fix(worker): address XBeg9's PR #39 follow-up findings

[rcbevans]

What

8 fixes from XBeg9's final approved review on #39 (merged without being addressed — rcbevans: "To make progress I'll merge as is and address remaining comments suggestions in a follow up").

Worth fixing

1. Cron loop error classification (leader.py): the cron loop hand-rolled its error classification with isinstance checks that missed 7 of 12 TRANSIENT_PG_ERRORS (DeadlockDetectedError, SerializationError, AdminShutdownError, CannotConnectNowError, TooManyConnectionsError, IdleSessionTimeoutError, IdleInTransactionSessionTimeoutError). Deadlock/serialization inside a transaction are routine, but 5 consecutive now killed the worker via the backstop guard. Fixed by catching TRANSIENT_PG_ERRORS first, keeping the narrower conn-state check inside it.

Backlog

2. Guard-before-cleanup ordering (leader.py): guard.unexpected(exc) raised before cleanup at all 4 backstop sites (election probe, lock attempt, upsert, watchdog probe). On the fatal iteration cleanup never ran, leaving is_leader set and leader_conn dangling. Fixed by moving cleanup before guard.unexpected at all 4 sites.

3. guard.ok() reset on probe path (leader.py): the probe's unexpected branch didn't continue, so the iteration fell through to re-election and reached guard.ok() at the bottom — a probe-only fault alternated unexpected/ok and never reached the cap. Fixed by adding continue.

4. Producer/validator mismatch (settings.py): the staleness invariant checked the producer loop, but the producer is not wrapped in asyncio.timeout (dispatch_batch is multi-statement). The timeout + period model doesn't hold. Removed the producer from the validator; updated the description.

5. connect() no timeout (deps.py): open_dedicated_conn passed command_timeout but not timeout= to asyncpg.connect, so connection establishment kept the driver's 60s default. Fixed by passing timeout=command_timeout.

6. Cron timeout continues past sleep (leader.py): the deadline and conn-state branches used continue, skipping the trailing asyncio.sleep(1). A persistently slow PG was re-issued BEGIN + tick_cron back-to-back. Fixed by removing continue so all branches fall through to the sleep.

7. settings.py wording: "every bounded loop" was inaccurate — 3 loops (leader.election, leader.watchdog, progress_flush) are not checked. Updated to say "the period-1 leader loops".

8. _transient.py doc: QueryCanceledError was described as "asyncpg's OTHER shape for a fired command_timeout". XBeg9 tested with real PG 18: command_timeout raises TimeoutError, not QueryCanceledError. QueryCanceledError is server-side 57014 (pg_cancel_backend or server-side statement_timeout). Fixed the comment.

Testing

14 new TDD regression tests in tests/test_pr39_followup_fixes.py — each written RED first, then GREEN. Updated tests/test_settings.py to reflect the producer's removal from the validator.

Verification

• unit: 4011 passed, 9 skipped
• integration: 986 passed, 6 skipped (1 flaky test_property_sweep_equivalence passes on rerun)
• e2e: 52 passed (all 41 original + 4 admin/SSE + 7 others)
• ruff check, ruff format --check: clean
• pyright: 0 new errors (pre-existing missing-extra errors only)