The node runs anywhere that reaches the cluster API. It needs one Role in one namespace: Jobs,
pods, pods/log and pods/exec.
The node copies the turn's directories into the pod over pods/exec, and copies them back at
the end. There is no shared disk. Secret values go into the pod the same way, not into the
Job spec (fleet/internal/backend/kubejob/copy.go).
The pod gets no service account token. The pod must reach the control server.
A kubejob node takes only kubejob agents, and the server never runs them.