Kubernetes PDB unhealthyPodEvictionPolicy Prompt

You are a senior reliability engineer whose node drain has been stuck for 20 minutes because a PodDisruptionBudget refuses to let an already-crashing pod be evicted. I want a fix and a policy that doesn't recreate this.

I will provide:
- The PDB spec (minAvailable / maxUnavailable, selector)
- The Deployment/StatefulSet it guards (replicas, current ready count)
- The drain or eviction error and which pods are unhealthy

Your job:

1. **Explain the deadlock**: with the default `IfHealthyBudget` policy, the eviction API only evicts a pod if doing so keeps the budget satisfied. If some pods are already not-ready, the budget is already at its limit, so even evicting a BROKEN pod is refused — and the drain hangs forever.
2. **Introduce `unhealthyPodEvictionPolicy`**:
   - `IfHealthyBudget` (default) — only evict unhealthy pods if the budget is currently met
   - `AlwaysAllow` — always permit eviction of pods that are not Ready, regardless of budget
   Explain when `AlwaysAllow` is correct (a broken pod isn't serving anyway, so evicting it can only help) and when it's risky.
3. **Check the budget math**: a PDB with `minAvailable` equal to (or above) the replica count makes EVERY voluntary eviction impossible; show the corrected value that allows one disruption at a time.
4. **Diagnose the specific stuck drain**: map ready/unhealthy counts to why the eviction API is returning 429/`Cannot evict`, and the precise change that unblocks it.
5. **Produce the fixed PDB** with `unhealthyPodEvictionPolicy` set and corrected budget, plus how to verify with `kubectl get pdb` (ALLOWED DISRUPTIONS column) and a dry-run eviction.
6. **Mark anything** that lowers availability guarantees, and confirm the workload can tolerate the disruptions the new budget permits.

Output format: deadlock explanation, the budget-math fix, the corrected PDB YAML, and verification commands. Do not drain anything — give me commands to inspect and the change to apply myself.

---

PDB spec:
```yaml
[PASTE]
```
Guarded workload (replicas, ready count): [DESCRIBE]
Drain/eviction error + unhealthy pods: [DESCRIBE]