Harden vLLM prompt-cache replay execution

[franciscojavierarceo]

Parent tracker: #69
Design context: #65

Summary

Harden the vLLM execution path for compact prefix replay using prompt_cache_ref + append_token_ids.

Scope

• Keep vLLM as the authority for rendering, tokenization, prompt-cache handle execution, and generation.
• Harden the prompt_cache_ref + append_token_ids path.
• Decide where marginal rendering lives: incremental Responses render API in vLLM, cached prefix IDs plus marginal messages in vLLM, or an agentic-api path that calls the same renderer/tokenizer source of truth.
• Scope prompt_cache_ref to the selected serving endpoint or cache epoch.
• Add handle miss fallback, reseeding, and restart behavior.
• Avoid returning full prompt_token_ids on every production turn.

Acceptance criteria

• Handle replay either executes safely or falls back to a full render path.
• vLLM restart or handle miss behavior is deterministic and tested.
• The production path can avoid full prompt-token arrays in steady state.
• Marginal rendering ownership is documented and implemented behind an explicit interface.