[rudycelekli]

Problem and change

UnifiedSwarmCoordinator.cancelTask() sent a task_fail notification to its assigned agent's bus subscription. That subscription interpreted the notification as a worker failure report and retried the task after cancellation returned. With no retries remaining, cancellation instead became a failure and incremented failure counters. Cancelling the old task again could also release an agent already assigned to new work.

Set the cancellation state before yielding to the bus, make cancellation of settled tasks idempotent, and ignore completion/failure reports for cancelled, completed, or failed tasks. This preserves the cancelled verdict and prevents duplicate completion accounting.

Validation

• Against main 1882006a02487590c8c66a7eec8d7504f6af33c3, the eight-case regression suite has 6 failures / 2 passes. Candidate: 8/8 pass.
• Three initial reproductions use only public coordinator methods and event-loop turns. Additional tests exercise real MessageBus report delivery, ordinary failure retries, duplicate completion, and both late-report and explicit-cancel cleanup after existing wait timeouts.
• 285/285 tests pass across all 19 swarm package test files using its checked-in Vitest configuration. Running from the repository root initially used a shorter default test timeout; rerunning through the package configuration resolves those timing failures.
• Independent review reran 38/38 coordination tests and verified the timeout cleanup controls.
• Full swarm package TypeScript build, strict targeted typecheck, and compiled-JavaScript public API smoke pass; git diff --check passes.

This fixes the canonical coordinator's cancellation lifecycle. Existing wait-timeout behavior remains unchanged: a wait timeout does not stop a worker, and late reports can still clean up that assignment. No new dependency or public method is introduced.