feat: assessment remediation — safety defaults, NER/OCR improvements, audit logging, compliance metadata

[SPerekrestova]

Summary

• Phase 1 (Safety): Replace zero-shot severity classification with deterministic template parsing for DrugBank interactions (severity_parser.py); unknown severity now defaults to major; confidence gating added; uncertain field propagated to API responses
• Phase 2 (NER/OCR): Swap NER model to PharmaDetect-BioPatient-108M; add OCR text preprocessor (ocr_cleaner.py) applied before NER; needs_confirmation flag added for low-confidence drug identifications
• Phase 3 (MCP server): Build-time severity parsing added to build-db.js (mirrors Python logic); VERSION file generated at build; severity field passed through drugbank-api.js and Python client. Note: This is a forward-looking optimization — the DB is built by the upstream CI pipeline at openpharma-org/drugbank-mcp-server which requires a DrugBank API key. Until that pipeline rebuilds, the severity field will be null in the DB, and the Python runtime fallback (template parser → classifier) handles classification transparently.
• Phase 4 (Observability): Structured audit logging middleware with contextvars (audit_log.py); wired into drug_analyzer.py and interaction_checker.py; cache TTL reduced to 4h; admin /admin/cache/clear endpoint added
• Phase 5 (Compliance): data_sources version stamps on all API responses; limitations field with scope disclaimer on interaction responses

Test Plan

• 148 tests passing (pytest tests/ -q)
• NER model OpenMed/OpenMed-NER-PharmaDetect-BioPatient-108M verified accessible and outputs B-CHEM/I-CHEM labels (compatible with existing filter)
• Deploy and smoke-test /interactions response for uncertain, data_sources, and limitations fields
• Verify audit log JSON appears in application logs on a real request

🤖 Generated with Claude Code