Add an inter-annotator-agreement scorer with Cohen and Fleiss kappa #934

Summary

When multiple annotators label the same corpus, OpenMed cannot quantify agreement, so the quality of golden fixtures is unverified. Inter-annotator agreement (kappa) is the standard reliability check for span-annotation tasks.

Scope

• Compute span-level Cohen kappa (two annotators) and Fleiss kappa (multiple annotators) over aligned annotations
• Add label-level and overall agreement breakdowns with disagreement examples by offset
• Emit an agreement report consumable by the active-learning queue

Acceptance criteria

• Kappa values match hand-computed reference results on a synthetic two- and three-annotator fixture
• Disagreement listings reference offsets and labels only, no raw PHI
• Perfect-agreement and total-disagreement edge cases return 1.0 and the expected low values

Out of scope

• Adjudication workflow (separate active-learning concern)
• Token-level alignment heuristics beyond exact and overlap matching

Files

• openmed/eval/annotation/agreement.py
• tests/unit/eval/test_agreement.py

Task: OM-577 · Milestone: v2.2 · Priority: P2 · Size: M
Depends on: OM-556 · Blocks: —
Roadmap: v2.2 trustworthy clinical data exchange follow-on wave
Spec: PLANS/V2/EXECUTION/tasks/OM-577.md