Add a clinical-domain label-map coverage eval suite over the synthetic fixtures #866

Summary

As clinical sub-domain label maps and synthetic fixtures multiply, there is no single eval that verifies every shipped domain has a matching fixture, every fixture label resolves to a catalog label, and per-label span coverage is non-empty. Without this gate, domain maps and fixtures drift apart silently.

Scope

• Add eval/domain_coverage.py that enumerates label_maps/defaults.json domains, loads each tests/fixtures/clinical/.jsonl, and reports per-domain/per-label span coverage
• Cross-check every fixture label against a catalog label in core/labels.py and flag orphans
• Emit a leakage-first-style coverage report (offsets/labels only, no raw PHI) and a machine-readable summary
• Wire the suite into the existing benchmark CLI and add a CI gate that fails on missing-fixture or orphan-label

Acceptance criteria

• Running the suite over all current clinical domains passes offline with zero orphan labels
• Adding a domain without a fixture, or a fixture label not in the catalog, fails the gate with a clear message
• Report contains only labels/offsets/counts, never raw fixture text in logs
• Suite registered in the eval harness and discoverable via the eval CLI

Out of scope

• No F1/accuracy scoring against a gold model (coverage and consistency only)
• No restricted datasets; operates on committed synthetic fixtures
• No new model training

Files

• openmed/eval/domain_coverage.py
• tests/unit/eval/test_domain_coverage.py
• .github/workflows/eval.yml
• docs/clinical-domains.md

Task: OM-508 · Milestone: v2.2 · Priority: P1 · Size: M
Depends on: OM-018, OM-176, OM-412 · Blocks: —
Roadmap: v2.2 trustworthy clinical data exchange follow-on wave
Spec: PLANS/V2/EXECUTION/tasks/OM-508.md