fix(history): persist query_history rows by encoding result_json as jsonb

[AttiR]

Summary

Fixes a silent data-loss bug where every query's row in query_history was dropped, while the per-user rate-limit counter still incremented. Users saw their budget badge decrement to 1 of 2 remaining but the History panel stayed empty and cached replays were impossible.

Root cause

_log_query_history in backend/app/agent/sql_agent.py passed a Python dict as the bind value for the result_json JSONB column. asyncpg has no built-in dict → jsonb encoder, so the INSERT raised DataError. The error was caught by a blanket except Exception: logger.exception("... non-fatal") block and never surfaced, so the regression was invisible in logs that weren't tailed carefully.
The earlier UPDATE users SET total_queries = ... (no JSON) committed normally — hence the asymmetry between budget and history.

Changes

• backend/app/agent/sql_agent.py
  • Serialise the result cache with json.dumps({"rows": ..., "columns": ...}, default=str).
  • Add explicit $4::jsonb cast in the INSERT.
  • Keep audit-log write best-effort (a successful query must not 500 because the audit log failed) but log structured context (user_id, row_count) so future regressions are visible in Azure Container Apps logs.
• tests/test_sql_agent.py — two new regression tests under TestLogQueryHistoryJsonbEncoding:
  • test_passes_json_string_and_jsonb_cast — asserts the bind value is a str and the SQL contains $4::jsonb.
  • test_db_failure_is_swallowed_not_propagated — verifies that a transient DB error during the audit-log write does not propagate to the user.
• tests/conftest.py — setdefault JWT_SECRET_KEY and GOOGLE_CLIENT_ID so unit tests load without a fully populated .env.

Test plan

• pytest tests/test_sql_agent.py -v — 27 passed (25 existing + 2 new)
• Manual: sign in to production, run one query, confirm the row appears in query_history and the History panel shows it
• Manual: click the history item, confirm result loads via GET /api/history/:id without a Claude call (from_cache: true)
• Manual: confirm users.total_queries matches COUNT(*) FROM query_history WHERE user_id = me after the fix

Risk

Low — change is scoped to one function and adds an explicit cast that PostgreSQL accepts unconditionally. Existing history rows are unaffected; new rows will be inserted correctly.

Rollback

Revert this commit. The previous behaviour was data-loss but not a crash, so rollback is safe at any time.

Closes #45