AIP-97 fast-follow: classify node drain and preemption as infra

[1fanwang]

Fast-follow on #66405 (stacked on its head). Kept out of that PR to hold its scope to the general mechanism.

Why

A node drain and a scheduler preemption are infrastructure taking the pod away, but classify_pod_failure() reports both as application, so neither gets the refund AIP-97 exists to give. Node drain covers cordon/drain and autoscaler scale-down, so this is the common case in a real cluster, not a corner.

Verified on a live kind cluster (k8s v1.35.0) by recording every watch event: both disruptions reach phase=Failed with pod.status.reason empty and the container reading only Error/exit 143, which is indistinguishable from an application crash. The DELETED event also arrives at phase=Failed, so the watcher's deleted-pod branch never fires either.

The signal is pod.status.conditions[type=DisruptionTarget]. The control plane sets it before the delete and it survives onto the terminal object, which is exactly where the classifier can read it.

What

collect_pod_failure_details() records the condition's reason, and classify_pod_failure() checks it before the reason sets. Gated on status: "True", matching Kubernetes' own podFailurePolicy matcher, since the writers update the condition in place. The four reasons are pinned to their Go definitions, same convention as the existing set.

Testing

Red/green against the pre-fix source, then the live cluster runs the fix is built from.

Red to green

$ git stash push -- providers/cncf/kubernetes/src/   # tests against pre-fix source
$ pytest .../test_classify_pod_failure.py -q -p no:randomly
8 failed, 16 passed

$ git stash pop
$ pytest .../test_classify_pod_failure.py -q -p no:randomly
24 passed

Live watch events, kind v1.35.0

Node drain, via kubectl taint node <node> aip97=evict:NoExecute:

  12.4s  type=MODIFIED  phase=Running  deletionTS=-    status.reason=None  DisruptionTarget=True/DeletionByTaintManager  container=-
  12.4s  type=MODIFIED  phase=Running  deletionTS=SET  status.reason=None  DisruptionTarget=True/DeletionByTaintManager  container=-
  17.8s  type=MODIFIED  phase=Failed   deletionTS=SET  status.reason=None  DisruptionTarget=True/DeletionByTaintManager  container=term:Error:exit=143
  18.1s  type=DELETED   phase=Failed   deletionTS=SET  status.reason=None  DisruptionTarget=True/DeletionByTaintManager  container=term:Error:exit=143

Scheduler preemption, via a higher-priority pod that cannot otherwise fit:

  12.5s  type=MODIFIED  phase=Running  deletionTS=-    status.reason=None  DisruptionTarget=True/PreemptionByScheduler  container=-
  12.5s  type=MODIFIED  phase=Running  deletionTS=SET  status.reason=None  DisruptionTarget=True/PreemptionByScheduler  container=-
  17.7s  type=MODIFIED  phase=Failed   deletionTS=SET  status.reason=None  DisruptionTarget=True/PreemptionByScheduler  container=term:Error:exit=143
  18.2s  type=DELETED   phase=Failed   deletionTS=SET  status.reason=None  DisruptionTarget=True/PreemptionByScheduler  container=term:Error:exit=143

status.reason is empty throughout and the container only ever says Error/143, so the condition is the sole signal. The end-to-end test builds this exact object.

ruff and mypy clean.

[1fanwang]

Folded into the consolidated AIP-97 PR as 8ae4f32a3a: apache#66405

The runtime cost is +42/-1 across two files, and the AIP already described a node drain and a preemption as infra, so shipping the reference classifier without the condition read left the text and the code disagreeing. Keeping the branch for the live evidence.