Persist reliable Responses continuation checkpoints

[franciscojavierarceo]

Parent tracker: #69
Design and measured evidence: #65

Summary

Make Responses and Conversation continuation checkpoints reliable enough to carry certified prefix/artifact metadata without promoting incorrect state.

This is the second milestone, after exact llm-d KV-affine routing is available. Correct routing can operate on ordinary rehydrated full requests; durable token artifacts must not be introduced until checkpoint correctness is defined.

Scope

• Persist immutable execution fingerprint, lineage, renderer-safe boundary proof, token-artifact identity, router-receipt identity, status, and expiration with the accepted response.
• Bind identity to model weights, tokenizer, renderer/template/parser configuration, instructions, tools, cache_salt, LoRA, multimodal inputs, and format versions.
• Fail closed for missing, corrupt, unknown, out-of-order, or partially rehydrated certified history.
• Atomically commit logical output and checkpoint metadata; streaming completion must not imply durable promotion before persistence succeeds.
• Add unique conversation ordering plus compare-and-swap or serialization for concurrent turns.
• Define cancellation, retry, and idempotency behavior so duplicate tool effects or checkpoints cannot be promoted.
• Carry ephemeral prefix/artifact state through gateway-tool subturns and promote only the accepted outer result.
• Measure checkpoint lookup, validation, and persistence overhead separately from vLLM latency.

Acceptance criteria

• Cancellation, retry, concurrent turns, corrupt history, and out-of-order items cannot promote an invalid checkpoint or duplicate tool side effects.
• Rehydration either reconstructs the complete certified model-visible history or returns a typed fallback/failure; it does not silently omit unknown items.
• Logical output and continuation metadata become durable atomically.
• Fingerprint changes invalidate reuse without invalidating the logical conversation fallback.
• Tests cover tool-loop intermediate state, streaming disconnects, persistence failure, and conversation races.