🚨 P0 HOTFIX — long-poll holds DB txn → pool exhaustion

Cross-port from cueapi/cueapi#925 (private, mergeCommit 359961b).

Incident: a downstream consumer (cue.dock.svc) experienced pool exhaustion in prod due to _run_long_poll_wait holding an idle in transaction connection for the full ~30s wait window.

Mechanism

At ~15 concurrent long-pollers (SQLAlchemy default pool_size=5 + max_overflow=10 = 15):

GET /v1/agents/{ref}/events?wait=long appears to BEGIN a transaction,
SELECT events, and then HOLD that transaction (and its connection) 
open for the full ~30s wait when there are no new events.

pg_stat_activity:
- 15 connections stuck in `idle in transaction`
- Each holding `SELECT events.id, events.event_type ...` for 14-20+ seconds
- Matches SQLAlchemy default pool ceiling
- App-side pool 100% saturated by waiting long-polls

Root cause: _run_long_poll_wait while loop runs await asyncio.sleep + await pull_events(db, ...) with no commit between iterations. Each pull_events opens an implicit txn that the session holds for the full ~30s wait.

Fix (+2 LOC)

1. await db.commit() at function entry — releases caller's initial-pull implicit txn before first sleep window
2. await db.commit() after each empty pull_events() iteration — releases per-iteration implicit txn before next sleep window

Empirical verification (load-bearing)

The fix relies on AsyncSession.commit() returning the connection to the pool. Verified empirically against real Postgres + asyncpg + SQLAlchemy 2.0.35 via engine.pool.checkedin()/checkedout() state inspection:

After SELECT 1: checked_in=0, checked_out=1
After commit:   checked_in=1, checked_out=0   ← RETURNED to pool ✓
After sleep:    checked_in=1, checked_out=0   ← stays in pool
After SELECT 2: checked_in=0, checked_out=1   ← re-acquired transparently

Session remains usable across commits; both Postgres idle in transaction AND SQLAlchemy pool slot exhaustion are addressed.

Tests

• 2 new regression-guard tests added pinning the commit invariant:
  • test_helper_commits_at_entry_and_between_empty_iterations
  • test_helper_commits_at_entry_even_when_event_arrives_immediately
• 13/13 long-poll tests pass locally

Tag plan

Suggesting messaging-v1.1.5-hotfix tag on merge. Single-purpose hotfix tag for downstream consumers to pin-bump on.

Parity reference

Same diff shape as private cueapi/cueapi#925 (mergeCommit 359961b). Per CLAUDE.md OSS-vs-private parity policy — app/routers/events.py:_run_long_poll_wait shares verbatim across both repos.

🤖 Generated with Claude Code