Kubernetes adapter: CRD operator with real cluster integration

[jeremymanning]

Description

adapters/kubernetes/src/main.rs (189 lines) has a CRD schema and CLI scaffold but no operator logic. No watch loop, no Pod creation, no result collection.

Requirements

• Implement K8s operator pattern: watch for ClusterDonation CRD changes
• Create Pods in configured namespace when tasks are assigned
• Enforce resource limits (CPU, memory, GPU) per CRD spec
• Collect results from completed Pods
• Clean up Pods on completion/failure
• Handle common K8s failure modes (OOMKill, Eviction, ImagePullBackOff)
• Deploy via Helm chart or kustomize manifests

Success Criteria

• Operator watches ClusterDonation CRD
• Tasks create real Pods with resource limits
• Results collected from Pod logs/volumes
• Cleanup removes completed Pods
• Helm chart or kustomize deployment manifests
• Integration test on real K8s cluster (minikube/kind acceptable for CI)

Testing (Principle V)

• Deploy operator on real K8s cluster
• Apply ClusterDonation CRD → verify capacity registered
• Submit job → verify Pod created → verify result collected
• Test resource limit enforcement (Pod exceeding limits → OOMKill handled)

[jeremymanning]

Partially addressed by spec 004 / PR #59 — not closing. ClusterDonation CRD + kube client scaffold exist. HOWEVER, no integration test deploys the operator to a real (or even minikube) cluster. The watch loop is not exercised end-to-end.

Remaining work tracked here.

[jeremymanning]

Spec 005 progress — PR #61

STATUS: No change beyond spec 004 — Kind-in-CI integration deferred.

K8s adapter code + CRD from spec 004 unchanged. T061 (Kind CI job + CRD apply + operator reconciliation) not done.

See notes/session-2026-04-19-spec-005-kickoff.md.