Runner hangs in job cancellation redelivery loop on v2.336.0 and never exits (#4312 not fully fixed)

Describe the bug

On v2.336.0, an ephemeral runner (ARC gha-runner-scale-set, dind mode) received a job cancellation and then hung indefinitely in a cancellation redelivery loop: the JobDispatcher logged Job cancellation request <id> received, cancellation timeout 5 minutes. several times per second, forever. The runner never completed the cancellation and never exited. GitHub deregistered it server side after the timeout, leaving a pod that Kubernetes considers healthy but that GitHub no longer counts as a registered runner.

This is the same failure mode as #4312, which was closed as completed on 2026-07-15, presumably on the basis of the v2.336.0 changes ("Exit ephemeral runners on broker acknowledge job-not-found" #4540, "Wait for worker to finish during cancel"). Our incident occurred on v2.336.0, roughly six hours after rolling it out, so those changes do not fully cover this path. Related: actions/actions-runner-controller#4307, actions/actions-runner-controller#4155.

To Reproduce

Not deterministic; observed occurrence:

1. Ephemeral runner (ARC scale set, dind) picks up a job that runs containers via the dind sidecar
2. The workflow run is cancelled while that job is executing (fail-fast from a matrix job failure)
3. Runner receives the cancellation but never finishes it; JobDispatcher loops on the redelivered cancellation message indefinitely

Expected behavior

After the 5 minute cancellation timeout, the runner force-terminates the worker and exits (it is ephemeral, and its registration is gone), so the controller can replace it.

Runner Version and Platform

v2.336.0, Linux x64 container (ARC gha-runner-scale-set 0.12.0 on GKE, dind mode).

What's not working?

Timeline (2026-07-22, UTC):

• 11:24 job starts on the ephemeral runner
• 11:29 the workflow run is cancelled
• 11:33:58 runner logs Job cancellation request d6802bf9-7688-5b26-acc6-8d3f9e224022 received, cancellation timeout 5 minutes.
• The same line then repeats several times per second, indefinitely (still looping 40+ minutes later when we intervened)
• End state: pod Running 2/2 in Kubernetes, runner absent from GitHub. Scale set stats: totalAssignedJobs: 1, totalRegisteredRunners: 0. The listener counted the zombie towards its target, so the next queued job waited ~25 minutes until we manually deleted the EphemeralRunner

Job Log Output

The cancelled job shows as cancelled in the UI; the hang is runner side only.

Runner and Worker's Diagnostic Logs

Runner container log excerpt (identical lines repeating at ~3 Hz until manual pod deletion):

[RUNNER 2026-07-22 11:33:58Z INFO JobDispatcher] Job cancellation request d6802bf9-7688-5b26-acc6-8d3f9e224022 received, cancellation timeout 5 minutes.
[RUNNER 2026-07-22 11:33:58Z INFO JobDispatcher] Job cancellation request d6802bf9-7688-5b26-acc6-8d3f9e224022 received, cancellation timeout 5 minutes.
[RUNNER 2026-07-22 11:33:59Z INFO JobDispatcher] Job cancellation request d6802bf9-7688-5b26-acc6-8d3f9e224022 received, cancellation timeout 5 minutes.
...

Happy to provide fuller diagnostic logs on request.

Start with the JobDispatcher cancellation path and compare the repeated cancellation logs with the v2.336.0 changes referenced in #4540 and #4312. Reproduce cancellation of a container-running job on an ephemeral Linux runner, then verify that the worker is force-terminated after five minutes and the runner exits so it can be replaced.