test: stabilize active writer latency gate

[douglasmonsky]

Summary

• replace the 15-sample pseudo-p95 (which was the single maximum) with five measured append-safe refreshes and at least 40 writer timings
• preserve the sustained p95 <=50 ms SLO
• add a separate <=150 ms catastrophic single-lock ceiling
• add semantic tests proving isolated runner noise passes while sustained or catastrophic regressions fail

Evidence

• eight fresh processes: p95 22.254–26.050 ms; max 23.333–38.109 ms
• 12-worker stress: p95 34.930 ms; max 44.688 ms
• Python 3.14 performance suite: 13 passed
• just v: 498 passed
• Ruff, MyPy, Pyright, release checks passed
• GitNexus: LOW risk, zero affected flows
• final reviewer: zero findings

No product ingest code or integrity/freshness semantics changed.