feat: make AI reviewer metadata-first

[ftchvs]

Summary

Makes the optional Ollama-compatible AI reviewer safer and more measurable:

• Keeps deterministic policy rules as the trusted baseline.
• Makes model review metadata-only by default.
• Adds explicit model_affects_score opt-in for letting valid model findings join policy_hits and affect scoring.
• Treats invalid model JSON/schema responses as ignored runtime metadata, not review hits.
• Passes bounded landing-page snapshot context into the model prompt with explicit untrusted-content boundaries.
• Passes extracted landing-page context into model-only evals as well as hybrid/product scans.
• Adds UI/API/CLI controls and docs for rule-only, metadata-only, and score-impact modes.
• Adds model reviewer usefulness metrics for measuring useful model-added notes vs false review burden.

Validation

• make test → 193 passed
• make eval → 58 examples, no FP/FN notes
• make benchmark → 213 examples, no FP/FN notes
• make pr-preflight → OK
• GitHub CI Python 3.11/3.12 → passed

Notes

This intentionally does not claim legal compliance or platform approval accuracy. The model layer is positioned as a local, optional reviewer assist; deterministic rules remain the default decision path.