Session indexing ignores failed FTS acknowledgment and leaves stale searchable content

[Calmingstorm]

Confirmed split-index success misreport

Reviewed master at 886c36d8ebe861aa987059a1744d45b78797baae (v4.7.0). Suggested priority: P2.

SessionVectorStore._write_session_sync ignores the boolean acknowledgment from FTS indexing, commits current metadata/vector state and reports success even when FTS rejected the write. Backfill checks missing IDs rather than stale content, so an existing stale FTS row does not recover.

Temporary-store reproduction

Index an archive with old text; make only FTS index_session return its documented failure value False; replace the archive with new text; restore the FTS method and run backfill.

replacement index return: True
metadata: new replacementphrase
FTS old text found: True
FTS new text found: False
backfill new count: 0

Independently reproduced by three reviewers using disposable SQLite stores. The fault is injected at the FTS acknowledgment boundary, modeling a split-store write failure, not a observed live incident.

Acceptance criteria

• Honor failed FTS acknowledgments and do not report the multi-index update fully successful.
• Coordinate rollback/repair across the separate stores without losing previously searchable data.
• Detect stale content/version mismatches, not only missing document IDs, during repair/backfill.
• Test FTS failure with successful metadata DB writes and successful retry repair.

No source changes were made.

[Calmingstorm]

Closing after independent verification at 886c36d.

The FTS acknowledgment is indeed ignored in SessionVectorStore._write_session_sync, but the harmful scenario this issue describes — re-indexing an existing archive whose text changed and leaving stale FTS content — cannot occur:

• Search documents are keyed by the archive file name (doc_id = archive_path.stem, i.e. <channel>_<timestamp>), and SessionManager._archive_session always writes a new, uniquely named archive atomically (it advances the timestamp until the name is free). Nothing rewrites an archive afterwards, so a document id's content never changes.
• The reachable case — a new archive whose FTS write fails while the metadata/vector write succeeds — heals itself: the startup archive backfill (SessionVectorStore.backfill, fired on every start) re-adds any session present in the metadata store but missing from FTS.

Honouring the boolean would only change a log line today. If archives ever become mutable, this should be reopened together with the stale-content repair requirement.