The "MCP conformance tests" job fails intermittently on unrelated PRs, always on `tasks-wire-fields`, `tasks-mrtr-input` and `tasks-dispatch-and-envelope`. The cause is in pydocket 0.24.1. When `docket.cancel()` reaches the worker after it has picked up a task but before the task has been claimed, the cancellation ends the task with a `CancelledError`. The worker loop awaits that finished task and treats the error as a request to shut down, so `run_forever()` returns without a log line and the worker never reads another task. The conformance run cancels a 60 second task right after creating it, so on a slow runner it hits this window; every task created afterwards stays in `working` until the scenario's own timeout. pydocket 0.26.0 fixes this (chrisguidry/docket#503), so this raises the floor in `fastmcp_tasks` and updates the lock. A regression test cancels a task while its claim is in flight and then checks that a second task still completes; it times out on 0.24.1 and passes on 0.26.0.
4900dd5c5a