fix(builder): replace lifecycle handlers instead of merging them

[giuliocalzo]

Summary

A NodeSet can already override the slurmd preStop command, but only while the handler stays an exec one. BuildContainer merges the handler from spec.slurmd.lifecycle into the operator default with a strategic merge patch, so switching the action to httpGet, tcpSocket, or sleep leaves the default exec in place alongside it. A lifecycle handler may specify only one action, so Kubernetes rejects the result.

A merged lifecycle handler now displaces the base handler rather than being merged into it, the way BuildContainer already treats probe handlers. The default slurmd preStop also moves out of an inline literal into a named function, so what a NodeSet replaces is legible at the call site.

Both new cases fail against main and pass with the fix: TestBuilder_BuildContainer/preStop_httpGet_replaces_exec and TestWorkerBuilder_slurmdContainerPreStop/httpGet_handler_replaces_the_default_exec.

Checklist

• I have read the
  CONTRIBUTING.md
  and the
  Code of Conduct.
• New or existing tests cover these changes (where applicable).
• Documentation is updated if user-visible behavior changes.

Breaking Changes

None. A NodeSet that does not set spec.slurmd.lifecycle renders an identical container.

Testing Notes

Deploy a NodeSet whose preStop uses a different action than the default exec:

spec:
  slurmd:
    lifecycle:
      preStop:
        httpGet:
          path: /drain
          port: slurmd

On main the API server rejects the pod, because the rendered handler carries both httpGet and the operator's exec. With this change the pod is admitted and carries only the httpGet handler:

kubectl get pod <nodeset-pod> -o jsonpath='{.spec.containers[?(@.name=="slurmd")].lifecycle.preStop}'

A NodeSet with no lifecycle should still show the default state=down hook.

Additional Context

Also submitted upstream on GitLab as slurm-operator!724, where the maintainers do their review; I will close whichever of the two you prefer not to keep.