eabf2e46f9
Under the 40-worker file runner, service.stop(timeout=1.0) / runtime.stop(timeout=0.5) and the 2s _wait_for lost the race to scheduler latency (a different test each run, green on retry). Bounds move to the 5s the other 25 sites already use; the bounded-stop invariant keeps a 2s ceiling, still far inside the join timeout.