feat: add global p95 latency guardrail

[brandonwise]

Summary

Adds a global max_p95_latency_ms guardrail so release decisions can fail on tail-latency regressions, not just average latency drift.

Why

Recent telemetry keeps highlighting production agent pain around slow long-tail responses and regression detection. Average latency can look healthy while p95 degrades.

What changed

• Added global.max_p95_latency_ms to gate spec parsing (GateSpec).
• Added p95 latency computation and gate enforcement in evaluator:
  • fail when p95 exceeds configured limit
  • fail when p95 limit is configured but latency telemetry is missing
• Added p95_latency_ms to JSON and markdown report output.
• Updated README + example spec for new config.
• Added evaluator tests for:
  • p95 limit catches tail outlier regressions
  • telemetry-required behavior when p95 guardrail is configured

Validation evidence

All checks run from repo root (/Users/bwise/.openclaw/workspace/projects/agent-release-gate):

• python3 -m pytest → ✅ 19 passed in 0.17s (and rerun after cleanup: 19 passed in 0.04s)
• python3 -m pytest tests/test_evaluator.py → ✅ 11 passed in 0.03s
• python3 -m ruff check . → ✅ All checks passed!
• python3 -m mypy src → ✅ Success: no issues found in 5 source files
• python3 -m build → ✅ built sdist + wheel successfully
• Smoke/integration (new behavior):
  • command: PYTHONPATH=src python3 -m agent_release_gate.cli evaluate --spec <tmp>/spec.yaml --results <tmp>/results.json
  • result: ✅ expected gate failure with exit_code=1, gate_passed=False, p95_latency_ms=2500.0, reason P95 latency 2500.0ms exceeds limit 1000ms

Risk

Low. Additive, opt-in config only; existing specs continue to behave as-is unless max_p95_latency_ms is set.