Run Codex Adversarial Review Loop

Orchestrate a bounded, self-verifying quality loop over a codebase using Codex as a fan-out adversarial reviewer. The loop is: explore the project → derive review lenses from what was found → fan out parallel Codex reviews → collect → independently verify every finding → re-evaluate → fix verified issues in grouped worktree subagents → merge → build-check → loop with fresh reviewers until a wave returns only noise.

Two ideas make this reliable and are non-negotiable:

1. Reviewers are derived, not dictated. Explorer subagents learn the actual project first; the review lenses come from that map. Do not hardcode "check X" — let the exploration decide what surfaces exist and matter.
2. Every Codex finding is independently verified before any fix. Codex under a "find something" framing sometimes misreads intentional code or invents issues. A separate verifier confirms/refutes each claim against the code — and is asked to report anything more it sees — before a fixer touches anything.