Each cluster has its own kubejob node. Each node takes turns for the agents spawned with
--kind kubejob.
A node finds its cluster from in-cluster config, or else from KUBECONFIG and the current
context (fleet/internal/cli/node.go, newKubejob).
Limit: node.yaml has no kubeconfig or context key at this time. To reach two clusters
from one machine, run two node processes, each with its own --dir and KUBECONFIG.
Not built: a way to choose a cluster for one agent. The scheduler sees only the kind
kubejob, so any kubejob node can take the turn.