feat: add per-case latency/cost guardrails

[brandonwise]

Summary

Adds optional per-case max_latency_ms and max_cost_usd guardrails so a single outlier can fail the gate even when global averages still look healthy.

Why this matters

Current gate checks only global average latency/cost. That can hide catastrophic case-level outliers (exactly the kind of failure teams hit in production with agent workflows).

What shipped

• Added max_latency_ms and max_cost_usd fields to GateCase.
• Enforced per-case limits during scoring:
  • fails case when metric exceeds limit
  • fails case when metric is missing but a limit is configured
• Added evaluator tests for:
  • latency/cost outlier failures
  • missing telemetry failures when case limits are set
• Updated README + example spec with case-level guardrail docs.

Validation

Command	Result
python3 -m pytest	✅ PASS (5 passed)
python3 -m pytest tests/test_evaluator.py -k case_limits -q	✅ PASS (2 passed)
python3 -m compileall src tests	✅ PASS
python3 -m build --sdist --wheel	✅ PASS
PYTHONPATH=src python3 -m agent_release_gate.cli evaluate --spec examples/spec.yaml --results examples/results.json --output /tmp/argate-report.json --markdown /tmp/argate-report.md	✅ PASS (gate_passed=true)
node /Users/bwise/.openclaw/workspace/projects/humanizer/src/cli.js score README.md	✅ PASS (13/100, mostly human-sounding)

Risk / rollback

Low risk, additive schema change only. Rollback with:

• Revert commit f16b25b