v1.21.0 — /simplify recommendation; model mismatch ends auto-resume

Follow-up to v1.20.0 from a week of dogfooding in two consumer repos. The review ratchet v1.20.0 targeted did not recur: no run hit the review cap, and there were no wording-only fix rounds and no rollback/recovery Fixes. Two gaps showed up instead.

What changed

• /simplify recommendation after hyper-implement-loop. Codex's "remove this" / "simplify this" findings usually arrive as Minor, so they are non-blocking and were never applied. The final report now lists them apart from the other residual findings and recommends Claude Code's built-in /simplify main...HEAD. When there is a review brief, its requirements are named as what /simplify must not change. hyper-auto relays the recommendation. It is never run inside the loop: simplification stays a human call, and its edits would land after the last review.
• Fix: --resume auto no longer resumes an unrelated older review. When the newest matching review had run under a different effective model, discovery skipped it and resumed an older artifact instead. For code-review --base main that older artifact was the previous feature's thread. Now a model mismatch ends the search and the run starts fresh, the same way a Claude reviewer-seat artifact breaks the chain. Explicit --resume <path> is unchanged.