Summary

Closes #3212

Fixes conversation context silently truncating to 4 messages after a gateway restart for sessions with long JSONL history.

Root Cause

_flush_messages_to_session_db deliberately skips messages already in conversation_history to avoid duplicate writes (the #860 fix). The assumption is that those messages are already persisted in SQLite. That assumption is false for:

• Sessions created before the SQLite layer was introduced
• Sessions after a deployment that reset or replaced the DB
• Any session where the DB was unavailable during earlier turns

Exact failure trace for the reporter's session (994 JSONL messages):

Turn N  (first after restart):
  load_transcript(id)          → SQLite: 0 rows → falls back to JSONL: 994 ✓
  _flush_messages_to_session_db:
    start_idx = len(conversation_history) = 994
    flush_from = max(994, 0) = 994
    writes messages[994:] = [user_msg, assistant_msg]  → SQLite now has 2 rows

Turn N+1:
  load_transcript(id)          → SQLite: 2 rows → returns immediately ✗
  agent sees 2 messages of history
  _save_session_log writes messages=[2 history + 2 new] = 4 to session JSON

This matches the reporter's exact evidence: JSONL 994 ✅, session JSON 4 ❌.

Fix

load_transcript now loads both SQLite and JSONL, then returns whichever has more messages.

if len(jsonl_messages) > len(db_messages):
    # Legacy session not yet fully migrated — use JSONL
    return jsonl_messages
return db_messages

For a fully-migrated session, SQLite will always be ≥ JSONL (JSONL is append-only and stops being written once SQLite takes over). There is no regression for normal sessions.

The JSONL read is a sequential scan; for very large transcripts it adds a small I/O cost, but load_transcript is called once per gateway turn and the file is already in the OS page cache for active sessions. The trade-off is acceptable.

Note on a related bug

This is a separate fix from #3210 / PR #3220 (cached agent session_id mismatch after session reset). That PR covers the case where a session reset causes the agent to write to the wrong session. This PR covers the case where a legacy JSONL session is never bootstrapped into SQLite, causing subsequent turns to see only the most recent 2 messages.

Tests

1453 gateway tests + 456 session tests pass. No regressions.