fix(state): repair duplicate session titles without data loss on startup (#65602)

[Tranquil-Flow]

Summary

SessionDB._init_schema() creates a UNIQUE partial index on sessions.title but only catches sqlite3.OperationalError. Existing databases that contain duplicate non-null titles (from imports, context-compression continuations, or concurrent creation) raise sqlite3.IntegrityError instead, which is uncaught — crashing Dashboard API calls that open the session database (/api/sessions/{id}/messages).

Root cause

Commit 60b6abefd (session naming with unique titles) added the CREATE UNIQUE INDEX guarded only by an OperationalError handler ("index already exists"). It does not handle the case where the index does not yet exist but the table already holds duplicate titles — which throws IntegrityError.

Fix

Catch sqlite3.IntegrityError and repair the duplicates without deleting any data:

• Of each duplicate-title group, retain the title on the newest inserted session (newer.rowid > older.rowid) and set title = NULL only on the older conflicting aliases.
• Log how many titles were cleared.
• Retry the unique-index creation.
• The existing OperationalError ("index already exists") handler is preserved.

Titles are optional aliases, so clearing an older alias does not lose the session or its messages.

Why not delete duplicate rows

The alternative (deleting duplicate sessions) causes severe transcript data loss. Competing PR #65628 uses:

DELETE FROM sessions WHERE rowid NOT IN (
  SELECT MAX(rowid) FROM sessions
  WHERE title IS NOT NULL GROUP BY title
)

That subquery only selects MAX(rowid) within non-null-title groups, so every untitled session (the majority of normal sessions) is deleted, along with all but the newest duplicate. This fix preserves every row.

Tests

New TestSessionTitleIndexRepair class in tests/test_hermes_state.py (3 production-path tests against a real SQLite database):

Test	What it asserts
test_duplicate_titles_are_repaired_without_deleting_sessions	All sessions + messages preserved; newest keeps the alias, older duplicate cleared; unique index exists after repair
test_repaired_index_rejects_future_duplicate_title	After repair, reusing the title via set_session_title is rejected (ValueError)
test_clean_legacy_database_keeps_existing_titles	A clean database (no duplicates) initializes unchanged

RED proof (on upstream/main before the production change): 2 of 3 tests fail with sqlite3.IntegrityError: UNIQUE constraint failed: sessions.title at _init_schema; the clean-database test passes.

GREEN after fix: focused class 3 passed; full tests/test_hermes_state.py 369 passed. ruff check clean.

Fixes #65602

---

Auto-published by Moonsong via Path B automated pipeline.

[teknium1]

Merged via PR #65689 — your commit was cherry-picked onto current main with your authorship preserved in git log (rebase-merge). We added one small follow-up on top wrapping the repair itself in a guard so index creation can never abort DB open. The no-data-loss UPDATE approach and the RED-proven tests were exactly right — thanks!