feat: add baseline latency/cost regression caps to release gate

[brandonwise]

Summary

Adds optional baseline drift guardrails so releases fail when average latency or cost regresses too far versus baseline, not just pass rate.

Why

Agent teams keep reporting quality can look stable while latency/cost quietly degrade. This closes that gap and keeps release gates aligned with production economics.

What changed

• Added max_avg_latency_regression_pct and max_avg_cost_regression_pct global spec options
• Enforced failures when current run exceeds configured baseline drift thresholds
• Added explicit failure reasons for missing/invalid baseline telemetry
• Added evaluator tests for latency drift, cost drift, and telemetry-missing paths
• Updated README + example spec with new config options

Validation evidence

• python3 -m pytest → 10 passed
• python3 -m pytest tests/test_evaluator.py -k baseline → 3 passed, 6 deselected
• python3 -m ruff check . → All checks passed
• python3 -m mypy src → Success: no issues found in 4 source files
• python3 -m build → Built sdist + wheel successfully
• Smoke/integration check (expected fail path):
  • Command: PYTHONPATH=src python3 -m agent_release_gate.cli evaluate --spec <tmp>/spec.yaml --results <tmp>/results.json --baseline <tmp>/baseline.json
  • Result: exit_code=1, gate_passed=false, reason includes Latency regression detected: current 1300.0ms vs baseline 1000.0ms (30.00% increase, allowed 10.00%)

Docs tone check

• node src/cli.js score -f /Users/bwise/.openclaw/workspace/projects/agent-release-gate/README.md (run from projects/humanizer) → 🟢 0/100