Why

A GPU fault event is only worth believing when the fault emitter on the pod's own node wrote it. The emitter's credentials only let it record events for its own node, so an event reported from any other node did not come from the emitter watching the pod's GPUs. The executor's failure classifier checked the event reason and the regarding UID, but not the reporting node, and its event watcher did not record the reporting node at all. A consumer that wanted a typed fault for an attached pod event had to parse the free-text message, with none of these checks.

What

• gpufault gains trust.go. FromPodEvent(PodEvent, Pod) returns the typed fault only for a trusted event: a GPUXidError or GPUSXidError reason, recorded against this very pod, and reported from the pod's node. The trade the classifier already made stays and is documented: when the pod is gone and only its name is left, the name match is accepted. A pod with a UID but no node was never scheduled and has no trusted faults. IsFaultReason is exported for callers that filter before logging.
• pluginsCore.ReasonInfo gains GpuFault, set only for an event that passed FromPodEvent.
• The executor's event watcher records the reporting node of each event: reportingInstance, else the deprecated source host.
• classifyGpuFailure uses FromPodEvent for its per-event checks, so a fault reported from another node no longer reclassifies a failure. RelevantToFailure still bounds the search in time.
• The executor sets ReasonInfo.GpuFault on the pod events it attaches to a phase, and toClusterEvents copies it onto ClusterEvent.gpu_fault. The message is never parsed for a fault there.

Tests

• New table in gpufault/trust_test.go covering each rule and each trade, a wrong reason with a valid-looking message, and an SXid.
• plugin_manager_test: a fault reported from another node does not classify, and attached reasons carry a fault only when trusted. event_watcher_test: the reporting node is recorded and survives aggregation.
• go test -race passes for gpufault, pluginmachinery/core and executor/pkg/plugin/k8s, and the lint diff against main is clean. The new toClusterEvents spec is in the controller's Ginkgo suite, which needs envtest binaries and was not run locally.

A downstream consumer that tags faults on pod events switches to FromPodEvent once this lands.

🤖 Generated with Claude Code

• main
  • GPU fault classification: cover the child pods of distributed tasks #7913
    • gpufault: trust a pod's fault event only from its own node 👈