Latency telemetry: add evidence-grade p99.9 retention and sufficiency

[proerror77]

Parent

• CEX HFT: define latency measurement and evidence guideline #394

PRD

#350

Change contract

Add a benchmark-grade latency artifact/reporting path that retains enough real observations to report p50, p95, p99, and p99.9 with sample sufficiency, while keeping the small operational health window explicitly separate.

Acceptance criteria

• Reports include count, p50, p95, p99, p99.9, max, exclusions, and capture provenance.
• p99.9 with fewer than 10,000 eligible observations is marked insufficient-sample and cannot pass a gate.
• The 200-sample operational window is not presented as benchmark evidence.
• Histogram ranges/buckets cover the declared venue-stage distributions without silent overflow.
• Deterministic tests cover skewed tails and insufficient samples.

Out of scope

• Universal latency SLO values
• Timestamp capture changes
• Live enablement

Blocked by

#394, #397