Stabilize experience latency admission

[CaspianG]

Summary

• measure WaveMind Core and Experience end-to-end latency in paired, alternating order
• use the median of three samples per held-out task so one shared-runner spike cannot flip the product verdict
• require the paired latency protocol in experience-quality admission and refresh all linked evidence

Evidence

• 60 training trajectories and the frozen 30-task held-out split are unchanged
• experience-quality admission remains 12/12 admitted
• task success: 100% vs 16.7% Core
• repeated observed errors: 0% vs 83.3% Core
• median tool steps: 3 vs 4
• median context: 36 vs 60 tokens
• p95: 19.57 ms vs 33.85 ms Core (-42.2%)

Validation

• python -m ruff check .
• python -m pytest -q -> 882 passed, 5 skipped
• six repeated benchmark runs -> 6/6 pass
• package build and twine check pass
• benchmark artifact audit passes

This does not weaken the existing 20% p95 regression gate. It makes the measurement protocol resistant to a single unrelated shared-runner scheduling spike while preserving end-to-end retrieval, tool, and trajectory-write cost.