M5: implement RAGcore knowledge integration

KnowledgeProvider protocol with a deterministic TF-IDF-weighted extractive demo provider (never generative, always cites real excerpts) and a RAGcore HTTP adapter that degrades cleanly to unavailable. Knowledge nav + chat-style Q&A UI with source cards and honest grounded/insufficient/unavailable states. 57 backend tests passing, ruff clean. Fixed a real relevance bug (generic terms like "vehicle" crowding out distinctive matches) via IDF weighting, found by testing the actual S6 scenario. Verified end-to-end in the browser: grounded damage question cites both expected procedures; unrelated question honestly returns insufficient evidence with no fabrication.

backend/Dockerfile

@@ -9,6 +9,7 @@ COPY backend/alembic ./alembic
 COPY backend/alembic.ini ./
 COPY backend/tests ./tests
 COPY seed ./seed
+COPY knowledge ./knowledge
 RUN pip install --no-cache-dir --no-deps -e .
 EXPOSE 8000
 CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]