Wire the cross-model advisory reviewer + required human approver

[ronimoe]

Context

advisory.py is a stub. Wire a different-model-family reviewer (comment-only, fail-open to non-blocking) plus a required human/reviewer-App approver so an all-agent fleet does not deadlock on require_code_owner_review.

Acceptance

Advisory comments post and never block; a risk-tier path requires a real approval bound to the head SHA (voided on force-push).

Seam

tools/conductor/advisory.py, risktier.py, ruleset required reviews.

[ronimoe]

Implemented (this commit).

advisory.py now has both halves, deliberately separate:

• Cross-model reviewer — CrossModelReviewer(infer, model, family): a different-model-family reviewer, comment-only, fail-open (any error/timeout → non-blocking unavailable), with a decorrelated flag when reviewer family ≠ author family. No LLM-derived field ever reaches a required check.
• Approval gate — approval_satisfied counts an approval only if it is bound to the exact head SHA (a force-push/rebase voids it) AND from a human or designated reviewer-app — an agent can never self-approve. deadlock_guard flags an all-agent CODEOWNERS that would wedge require_code_owner_review. approval_aware_overlay bridges verifier status into the verdict seam: FAIL→red, NEEDS_HUMAN→held until a valid head-bound approval, then green.

Tested end to end on real git: a risk-tier (auth) change goes NEEDS_HUMAN, the batch holds with CI green, and lands only after a head-bound human approval. 10 unit tests; full suite now 130. Remaining for full close: wire the reviewer to a real different-family inference endpoint and the approval source to GitHub reviews / a reviewer-App, and add the reviewer-App to CODEOWNERS in governance-setup (so the deadlock guard passes on a real repo).