Fail closed on untrusted AI signal provenance

[Pigbibi]

Summary

• require ResearchSignalContextPipelines manifest v2 before AI context can affect advisory scoring
• verify the signal and sibling manifest against the current upstream checkout HEAD blobs and expected origin repository
• prevent untrusted AI inputs from expanding the market-confirmation universe
• return one sanitized ai_signal_provenance_untrusted warning for missing, legacy, unknown, or mismatched provenance

Boundaries

• no workflow changes
• no GitHub API, provider, credential, deployment, or execution calls
• no new dependencies or registry

Validation

• RED first: 7 provenance/universe tests failed before implementation
• python3 -m pytest -q tests/test_ai_ingestion_v6.py: 39 passed
• python3 -m pytest -q tests/test_build_pipeline.py: 7 passed
• focused Ruff, compileall, and diff check passed
• full suite: 361 passed, 5 failed because existing tests outside the authorized write set still expect manifest-less/v1 AI inputs to affect scoring; production compatibility was intentionally not added because it would violate fail-closed

Migration note

CI remains blocked until those stale expectations are migrated to trusted manifest-v2 fixtures or no-op assertions in a separately authorized test-only follow-up.