[iris] Controller-side grace window for StartTasks/PollTasks race

[rjpower]

When a PollTasks RPC's DB snapshot is taken before a concurrent StartTasks' assignment commit, the poll omits the new task from expected_tasks and the worker kills it as unexpected (or reports WORKER_FAILED for a task it has not yet received). Track recent StartTasks dispatches in-memory for 30s, merge them into the per-worker expected set, and drop any updates for tasks still inside the grace window so the next poll settles them cleanly. Complements #5043.

[claude]

Claude finished @rjpower's task in 4m 44s —— View job

---

Code review

• Check PR state and prior reviews
• Identify relevant CLAUDE.md/AGENTS.md files
• Read diff and surrounding controller context
• Analyze for bugs and CLAUDE.md compliance
• Post final review

No issues found. Checked for bugs and CLAUDE.md compliance.

Notes from the analysis (informational, not blocking):

• The time.monotonic() use in controller.py is consistent with existing usage at controller.py:1431. The Iris AGENTS.md note about preferring rigging.timing is scoped to Timestamp/Duration/Deadline/Timer/ExponentialBackoff, not monotonic-clock grace windows.
• Race ordering in _poll_all_workers looks correct: DB snapshot first, then _prune_and_snapshot_recent_dispatches(), then the merge into the per-worker expected set, then the poll_workers RPC. The suppression of poll-side updates is keyed off the same recent snapshot, so any task whose dispatch was visible at snapshot time is also exempt from the kill-as-unexpected report.
• Stale entries left behind on StartTasks-failure or ack-rejection paths are harmless: they age out after _RECENT_DISPATCH_GRACE_SECONDS and any subsequent attempt uses a different attempt_id.
  | Branch: fix-controller-race

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

[ravwojdyla]

🚀

[ravwojdyla]

I'm scared of these magic numbers - but it makes sense!

[rjpower]

yeah I want to get rid of this soon, we'll switch to a more sensible model without the race

[rjpower]

🤖 Superseded by #5090, which fixes the same race at the source by moving PollTasks inline in the scheduling loop (single thread now owns both the assignment commit and the running-tasks snapshot). The _recent_dispatches bookkeeping in this PR becomes unnecessary — closing.