spike: re-evaluate Pod checkpoint/restore once a workload misses its measured cold-start budget #4089

🤖 Generated by the Agentic Engineer (interactive Claude Code session)

The trigger that the #3420 decision deferred to.

Not actionable until the cold-start measurement exists. devantler-tech/wedding-app#291 must show a workload that still misses its agreed cold-start budget after the cheaper measures from #3418: request holding, image caching and startup work. Also re-check only once prod runs Kubernetes 1.37 or later (#3850 → #3534), because checkpoint/restore does not exist in 1.36.

What to do then

Re-run the #3420 strong-fit gate against the then-current state:

• The feature's maturity in the Kubernetes version prod runs: stop-and-release semantics and workload-controller or HPA integration.
• Whether Talos's containerd implements the CheckpointPod and RestorePod CRI calls, proven by a direct runtime test.
• Checkpoint storage across the node lifecycle of the min-zero overflow pools.
• Security, retention and cleanup of checkpoint artifacts.
• Measured restore time against an image-cached ordinary cold start for the same user path.

Acceptance criteria

• A recorded decision: adopt through follow-up issues, defer again with named triggers, or reject.
• Any adopt decision cites direct runtime evidence and comparable cold-start numbers.

Size: S–M (investigation; no production rollout).