fix(stats): drop first_seen repair from lexical rollup rebuild

A stale first_seen is indistinguishable from real history whose occurrences
retention pruned, and frequency does not reliably separate them (it diverges
from surviving occurrences on databases with neither retention nor
pre-occurrence history), so the guarded repair either did nothing or risked
overwriting correct dates. The delete path already keeps first_seen exact for
new deletes; the rebuild now repairs only last_seen.

src/core/services/immersion-tracker/lexical-rollups.test.ts

@@ -182,7 +182,7 @@ test('re-seeing a filtered legacy word keeps charts consistent with the summary'
   }
 });
 
-test('lexical rollup rebuild repairs seen bounds left pointing at deleted history', () => {
+test('lexical rollup rebuild repairs last_seen left pointing at deleted history', () => {
   const db = createDb();
 
   try {
@@ -197,21 +197,20 @@ test('lexical rollup rebuild repairs seen bounds left pointing at deleted histor
         .run(sessionId, videoId, BASE_MS, BASE_MS).lastInsertRowid,
     );
     const seen = (BASE_MS + 251) / 1000;
-    const staleFirst = seen - 86_400.5;
+    // first_seen this early may be real history pruned by retention; never touched.
+    const earlyFirst = seen - 86_400.5;
     const staleLast = seen + 86_400.631;
     db.prepare(
       `INSERT INTO imm_words(id, headword, word, reading, first_seen, last_seen, frequency)
-       VALUES (1, '猫', '猫', 'ねこ', ?, ?, 1), (2, '犬', '犬', 'いぬ', ?, ?, 1),
+       VALUES (1, '猫', '猫', 'ねこ', ?, ?, 1), (2, '犬', '犬', 'いぬ', ?, ?, 1)`,
-         (3, '鳥', '鳥', 'とり', ?, ?, 5)`,
+    ).run(earlyFirst, staleLast, seen - 0.1, seen - 0.1);
-      // Word 3's earlier history was pruned by retention: frequency still counts it.
-    ).run(staleFirst, staleLast, seen - 0.1, seen - 0.1, staleFirst, seen);
     db.prepare(
       `INSERT INTO imm_kanji(id, kanji, first_seen, last_seen, frequency) VALUES (1, '猫', ?, ?, 1)`,
-    ).run(staleFirst, staleLast);
+    ).run(earlyFirst, staleLast);
     db.prepare(
       `INSERT INTO imm_word_line_occurrences(line_id, word_id, occurrence_count, seen_ms)
-       VALUES (?, 1, 1, ?), (?, 2, 1, ?), (?, 3, 1, ?)`,
+       VALUES (?, 1, 1, ?), (?, 2, 1, ?)`,
-    ).run(lineId, BASE_MS + 251, lineId, BASE_MS + 251, lineId, BASE_MS + 251);
+    ).run(lineId, BASE_MS + 251, lineId, BASE_MS + 251);
     db.prepare(
       `INSERT INTO imm_kanji_line_occurrences(line_id, kanji_id, occurrence_count, seen_ms)
        VALUES (?, 1, 1, ?)`,
@@ -231,13 +230,12 @@ test('lexical rollup rebuild repairs seen bounds left pointing at deleted histor
     assert.deepEqual(
       bounds('imm_words'),
       [
-        [seen, seen],
+        [earlyFirst, seen],
         [seen - 0.1, seen - 0.1],
-        [staleFirst, seen],
       ],
-      'flush-lag and retention-pruned rows keep their first_seen',
+      'healthy rows are untouched',
     );
-    assert.deepEqual(bounds('imm_kanji'), [[seen, seen]]);
+    assert.deepEqual(bounds('imm_kanji'), [[earlyFirst, seen]]);
   } finally {
     db.close();
   }

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

@@ -19,49 +19,37 @@ const LOCAL_EPOCH_DAY_SQL = `
 `;
 
 // v3 re-derives visibility for rows a re-seen word forced visible and repairs
-// first_seen/last_seen values left behind by session deletes.
+// last_seen values left behind by session deletes.
 const LEXICAL_DAILY_ROLLUP_VERSION = '3';
 const LEXICAL_DAILY_ROLLUP_VERSION_KEY = 'lexical_daily_rollups_version';
 const VOCABULARY_VISIBILITY_SCAN_BATCH_SIZE = 5_000;
-// Healthy first_seen trails its oldest seen_ms only by queue-to-flush lag.
-const FIRST_SEEN_FLUSH_LAG_SLACK_SEC = 60;
 
 export function localEpochDaySql(value: string): string {
   return LOCAL_EPOCH_DAY_SQL.replaceAll('%VALUE%', value);
 }
 
-// Session deletes once compared rounded timestamps against stored bounds and
+// Session deletes once rounded removed timestamps to whole seconds, missing
-// could leave first_seen/last_seen pointing at deleted history. Bounds are
+// fractional last_seen extremes that pointed at deleted history. last_seen is
-// stamped at queue time, before seen_ms is stamped at flush, so healthy rows
+// stamped at queue time, before seen_ms is stamped at flush, and retention
-// never exceed their newest occurrence and only trail their oldest by flush lag.
+// prunes the oldest sessions first, so healthy rows never exceed their newest
-// Rows with undated occurrences are skipped.
+// surviving occurrence. Rows with undated occurrences are skipped.
 //
-// first_seen can also legitimately predate every surviving occurrence: retention
+// first_seen is deliberately not repaired: a stale bound is indistinguishable
-// pruning cascades away old lines, and early history predates the occurrence
+// from real history whose occurrences retention pruned, and frequency does not
-// tables, while both keep the lifetime frequency. Deletes subtract frequency
+// reliably tell them apart.
-// exactly, so only rows whose frequency matches their surviving occurrences have
+function repairStaleLastSeen(db: DatabaseSync): void {
-// their whole history on hand and a provably stale first_seen.
-function repairStaleSeenBounds(db: DatabaseSync): void {
   for (const [table, occurrenceTable, col] of [
     ['imm_words', 'imm_word_line_occurrences', 'word_id'],
     ['imm_kanji', 'imm_kanji_line_occurrences', 'kanji_id'],
   ] as const) {
-    const seenSec = (fn: 'MIN' | 'MAX'): string =>
+    const newestSeen = `(SELECT MAX(o.seen_ms) FROM ${occurrenceTable} o WHERE o.${col} = ${table}.id) / 1000.0`;
-      `(SELECT ${fn}(o.seen_ms) FROM ${occurrenceTable} o WHERE o.${col} = ${table}.id) / 1000.0`;
-    const hasNoUndated = `NOT EXISTS (
-      SELECT 1 FROM ${occurrenceTable} o WHERE o.${col} = ${table}.id AND o.seen_ms IS NULL
-    )`;
     db.exec(`
       UPDATE ${table}
-      SET first_seen = ${seenSec('MIN')}
+      SET last_seen = ${newestSeen}
-      WHERE first_seen < ${seenSec('MIN')} - ${FIRST_SEEN_FLUSH_LAG_SLACK_SEC}
+      WHERE last_seen > ${newestSeen}
-        AND ${hasNoUndated}
+        AND NOT EXISTS (
-        AND frequency = (
+          SELECT 1 FROM ${occurrenceTable} o WHERE o.${col} = ${table}.id AND o.seen_ms IS NULL
-          SELECT SUM(o.occurrence_count) FROM ${occurrenceTable} o WHERE o.${col} = ${table}.id
+        )
-        );
-      UPDATE ${table}
-      SET last_seen = ${seenSec('MAX')}
-      WHERE last_seen > ${seenSec('MAX')} AND ${hasNoUndated};
     `);
   }
 }
@@ -226,7 +214,7 @@ export function rebuildLexicalDailyRollups(db: DatabaseSync): void {
       lastId = vocabularyRows[vocabularyRows.length - 1]!.id;
       if (vocabularyRows.length < VOCABULARY_VISIBILITY_SCAN_BATCH_SIZE) break;
     }
-    repairStaleSeenBounds(db);
+    repairStaleLastSeen(db);
     db.exec('DELETE FROM imm_lexical_daily_rollups');
     db.exec(`
       INSERT INTO imm_lexical_daily_rollups(epoch_day, word_count, word_count_without_names, kanji_count)