Persist and validate prompt-prefix state in agentic-api

[franciscojavierarceo]

Parent tracker: #69
Design context: #65

Summary

Persist enough prompt-token prefix metadata in agentic-api to decide whether a cached prefix can be safely replayed for a Responses or Conversation continuation.

Scope

• Store prefix metadata with the response or conversation checkpoint.
• Include model, tokenizer, renderer/template, effective instructions/tools identity, prefix hash, prefix token count, and safe-boundary proof.
• Add strict-prefix validation against a fresh full render before replay is enabled.
• Replace Harmony-specific boundary checks with renderer/template-specific safe-boundary proof.
• Measure latest-span lookup and span persistence overhead in the state store.

Acceptance criteria

• Replay planning can prove that a cached prefix still matches the active model-visible history.
• Replay is refused when model/tokenizer/renderer/template/instructions/tools fingerprints differ.
• Tests cover strict-prefix validation and unsafe append-boundary rejection.
• Storage lookup/persistence overhead is measured and reported.