Roadmap: improve realtime latency p95 evidence

[ykshv]

Goal

Bring first-audio and barge-in p95 closer to the release targets under a long-running local GPU profile.

Scope

• Separate cached deterministic responses from normal LLM/TTS responses.
• Measure first audio p50/p95 and underruns over a longer soak.
• Review speculative turn-taking thresholds and TTS first-clause budget.
• Track GPU headroom and thermal behavior.

Acceptance

• latency_report.py output clearly separates response classes.
• p95 regressions are visible in CI/manual release evidence.
• Any changed threshold is justified by measured behavior.