kserve: the unload whose LLMInferenceService delete did not take effect #281

Problem

On an AWS test installation an unload_model for a kserve-served model answered loaded: false while the LLMInferenceService stayed Ready and kept serving (#242). The symptom is fixed in v1.14.1-rc.1 (#277): Stop reads the object back after its delete and answers conflict naming it when the delete did not take effect. Why the delete did not take effect is unknown: no error surfaced, the object carried no deletionTimestamp afterwards, and an earlier unload of the same model on the same pool had removed it. The case was not reproduced.

Proposed solution

This issue holds the symptom until the next occurrence, which rc.1 makes diagnosable: the conflict answer names the object, and whoever meets it records here the LLMInferenceService's state at that moment (finalizers, deletionTimestamp, owner references) and the API server's audit trail of the DELETE, with any webhook or controller that acted on the object. With that the root cause is fixed, or the issue is closed as not reproduced.

Acceptance criteria

• The next conflict from an unload is recorded here with the object's state and the delete's audit trail.
• The root cause is named and fixed, or the issue is closed as not reproduced after 30 days without a new occurrence.

Written by an agent.