Benchmark cached-prefix replay end to end

[franciscojavierarceo]

Parent tracker: #69
Design context: #65

Summary

Build the benchmark coverage needed to validate cached-prefix replay against realistic long-context Responses and Conversation workloads.

Scope

• Re-run benchmarks with APC disabled, cold prefixes, non-Harmony Responses models, and real agentic traffic profiles.
• Add server-side render/tokenize timing instrumentation.
• Add an agentic-api end-to-end benchmark that sends prompt_cache_ref + append_token_ids.
• Measure storage lookup/persistence overhead separately from vLLM TTFT.
• Compare short-context, long-context APC-hot, and APC-unstable regimes.

Acceptance criteria

• Benchmark output includes TTFT, total latency, prompt tokens, generated tokens, cached-token counts, and fallback reasons.
• Long APC-hot agentic loops show lower TTFT without changing model-visible token IDs.
• Neutral or negative short-context results are explicitly reported rather than hidden.
• Results can be attached back to ADR-04 or a follow-up report.