test: gate operational self-knowledge contradictions - #23

Summary

• derive operational expected facts from /health or an independent fixture instead of encoding one model name
• evaluate 24 Chinese/English, typo, follow-up, misleading-premise, and stack-collaboration questions
• fail on missing required facts, forbidden runtime claims, used_model mismatch, missing Support, or wrong decision route
• record trace, stage timings, citation coverage, and contradiction score in a JSON artifact
• make the semantic gate part of hosted deployment verification
• expose structured speculative non-activation reason and platform-aware SAGE/vLLM-HUST/backend roles

Validation

• uv run ruff check src tests tools/validate_operational_self_knowledge.py tools/verify_hosted_web_deploy.py
• uv run pytest -q — 721 passed, 28 skipped
• focused operational/runtime/receipt suite — 78 passed
• isolated HTTP gate — 24/24 passed, reference coverage 1.0, max contradiction score 0.0

Closes #12