fix: unify fuzzy_match metrics into shared module

[abrichr]

Summary

• Create evaluation/metrics.py as single source of truth for metric functions
• client.py fallback now delegates to shared module (was word-set overlap, now rapidfuzz with fallback)
• evaluate_endpoint.py StandaloneMetrics now wraps shared module
• Add 15 tests for the shared metrics module

Resolves the eval path divergence documented in client.py and STATUS.md (Tier 2).

What changed

Metric	Before (client.py)	Before (endpoint)	After (both)
fuzzy_match	word-set intersection	rapidfuzz + substring fallback	rapidfuzz + substring fallback
exact_match	actual == expected	actual == expected + strip	actual == expected + strip
contains	case-insensitive substring	case-insensitive substring	case-insensitive substring

Test plan

• All 26 existing evaluate_endpoint tests pass
• All 15 new metrics tests pass
• 507 total tests pass (22 pre-existing failures in test_setup_handlers.py)

🤖 Generated with Claude Code