feat: metrics parser — connect gates to real test data

[Wool-xing]

Summary

Gate enforcement now reads real test output instead of empty dicts.

Changes

New: runtime/orchestrator/metrics/parser.py

• parse_junit(xml) — extract total/passed/failed/rate from junit XML
• parse_jmeter_jtl(csv) — extract samples/failures/avg_ms/p95_ms from JMeter JTL
• extract_metrics(outcome) — auto-detect format and parse

Modified: runtime/orchestrator/workflows/test_coordinator.py

• _execute_node() — extracts metrics from stdout after each step
• _check_gates() — uses live metrics dict (was TODO stub)

Flow

execute_node() → stdout (junit XML / JMeter JTL)
     ↓
extract_metrics() → {total, passed, rate, avg_ms, p95_ms, ...}
     ↓
_check_gates() → PASS / BLOCK (real data)

Tests

test_metrics_parser.py: 12 passed
Total: 439 passed, 0 failed