleteSession(db, 1);
+
+    const word = readWord(db, 20);
+    assert.equal(word?.frequency, 1);
+    assert.equal(
+      word?.firstSeen,
+      Math.floor((BASE_MS + 6 * DAY_MS) / 1000),
+      'falls back to the subtitle line timestamp',
+    );
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('upgrading an older database backfills seen_ms from the subtitle lines', () => {
+  const { db, dbPath } = createDb(
+    [
+      { session: 1, wordId: 21, dayOffset: 0 },
+      { session: 2, wordId: 21, dayOffset: 2 },
+    ],
+    { legacyRows: true },
+  );
+
+  try {
+    // Re-run ensureSchema the way a pre-19 database would be opened.
+    db.exec('DELETE FROM imm_schema_version');
+    db.exec(`INSERT INTO imm_schema_version(schema_version, applied_at_ms) VALUES (18, '0')`);
+    ensureSchema(db);
+
+    const rows = db
+      .prepare('SELECT line_id AS lineId, seen_ms AS seenMs FROM imm_word_line_occurrences')
+      .all() as Array<{ lineId: number; seenMs: number | null }>;
+    assert.equal(rows.length, 2);
+    for (const row of rows) {
+      assert.ok(row.seenMs, `line ${row.lineId} should have a backfilled timestamp`);
+    }
+    assert.deepEqual(
+      rows.map((row) => row.seenMs).sort((a, b) => Number(a) - Number(b)),
+      [BASE_MS, BASE_MS + 2 * DAY_MS],
+    );
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});

src/core/services/immersion-tracker/__tests__/query.test.ts

@@ -4982,3 +4982,145 @@ test('getTrendsDashboard librarySummary is empty when no rollups exist', () => {
     cleanupDbPath(dbPath);
   }
 });
+
+test('getVocabularyStats counts the distinct anime each word appeared in', () => {
+  const dbPath = makeDbPath();
+  const db = new Database(dbPath);
+
+  try {
+    ensureSchema(db);
+    const startedAtMs = 1_700_000_000_000;
+    db.exec(`
+      INSERT INTO imm_anime(anime_id, normalized_title_key, canonical_title, CREATED_DATE, LAST_UPDATE_DATE)
+        VALUES (11, 'vocab-a', 'Vocab A', ${startedAtMs}, ${startedAtMs}),
+               (22, 'vocab-b', 'Vocab B', ${startedAtMs}, ${startedAtMs});
+    `);
+    const videoIds = [1, 2].map((n) =>
+      getOrCreateVideoRecord(db, `local:/tmp/vocab-anime-${n}.mkv`, {
+        canonicalTitle: `Vocab Anime ${n}`,
+        sourcePath: `/tmp/vocab-anime-${n}.mkv`,
+        sourceUrl: null,
+        sourceType: SOURCE_TYPE_LOCAL,
+      }),
+    );
+    const sessionIds = videoIds.map(
+      (videoId) => startSessionRecord(db, videoId, startedAtMs).sessionId,
+    );
+
+    // 猫 spans both titles, twice in the first, so the count has to deduplicate
+    // by anime rather than just tallying occurrence rows. 犬 stays in one title.
+    insertFilteredWordOccurrence(db, {
+      sessionId: sessionIds[0]!,
+      videoId: videoIds[0]!,
+      animeId: 11,
+      lineIndex: 1,
+      occurrenceCount: 5,
+      startedAtMs,
+      word: '猫',
+      reading: 'ねこ',
+    });
+    insertFilteredWordOccurrence(db, {
+      sessionId: sessionIds[0]!,
+      videoId: videoIds[0]!,
+      animeId: 11,
+      lineIndex: 2,
+      occurrenceCount: 1,
+      startedAtMs,
+      word: '猫',
+      reading: 'ねこ',
+    });
+    insertFilteredWordOccurrence(db, {
+      sessionId: sessionIds[1]!,
+      videoId: videoIds[1]!,
+      animeId: 22,
+      lineIndex: 3,
+      occurrenceCount: 4,
+      startedAtMs,
+      word: '猫',
+      reading: 'ねこ',
+    });
+    insertFilteredWordOccurrence(db, {
+      sessionId: sessionIds[0]!,
+      videoId: videoIds[0]!,
+      animeId: 11,
+      lineIndex: 4,
+      occurrenceCount: 2,
+      startedAtMs,
+      word: '犬',
+      reading: 'いぬ',
+    });
+
+    const rows = getVocabularyStats(db, 10);
+    const cat = rows.find((row) => row.headword === '猫');
+    const dog = rows.find((row) => row.headword === '犬');
+
+    assert.equal(cat?.animeCount, 2, '猫 was seen in two anime across three lines');
+    assert.equal(dog?.animeCount, 1, '犬 was seen in one');
+    assert.equal(cat?.frequency, 10, 'frequency still aggregates across both titles');
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});
+
+test('getVocabularyStats still applies part-of-speech exclusions', () => {
+  const dbPath = makeDbPath();
+  const db = new Database(dbPath);
+
+  try {
+    ensureSchema(db);
+    const startedAtMs = 1_700_000_000_000;
+    db.exec(`
+      INSERT INTO imm_anime(anime_id, normalized_title_key, canonical_title, CREATED_DATE, LAST_UPDATE_DATE)
+        VALUES (11, 'vocab-a', 'Vocab A', ${startedAtMs}, ${startedAtMs});
+    `);
+    const videoId = getOrCreateVideoRecord(db, 'local:/tmp/vocab-exclude.mkv', {
+      canonicalTitle: 'Vocab Exclude',
+      sourcePath: '/tmp/vocab-exclude.mkv',
+      sourceUrl: null,
+      sourceType: SOURCE_TYPE_LOCAL,
+    });
+    const { sessionId } = startSessionRecord(db, videoId, startedAtMs);
+
+    insertFilteredWordOccurrence(db, {
+      sessionId,
+      videoId,
+      animeId: 11,
+      lineIndex: 1,
+      occurrenceCount: 9,
+      startedAtMs,
+      word: '走る',
+      reading: 'はしる',
+      partOfSpeech: 'verb',
+    });
+    insertFilteredWordOccurrence(db, {
+      sessionId,
+      videoId,
+      animeId: 11,
+      lineIndex: 2,
+      occurrenceCount: 3,
+      startedAtMs,
+      word: '猫',
+      reading: 'ねこ',
+      partOfSpeech: 'noun',
+    });
+
+    const all = getVocabularyStats(db, 10);
+    assert.deepEqual(
+      all.map((row) => row.headword).sort(),
+      ['猫', '走る'],
+      'precondition: both parts of speech are present',
+    );
+
+    const nounsOnly = getVocabularyStats(db, 10, ['verb']);
+    assert.deepEqual(
+      nounsOnly.map((row) => row.headword),
+      ['猫'],
+      'excluded part of speech is filtered out even though it ranks higher',
+    );
+    assert.equal(nounsOnly[0]?.animeCount, 1, 'surviving rows keep their anime count');
+  } finally {
+    db.close();
+    cleanupDbPath(dbPath);
+  }
+});

src/core/services/immersion-tracker/query-lexical.ts

@@ -115,18 +115,29 @@ export function getVocabularyStats(
   const whereClause = hasExclude
     ? `WHERE (part_of_speech IS NULL OR part_of_speech NOT IN (${placeholders}))`
     : '';
+  // The page is selected before the join so `animeCount` is only computed for the
+  // rows being returned. Aggregating first made every request walk each word's
+  // entire occurrence history — seconds of blocked event loop on a large library,
+  // because only the ordering, not the aggregate, decides which rows survive.
   const stmt = db.prepare(`
-    SELECT w.id AS wordId, w.headword, w.word, w.reading,
-      w.part_of_speech AS partOfSpeech, w.pos1, w.pos2, w.pos3,
-      w.frequency, w.frequency_rank AS frequencyRank,
-      w.first_seen AS firstSeen, w.last_seen AS lastSeen,
+    WITH page AS (
+      SELECT id, headword, word, reading, part_of_speech, pos1, pos2, pos3,
+             frequency, frequency_rank, first_seen, last_seen
+      FROM imm_words
+      ${whereClause}
+      ORDER BY frequency DESC
+      LIMIT ? OFFSET ?
+    )
+    SELECT p.id AS wordId, p.headword, p.word, p.reading,
+      p.part_of_speech AS partOfSpeech, p.pos1, p.pos2, p.pos3,
+      p.frequency, p.frequency_rank AS frequencyRank,
+      p.first_seen AS firstSeen, p.last_seen AS lastSeen,
       COUNT(DISTINCT sl.anime_id) AS animeCount
-    FROM imm_words w
-    LEFT JOIN imm_word_line_occurrences o ON o.word_id = w.id
+    FROM page p
+    LEFT JOIN imm_word_line_occurrences o ON o.word_id = p.id
     LEFT JOIN imm_subtitle_lines sl ON sl.line_id = o.line_id AND sl.anime_id IS NOT NULL
-    ${whereClause ? whereClause.replace('part_of_speech', 'w.part_of_speech') : ''}
-    GROUP BY w.id
-    ORDER BY w.frequency DESC LIMIT ? OFFSET ?
+    GROUP BY p.id
+    ORDER BY p.frequency DESC
   `);
   const visibleRows: VocabularyStatsRow[] = [];
   let offset = 0;

src/core/services/immersion-tracker/query-maintenance.ts

@@ -9,14 +9,12 @@ import { PartOfSpeech, type MergedToken } from '../../../types';
 import { shouldExcludeTokenFromVocabularyPersistence } from '../tokenizer/annotation-stage';
 import { deriveStoredPartOfSpeech } from '../tokenizer/part-of-speech';
 import {
+  applyLexicalRemovals,
   cleanupUnusedCoverArtBlobHash,
   deleteSessionsByIds,
   findSharedCoverBlobHash,
-  getAffectedKanjiIdsForSessions,
-  getAffectedKanjiIdsForVideo,
-  getAffectedWordIdsForSessions,
-  getAffectedWordIdsForVideo,
-  refreshLexicalAggregates,
+  planLexicalRemovalsForSessions,
+  planLexicalRemovalsForVideos,
   toDbMs,
   toDbTimestamp,
 } from './query-shared';
@@ -232,12 +230,13 @@ export async function cleanupVocabularyStats(
      WHERE id = ?`,
   );
   const moveOccurrencesStmt = db.prepare(
-    `INSERT INTO imm_word_line_occurrences (line_id, word_id, occurrence_count)
-     SELECT line_id, ?, occurrence_count
+    `INSERT INTO imm_word_line_occurrences (line_id, word_id, occurrence_count, seen_ms)
+     SELECT line_id, ?, occurrence_count, seen_ms
      FROM imm_word_line_occurrences
      WHERE word_id = ?
      ON CONFLICT(line_id, word_id) DO UPDATE SET
-       occurrence_count = imm_word_line_occurrences.occurrence_count + excluded.occurrence_count`,
+       occurrence_count = imm_word_line_occurrences.occurrence_count + excluded.occurrence_count,
+       seen_ms = COALESCE(imm_word_line_occurrences.seen_ms, excluded.seen_ms)`,
   );
   const deleteOccurrencesStmt = db.prepare(
     'DELETE FROM imm_word_line_occurrences WHERE word_id = ?',
@@ -484,14 +483,13 @@ export function isVideoWatched(db: DatabaseSync, videoId: number): boolean {
 
 export function deleteSession(db: DatabaseSync, sessionId: number): void {
   const sessionIds = [sessionId];
-  const affectedWordIds = getAffectedWordIdsForSessions(db, sessionIds);
-  const affectedKanjiIds = getAffectedKanjiIdsForSessions(db, sessionIds);
+  const lexicalRemovals = planLexicalRemovalsForSessions(db, sessionIds);
   const affectedRollupGroups = getRollupGroupsForSessions(db, sessionIds);
 
   db.exec('BEGIN IMMEDIATE');
   try {
     deleteSessionsByIds(db, sessionIds);
-    refreshLexicalAggregates(db, affectedWordIds, affectedKanjiIds);
+    applyLexicalRemovals(db, lexicalRemovals);
     rebuildLifetimeSummariesInTransaction(db);
     refreshRollupsForGroupsInTransaction(db, affectedRollupGroups);
     db.exec('COMMIT');
@@ -503,14 +501,13 @@ export function deleteSession(db: DatabaseSync, sessionId: number): void {
 
 export function deleteSessions(db: DatabaseSync, sessionIds: number[]): void {
   if (sessionIds.length === 0) return;
-  const affectedWordIds = getAffectedWordIdsForSessions(db, sessionIds);
-  const affectedKanjiIds = getAffectedKanjiIdsForSessions(db, sessionIds);
+  const lexicalRemovals = planLexicalRemovalsForSessions(db, sessionIds);
   const affectedRollupGroups = getRollupGroupsForSessions(db, sessionIds);
 
   db.exec('BEGIN IMMEDIATE');
   try {
     deleteSessionsByIds(db, sessionIds);
-    refreshLexicalAggregates(db, affectedWordIds, affectedKanjiIds);
+    applyLexicalRemovals(db, lexicalRemovals);
     rebuildLifetimeSummariesInTransaction(db);
     refreshRollupsForGroupsInTransaction(db, affectedRollupGroups);
     db.exec('COMMIT');
@@ -520,6 +517,66 @@ export function deleteSessions(db: DatabaseSync, sessionIds: number[]): void {
   }
 }
 
+/**
+ * Delete an entire library entry: every episode of the anime, all of their
+ * sessions and derived stats, and the anime row itself.
+ *
+ * Mirrors {@link deleteVideo} per episode, but batches the lexical refresh and
+ * lifetime rebuild into a single transaction so a multi-episode title doesn't
+ * pay for one full rebuild per episode.
+ */
+export function deleteAnime(db: DatabaseSync, animeId: number): void {
+  const videoIds = (
+    db.prepare('SELECT video_id FROM imm_videos WHERE anime_id = ?').all(animeId) as Array<{
+      video_id: number;
+    }>
+  ).map((row) => row.video_id);
+
+  const lexicalRemovals = planLexicalRemovalsForVideos(db, videoIds);
+  const coverBlobHashes: string[] = [];
+  const sessionIds: number[] = [];
+  for (const videoId of videoIds) {
+    const artRow = db
+      .prepare('SELECT cover_blob_hash AS coverBlobHash FROM imm_media_art WHERE video_id = ?')
+      .get(videoId) as { coverBlobHash: string | null } | undefined;
+    if (artRow?.coverBlobHash) {
+      coverBlobHashes.push(artRow.coverBlobHash);
+    }
+    const sessions = db
+      .prepare('SELECT session_id FROM imm_sessions WHERE video_id = ?')
+      .all(videoId) as Array<{ session_id: number }>;
+    sessionIds.push(...sessions.map((session) => session.session_id));
+  }
+
+  db.exec('BEGIN IMMEDIATE');
+  try {
+    deleteSessionsByIds(db, sessionIds);
+    const deleteLinesStmt = db.prepare('DELETE FROM imm_subtitle_lines WHERE video_id = ?');
+    const deleteDailyStmt = db.prepare('DELETE FROM imm_daily_rollups WHERE video_id = ?');
+    const deleteMonthlyStmt = db.prepare('DELETE FROM imm_monthly_rollups WHERE video_id = ?');
+    const deleteArtStmt = db.prepare('DELETE FROM imm_media_art WHERE video_id = ?');
+    const deleteVideoStmt = db.prepare('DELETE FROM imm_videos WHERE video_id = ?');
+    for (const videoId of videoIds) {
+      deleteLinesStmt.run(videoId);
+      deleteDailyStmt.run(videoId);
+      deleteMonthlyStmt.run(videoId);
+      deleteArtStmt.run(videoId);
+      deleteVideoStmt.run(videoId);
+    }
+    for (const coverBlobHash of new Set(coverBlobHashes)) {
+      cleanupUnusedCoverArtBlobHash(db, coverBlobHash);
+    }
+    db.prepare('DELETE FROM imm_lifetime_anime WHERE anime_id = ?').run(animeId);
+    db.prepare('DELETE FROM imm_anime WHERE anime_id = ?').run(animeId);
+    applyLexicalRemovals(db, lexicalRemovals);
+    rebuildLifetimeSummariesInTransaction(db);
+    db.exec('COMMIT');
+  } catch (error) {
+    db.exec('ROLLBACK');
+    throw error;
+  }
+}
+
 export function deleteVideo(db: DatabaseSync, videoId: number): void {
   const artRow = db
     .prepare(
@@ -530,8 +587,7 @@ export function deleteVideo(db: DatabaseSync, videoId: number): void {
       `,
     )
     .get(videoId) as { coverBlobHash: string | null } | undefined;
-  const affectedWordIds = getAffectedWordIdsForVideo(db, videoId);
-  const affectedKanjiIds = getAffectedKanjiIdsForVideo(db, videoId);
+  const lexicalRemovals = planLexicalRemovalsForVideos(db, [videoId]);
   const sessions = db
     .prepare('SELECT session_id FROM imm_sessions WHERE video_id = ?')
     .all(videoId) as Array<{ session_id: number }>;
@@ -548,7 +604,7 @@ export function deleteVideo(db: DatabaseSync, videoId: number): void {
     db.prepare('DELETE FROM imm_media_art WHERE video_id = ?').run(videoId);
     cleanupUnusedCoverArtBlobHash(db, artRow?.coverBlobHash ?? null);
     db.prepare('DELETE FROM imm_videos WHERE video_id = ?').run(videoId);
-    refreshLexicalAggregates(db, affectedWordIds, affectedKanjiIds);
+    applyLexicalRemovals(db, lexicalRemovals);
     rebuildLifetimeSummariesInTransaction(db);
     db.exec('COMMIT');
   } catch (error) {

src/core/services/immersion-tracker/query-shared.ts

@@ -203,6 +203,167 @@ export function getAffectedKanjiIdsForVideo(db: DatabaseSync, videoId: number):
   return getAffectedIdsForVideo(db, 'kanji', videoId);
 }
 
+/** Per-entity totals that a pending delete is about to remove. */
+interface LexicalRemoval {
+  id: number;
+  removedFrequency: number;
+  removedFirstSeenMs: number | null;
+  removedLastSeenMs: number | null;
+}
+
+/** What a pending delete removes from `imm_words` and `imm_kanji`. */
+export interface LexicalRemovalPlan {
+  words: LexicalRemoval[];
+  kanji: LexicalRemoval[];
+}
+
+export const EMPTY_LEXICAL_REMOVAL_PLAN: LexicalRemovalPlan = { words: [], kanji: [] };
+
+function collectLexicalRemovals(
+  db: DatabaseSync,
+  entity: LexicalEntity,
+  lineScopeSql: string,
+  params: number[],
+): LexicalRemoval[] {
+  const table = entity === 'word' ? 'imm_word_line_occurrences' : 'imm_kanji_line_occurrences';
+  const col = `${entity}_id`;
+  return db
+    .prepare(
+      `SELECT
+         o.${col} AS id,
+         COALESCE(SUM(o.occurrence_count), 0) AS removedFrequency,
+         MIN(COALESCE(o.seen_ms, sl.CREATED_DATE, sl.LAST_UPDATE_DATE)) AS removedFirstSeenMs,
+         MAX(COALESCE(o.seen_ms, sl.LAST_UPDATE_DATE, sl.CREATED_DATE)) AS removedLastSeenMs
+       FROM imm_subtitle_lines sl
+       JOIN ${table} o ON o.line_id = sl.line_id
+       WHERE ${lineScopeSql}
+       GROUP BY o.${col}`,
+    )
+    .all(...params) as LexicalRemoval[];
+}
+
+function planLexicalRemovals(
+  db: DatabaseSync,
+  lineScopeSql: string,
+  params: number[],
+): LexicalRemovalPlan {
+  return {
+    words: collectLexicalRemovals(db, 'word', lineScopeSql, params),
+    kanji: collectLexicalRemovals(db, 'kanji', lineScopeSql, params),
+  };
+}
+
+/**
+ * Measure what deleting these sessions removes from the vocabulary tables.
+ *
+ * Must run before the rows are deleted. Reads only the subtitle lines in scope,
+ * unlike {@link refreshLexicalAggregates}, which re-reads every occurrence of
+ * every affected word across the whole library.
+ */
+export function planLexicalRemovalsForSessions(
+  db: DatabaseSync,
+  sessionIds: number[],
+): LexicalRemovalPlan {
+  if (sessionIds.length === 0) return EMPTY_LEXICAL_REMOVAL_PLAN;
+  return planLexicalRemovals(db, `sl.session_id IN (${makePlaceholders(sessionIds)})`, sessionIds);
+}
+
+/** Measure what deleting these videos removes from the vocabulary tables. */
+export function planLexicalRemovalsForVideos(
+  db: DatabaseSync,
+  videoIds: number[],
+): LexicalRemovalPlan {
+  if (videoIds.length === 0) return EMPTY_LEXICAL_REMOVAL_PLAN;
+  return planLexicalRemovals(db, `sl.video_id IN (${makePlaceholders(videoIds)})`, videoIds);
+}
+
+function toStoredSeenSeconds(ms: number | null): number | null {
+  if (ms === null || !Number.isFinite(ms)) return null;
+  return Math.floor(ms / 1000);
+}
+
+/**
+ * Apply a removal plan to the vocabulary aggregates.
+ *
+ * Frequencies are adjusted by subtraction, which is exact and touches only the
+ * affected rows. `first_seen`/`last_seen` only need a rescan when the removed
+ * lines held the current extreme, and rows whose frequency reaches zero are
+ * verified against the surviving occurrences before deletion — so stored counts
+ * that have drifted still converge on the truth instead of dropping a live row.
+ */
+export function applyLexicalRemovals(db: DatabaseSync, plan: LexicalRemovalPlan): void {
+  applyRemovalsForEntity(db, 'word', plan.words);
+  applyRemovalsForEntity(db, 'kanji', plan.kanji);
+}
+
+function applyRemovalsForEntity(
+  db: DatabaseSync,
+  entity: LexicalEntity,
+  removals: LexicalRemoval[],
+): void {
+  if (removals.length === 0) return;
+
+  const entityTable = entity === 'word' ? 'imm_words' : 'imm_kanji';
+  const occurrenceTable =
+    entity === 'word' ? 'imm_word_line_occurrences' : 'imm_kanji_line_occurrences';
+  const col = `${entity}_id`;
+
+  const selectStmt = db.prepare(
+    `SELECT frequency, first_seen AS firstSeen, last_seen AS lastSeen
+     FROM ${entityTable}
+     WHERE id = ?`,
+  );
+  const updateFrequencyStmt = db.prepare(`UPDATE ${entityTable} SET frequency = ? WHERE id = ?`);
+  const hasOccurrencesStmt = db.prepare(
+    `SELECT 1 AS found FROM ${occurrenceTable} WHERE ${col} = ? LIMIT 1`,
+  );
+  const deleteStmt = db.prepare(`DELETE FROM ${entityTable} WHERE id = ?`);
+
+  const needsExactRefresh: number[] = [];
+
+  for (const removal of removals) {
+    const current = selectStmt.get(removal.id) as {
+      frequency: number | null;
+      firstSeen: number | null;
+      lastSeen: number | null;
+    } | null;
+    if (!current) continue;
+
+    const nextFrequency = (current.frequency ?? 0) - removal.removedFrequency;
+    if (nextFrequency <= 0) {
+      // The rows in scope are already gone by now, so anything still pointing at
+      // this entity means the stored frequency was stale rather than exhausted.
+      if (hasOccurrencesStmt.get(removal.id)) {
+        needsExactRefresh.push(removal.id);
+      } else {
+        deleteStmt.run(removal.id);
+      }
+      continue;
+    }
+
+    const removedFirstSeen = toStoredSeenSeconds(removal.removedFirstSeenMs);
+    const removedLastSeen = toStoredSeenSeconds(removal.removedLastSeenMs);
+    const firstSeenMayHaveMoved =
+      current.firstSeen === null ||
+      (removedFirstSeen !== null && removedFirstSeen <= current.firstSeen);
+    const lastSeenMayHaveMoved =
+      current.lastSeen === null ||
+      (removedLastSeen !== null && removedLastSeen >= current.lastSeen);
+    if (firstSeenMayHaveMoved || lastSeenMayHaveMoved) {
+      needsExactRefresh.push(removal.id);
+      continue;
+    }
+
+    updateFrequencyStmt.run(nextFrequency, removal.id);
+  }
+
+  if (entity === 'word') {
+    refreshWordAggregates(db, needsExactRefresh);
+  } else {
+    refreshKanjiAggregates(db, needsExactRefresh);
+  }
+}
+
 function refreshWordAggregates(db: DatabaseSync, wordIds: number[]): void {
   if (wordIds.length === 0) {
     return;
@@ -214,11 +375,18 @@ function refreshWordAggregates(db: DatabaseSync, wordIds: number[]): void {
         SELECT
           w.id AS wordId,
           COALESCE(SUM(o.occurrence_count), 0) AS frequency,
-          MIN(COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)) AS firstSeen,
-          MAX(COALESCE(sl.LAST_UPDATE_DATE, sl.CREATED_DATE)) AS lastSeen
+          MIN(COALESCE(o.seen_ms, (
+            SELECT COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)
+            FROM imm_subtitle_lines sl
+            WHERE sl.line_id = o.line_id
+          ))) AS firstSeen,
+          MAX(COALESCE(o.seen_ms, (
+            SELECT COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)
+            FROM imm_subtitle_lines sl
+            WHERE sl.line_id = o.line_id
+          ))) AS lastSeen
         FROM imm_words w
         LEFT JOIN imm_word_line_occurrences o ON o.word_id = w.id
-        LEFT JOIN imm_subtitle_lines sl ON sl.line_id = o.line_id
         WHERE w.id IN (${makePlaceholders(wordIds)})
         GROUP BY w.id
       `,
@@ -263,11 +431,18 @@ function refreshKanjiAggregates(db: DatabaseSync, kanjiIds: number[]): void {
         SELECT
           k.id AS kanjiId,
           COALESCE(SUM(o.occurrence_count), 0) AS frequency,
-          MIN(COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)) AS firstSeen,
-          MAX(COALESCE(sl.LAST_UPDATE_DATE, sl.CREATED_DATE)) AS lastSeen
+          MIN(COALESCE(o.seen_ms, (
+            SELECT COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)
+            FROM imm_subtitle_lines sl
+            WHERE sl.line_id = o.line_id
+          ))) AS firstSeen,
+          MAX(COALESCE(o.seen_ms, (
+            SELECT COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)
+            FROM imm_subtitle_lines sl
+            WHERE sl.line_id = o.line_id
+          ))) AS lastSeen
         FROM imm_kanji k
         LEFT JOIN imm_kanji_line_occurrences o ON o.kanji_id = k.id
-        LEFT JOIN imm_subtitle_lines sl ON sl.line_id = o.line_id
         WHERE k.id IN (${makePlaceholders(kanjiIds)})
         GROUP BY k.id
       `,

src/core/services/immersion-tracker/storage.ts

@@ -192,6 +192,33 @@ function addColumnIfMissing(
   }
 }
 
+/**
+ * Copy each subtitle line's timestamp onto its word/kanji occurrence rows.
+ *
+ * Vocabulary aggregates used to be recomputed by joining every occurrence back
+ * to `imm_subtitle_lines` just to read two timestamps, which meant one random
+ * read into the widest table per occurrence — the reason deleting a session cost
+ * seconds on a large library. With the timestamp stored alongside the count, the
+ * covering index answers those aggregates on its own.
+ */
+function backfillLexicalOccurrenceSeenMs(db: DatabaseSync): void {
+  for (const table of ['imm_word_line_occurrences', 'imm_kanji_line_occurrences']) {
+    addColumnIfMissing(db, table, 'seen_ms', 'INTEGER');
+    db.exec(`
+      UPDATE ${table}
+      SET seen_ms = (
+        SELECT COALESCE(sl.CREATED_DATE, sl.LAST_UPDATE_DATE)
+        FROM imm_subtitle_lines sl
+        WHERE sl.line_id = ${table}.line_id
+      )
+      WHERE seen_ms IS NULL
+    `);
+  }
+  // Superseded by the covering (entity, seen_ms, occurrence_count, line_id) indexes.
+  db.exec('DROP INDEX IF EXISTS idx_word_line_occurrences_word');
+  db.exec('DROP INDEX IF EXISTS idx_kanji_line_occurrences_kanji');
+}
+
 function dropColumnIfExists(db: DatabaseSync, tableName: string, columnName: string): void {
   if (hasColumn(db, tableName, columnName)) {
     db.exec(`ALTER TABLE ${tableName} DROP COLUMN ${columnName}`);
@@ -923,6 +950,7 @@ export function ensureSchema(db: DatabaseSync): void {
       line_id INTEGER NOT NULL,
       word_id INTEGER NOT NULL,
       occurrence_count INTEGER NOT NULL,
+      seen_ms INTEGER,
       PRIMARY KEY(line_id, word_id),
       FOREIGN KEY(line_id) REFERENCES imm_subtitle_lines(line_id) ON DELETE CASCADE,
       FOREIGN KEY(word_id) REFERENCES imm_words(id) ON DELETE CASCADE
@@ -933,6 +961,7 @@ export function ensureSchema(db: DatabaseSync): void {
       line_id INTEGER NOT NULL,
       kanji_id INTEGER NOT NULL,
       occurrence_count INTEGER NOT NULL,
+      seen_ms INTEGER,
       PRIMARY KEY(line_id, kanji_id),
       FOREIGN KEY(line_id) REFERENCES imm_subtitle_lines(line_id) ON DELETE CASCADE,
       FOREIGN KEY(kanji_id) REFERENCES imm_kanji(id) ON DELETE CASCADE
@@ -1093,6 +1122,7 @@ export function ensureSchema(db: DatabaseSync): void {
         line_id INTEGER NOT NULL,
         word_id INTEGER NOT NULL,
         occurrence_count INTEGER NOT NULL,
+        seen_ms INTEGER,
         PRIMARY KEY(line_id, word_id),
         FOREIGN KEY(line_id) REFERENCES imm_subtitle_lines(line_id) ON DELETE CASCADE,
         FOREIGN KEY(word_id) REFERENCES imm_words(id) ON DELETE CASCADE
@@ -1103,6 +1133,7 @@ export function ensureSchema(db: DatabaseSync): void {
         line_id INTEGER NOT NULL,
         kanji_id INTEGER NOT NULL,
         occurrence_count INTEGER NOT NULL,
+        seen_ms INTEGER,
         PRIMARY KEY(line_id, kanji_id),
         FOREIGN KEY(line_id) REFERENCES imm_subtitle_lines(line_id) ON DELETE CASCADE,
         FOREIGN KEY(kanji_id) REFERENCES imm_kanji(id) ON DELETE CASCADE
@@ -1349,13 +1380,19 @@ export function ensureSchema(db: DatabaseSync): void {
     CREATE INDEX IF NOT EXISTS idx_subtitle_lines_anime_line
     ON imm_subtitle_lines(anime_id, line_index)
   `);
+  if (currentVersion?.schema_version && currentVersion.schema_version < 19) {
+    backfillLexicalOccurrenceSeenMs(db);
+  }
+
+  // Covering indexes: vocabulary aggregates (frequency, first/last seen) are
+  // answered from the index alone, without reading the wide subtitle-line rows.
   db.exec(`
-    CREATE INDEX IF NOT EXISTS idx_word_line_occurrences_word
-    ON imm_word_line_occurrences(word_id, line_id)
+    CREATE INDEX IF NOT EXISTS idx_word_line_occurrences_word_seen
+    ON imm_word_line_occurrences(word_id, seen_ms, occurrence_count, line_id)
   `);
   db.exec(`
-    CREATE INDEX IF NOT EXISTS idx_kanji_line_occurrences_kanji
-    ON imm_kanji_line_occurrences(kanji_id, line_id)
+    CREATE INDEX IF NOT EXISTS idx_kanji_line_occurrences_kanji_seen
+    ON imm_kanji_line_occurrences(kanji_id, seen_ms, occurrence_count, line_id)
   `);
   db.exec(`
     CREATE INDEX IF NOT EXISTS idx_media_art_cover_blob_hash
@@ -1475,21 +1512,23 @@ export function createTrackerPreparedStatements(db: DatabaseSync): TrackerPrepar
     `),
     wordLineOccurrenceUpsertStmt: db.prepare(`
       INSERT INTO imm_word_line_occurrences (
-        line_id, word_id, occurrence_count
+        line_id, word_id, occurrence_count, seen_ms
       ) VALUES (
-        ?, ?, ?
+        ?, ?, ?, ?
       )
       ON CONFLICT(line_id, word_id) DO UPDATE SET
-        occurrence_count = imm_word_line_occurrences.occurrence_count + excluded.occurrence_count
+        occurrence_count = imm_word_line_occurrences.occurrence_count + excluded.occurrence_count,
+        seen_ms = COALESCE(imm_word_line_occurrences.seen_ms, excluded.seen_ms)
     `),
     kanjiLineOccurrenceUpsertStmt: db.prepare(`
       INSERT INTO imm_kanji_line_occurrences (
-        line_id, kanji_id, occurrence_count
+        line_id, kanji_id, occurrence_count, seen_ms
       ) VALUES (
-        ?, ?, ?
+        ?, ?, ?, ?
       )
       ON CONFLICT(line_id, kanji_id) DO UPDATE SET
-        occurrence_count = imm_kanji_line_occurrences.occurrence_count + excluded.occurrence_count
+        occurrence_count = imm_kanji_line_occurrences.occurrence_count + excluded.occurrence_count,
+        seen_ms = COALESCE(imm_kanji_line_occurrences.seen_ms, excluded.seen_ms)
     `),
     videoAnimeIdSelectStmt: db.prepare(`
       SELECT anime_id FROM imm_videos
@@ -1630,11 +1669,16 @@ export function executeQueuedWrite(write: QueuedWrite, stmts: TrackerPreparedSta
     const lineId = Number(lineResult.lastInsertRowid);
     for (const occurrence of write.wordOccurrences) {
       const wordId = incrementWordAggregate(stmts, occurrence, write.firstSeen, write.lastSeen);
-      stmts.wordLineOccurrenceUpsertStmt.run(lineId, wordId, occurrence.occurrenceCount);
+      stmts.wordLineOccurrenceUpsertStmt.run(lineId, wordId, occurrence.occurrenceCount, currentMs);
     }
     for (const occurrence of write.kanjiOccurrences) {
       const kanjiId = incrementKanjiAggregate(stmts, occurrence, write.firstSeen, write.lastSeen);
-      stmts.kanjiLineOccurrenceUpsertStmt.run(lineId, kanjiId, occurrence.occurrenceCount);
+      stmts.kanjiLineOccurrenceUpsertStmt.run(
+        lineId,
+        kanjiId,
+        occurrence.occurrenceCount,
+        currentMs,
+      );
     }
     return;
   }

src/core/services/immersion-tracker/types.ts

@@ -1,4 +1,4 @@
-export const SCHEMA_VERSION = 18;
+export const SCHEMA_VERSION = 19;
 export const DEFAULT_QUEUE_CAP = 1_000;
 export const DEFAULT_BATCH_SIZE = 25;
 export const DEFAULT_FLUSH_INTERVAL_MS = 500;

src/core/services/stats-server/library-routes.ts

@@ -190,4 +190,11 @@ export function registerStatsLibraryRoutes(
     await tracker.deleteVideo(videoId);
     return c.json(statsJson('deleteVideo', { ok: true }));
   });
+
+  app.delete('/api/stats/anime/:animeId', async (c) => {
+    const animeId = parseIntQuery(c.req.param('animeId'), 0);
+    if (animeId <= 0) return c.body(null, 400);
+    await tracker.deleteAnime(animeId);
+    return c.json(statsJson('deleteAnime', { ok: true }));
+  });
 }

src/core/services/stats-sync/merge-occurrences.test.ts

@@ -0,0 +1,100 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import { Database } from '../immersion-tracker/sqlite';
+import { ensureSchema } from '../immersion-tracker/storage';
+import { mergeSnapshotIntoDb } from './merge';
+
+const BASE_MS = 1_700_000_000_000;
+const DAY_MS = 86_400_000;
+
+/**
+ * Build a database holding one anime, one episode, one ended session and a
+ * single word occurrence.
+ *
+ * `legacyOccurrences` reproduces a peer whose schema predates the denormalised
+ * `seen_ms` column, so the merge has to recover the timestamp from the line.
+ */
+function buildDb(
+  dir: string,
+  name: string,
+  options: { word: string; seenMs: number; legacyOccurrences: boolean },
+): string {
+  const dbPath = path.join(dir, name);
+  const db = new Database(dbPath);
+  ensureSchema(db);
+  db.exec(`
+    INSERT INTO imm_anime(anime_id, normalized_title_key, canonical_title, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'key-${name}', 'Show ${name}', ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_videos(video_id, video_key, anime_id, canonical_title, source_type, watched, duration_ms, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'video-${name}', 1, 'Episode 1', 1, 1, 1440000, ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_sessions(session_id, session_uuid, video_id, started_at_ms, ended_at_ms, status, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 'uuid-${name}', 1, '${options.seenMs}', '${options.seenMs + 1000}', 2, ${BASE_MS}, ${BASE_MS});
+    INSERT INTO imm_subtitle_lines(line_id, session_id, video_id, anime_id, line_index, text, CREATED_DATE, LAST_UPDATE_DATE)
+      VALUES (1, 1, 1, 1, 0, 'line', ${options.seenMs}, ${options.seenMs});
+    INSERT INTO imm_words(id, headword, word, reading, part_of_speech, pos1, first_seen, last_seen, frequency)
+      VALUES (1, '${options.word}', '${options.word}', '', 'noun', '名詞',
+              ${Math.floor(options.seenMs / 1000)}, ${Math.floor(options.seenMs / 1000)}, 1);
+  `);
+  db.exec(
+    options.legacyOccurrences
+      ? 'INSERT INTO imm_word_line_occurrences(line_id, word_id, occurrence_count) VALUES (1, 1, 1)'
+      : `INSERT INTO imm_word_line_occurrences(line_id, word_id, occurrence_count, seen_ms) VALUES (1, 1, 1, ${options.seenMs})`,
+  );
+  db.close();
+  return dbPath;
+}
+
+function mergedOccurrences(dbPath: string): Array<{ word: string; seenMs: number | null }> {
+  const db = new Database(dbPath);
+  try {
+    return db
+      .prepare(
+        `SELECT w.word AS word, o.seen_ms AS seenMs
+         FROM imm_word_line_occurrences o
+         JOIN imm_words w ON w.id = o.word_id
+         ORDER BY w.word`,
+      )
+      .all() as Array<{ word: string; seenMs: number | null }>;
+  } finally {
+    db.close();
+  }
+}
+
+for (const legacyOccurrences of [false, true]) {
+  const label = legacyOccurrences ? 'a peer predating the seen_ms column' : 'a current peer';
+
+  test(`sync merge carries occurrence timestamps in from ${label}`, () => {
+    const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'subminer-merge-occurrences-test-'));
+    try {
+      const local = buildDb(dir, 'local.sqlite', {
+        word: '猫',
+        seenMs: BASE_MS,
+        legacyOccurrences: false,
+      });
+      const remoteSeenMs = BASE_MS + 5 * DAY_MS;
+      const remote = buildDb(dir, 'remote.sqlite', {
+        word: '犬',
+        seenMs: remoteSeenMs,
+        legacyOccurrences,
+      });
+
+      mergeSnapshotIntoDb(local, remote);
+
+      const rows = mergedOccurrences(local);
+      assert.equal(rows.length, 2, 'local and remote occurrences both survive the merge');
+      for (const row of rows) {
+        assert.ok(row.seenMs, `${row.word} must carry a timestamp so aggregates stay index-only`);
+      }
+      assert.equal(
+        Number(rows.find((row) => row.word === '犬')?.seenMs),
+        remoteSeenMs,
+        "the remote line's own timestamp is preserved rather than stamped with merge time",
+      );
+    } finally {
+      fs.rmSync(dir, { recursive: true, force: true });
+    }
+  });
+}

src/core/services/stats-sync/merge-sessions.ts

@@ -215,12 +215,16 @@ function copySubtitleLines(
     'SELECT kanji_id, occurrence_count FROM imm_kanji_line_occurrences WHERE line_id = ?',
   );
   const insertWordOccurrence = local.query(
-    `INSERT INTO imm_word_line_occurrences (line_id, word_id, occurrence_count) VALUES (?, ?, ?)
-     ON CONFLICT(line_id, word_id) DO UPDATE SET occurrence_count = occurrence_count + excluded.occurrence_count`,
+    `INSERT INTO imm_word_line_occurrences (line_id, word_id, occurrence_count, seen_ms) VALUES (?, ?, ?, ?)
+     ON CONFLICT(line_id, word_id) DO UPDATE SET
+       occurrence_count = occurrence_count + excluded.occurrence_count,
+       seen_ms = COALESCE(seen_ms, excluded.seen_ms)`,
   );
   const insertKanjiOccurrence = local.query(
-    `INSERT INTO imm_kanji_line_occurrences (line_id, kanji_id, occurrence_count) VALUES (?, ?, ?)
-     ON CONFLICT(line_id, kanji_id) DO UPDATE SET occurrence_count = occurrence_count + excluded.occurrence_count`,
+    `INSERT INTO imm_kanji_line_occurrences (line_id, kanji_id, occurrence_count, seen_ms) VALUES (?, ?, ?, ?)
+     ON CONFLICT(line_id, kanji_id) DO UPDATE SET
+       occurrence_count = occurrence_count + excluded.occurrence_count,
+       seen_ms = COALESCE(seen_ms, excluded.seen_ms)`,
   );
 
   for (const row of rows) {
@@ -241,17 +245,20 @@ function copySubtitleLines(
       ],
     );
     summary.subtitleLinesAdded += 1;
+    // Taken from the line rather than the remote occurrence row so this also
+    // works when the remote database predates the seen_ms column.
+    const seenMs = row.CREATED_DATE ?? row.LAST_UPDATE_DATE ?? null;
 
     for (const occurrence of wordOccurrences.all(row.line_id) as SqlRow[]) {
       const localWordId = lexicon.resolveWord(Number(occurrence.word_id));
       const count = Number(occurrence.occurrence_count);
-      insertWordOccurrence.run(localLineId, localWordId, count);
+      insertWordOccurrence.run(localLineId, localWordId, count, seenMs);
       lexicon.addWordOccurrences(Number(occurrence.word_id), count);
     }
     for (const occurrence of kanjiOccurrences.all(row.line_id) as SqlRow[]) {
       const localKanjiId = lexicon.resolveKanji(Number(occurrence.kanji_id));
       const count = Number(occurrence.occurrence_count);
-      insertKanjiOccurrence.run(localLineId, localKanjiId, count);
+      insertKanjiOccurrence.run(localLineId, localKanjiId, count, seenMs);
       lexicon.addKanjiOccurrences(Number(occurrence.kanji_id), count);
     }
   }

src/types/stats-http-contract.ts

@@ -141,6 +141,7 @@ export interface StatsJsonResponseMap {
   deleteSessions: StatsOkResponse;
   deleteSession: StatsOkResponse;
   deleteVideo: StatsOkResponse;
+  deleteAnime: StatsOkResponse;
   anilistSearch: StatsAnilistSearchResult[];
   knownWords: string[];
   knownWordsSummary: StatsKnownWordsSummary;
@@ -219,6 +220,7 @@ export interface StatsHttpClient {
   deleteSession: (sessionId: number) => Promise<void>;
   deleteSessions: (sessionIds: number[]) => Promise<void>;
   deleteVideo: (videoId: number) => Promise<void>;
+  deleteAnime: (animeId: number) => Promise<void>;
   getKnownWords: () => Promise<string[]>;
   getKnownWordsSummary: () => Promise<StatsKnownWordsSummary>;
   getAnimeKnownWordsSummary: (animeId: number) => Promise<StatsKnownWordsSummary>;