Cache-locality-aware worker claim + reap

[benben]

Summary

Both worker scheduling paths in `K8sWorkerPool` iterate `p.workers` map in Go's randomized order, so claims and reaps are non-deterministic. That churns nodes: a new query may land on a cold-cache node while an idle reap may kill the oldest, warmest one — defeating the NVMe cache proxy we just rolled out.

Fix: track the first time this CP saw a worker on each node (`nodeFirstSeen map[string]time.Time`, stamped when a pod reports Ready with `Spec.NodeName`) and use it to order both paths:

• `findIdleWorkerLocked` — prefer the idle worker on the oldest-seen node so sessions land where the NVMe cache is warmest.
• `reapIdleWorkers` — sort reap candidates by nodeFirstSeen desc so excess idle workers are evicted from the newest-seen nodes first, letting Karpenter consolidate those and keeping old warm nodes alive longer.

Workers missing a nodeName (informer race on the first few ms after spawn) sort as "new" — reaped first, claimed last — which is the safe direction.

Design choices

• Per-CP state, no coordination. Each CP only decides about workers it owns. The shared store already prevents over-spawn; the local `idleCount > minWorkers` gate already prevents over-reap. Cross-CP "oldest node absolutely" would need a shared store round-trip or a cluster-scoped `Node.Get` (extra RBAC) for marginal benefit.
• No new k8s API calls, no new RBAC. nodeName comes from the pod the informer already watches; no `Nodes().Get` needed.
• Map is pruned when the last worker on a given nodeName retires (O(N) scan inside pool mutex, N = worker count per CP, small).

Test plan

• 8 new unit tests in `k8s_pool_test.go` covering:
  • `stampNodeFirstSeen` idempotency + empty-node guard
  • `pruneNodeFirstSeen` only drops entries with no remaining worker
  • claim prefers oldest node, falls back cleanly when nodeName is empty
  • reap picks newest-node first
  • reap stops at minWorkers
  • reap skips workers still within idleTimeout
• Full `go test -tags kubernetes ./controlplane/` passes
• After merge: watch `duckgres_org_workers_active` by node and Karpenter node churn — expect fewer long-lived nodes getting recycled, more newest nodes evicted.

🤖 Generated with Claude Code