D860 — A defect that escapes review becomes an eval case for the reviewers

Decision. A fix: PR that repairs a regression names the PR that introduced it, Regression-of: #<pr>. That PR becomes a row of .agents/skills/review-pr/evals/cases.md, and a change to a reviewer definition, its model pin or review-pr §3-§5 is scored on the rows (and on clean PRs, for noise) before it lands; the score goes in evals/scores.md and in the PR's ## Self-review. A case is caught when a high or medium finding names the defect's file and its class. Recall comes first, noise second.

Why. Reviewer definitions and model pins change, and nothing said whether they still catch the bugs that matter; a defect that escaped left no trace that could be learned from. Every miss now gets one question, "what did the reviewer not have in front of it?", answered where it bites (D214).

Alternatives rejected. Running the evals in CI: D856 rejected a review bot for its cost, and a run is N reviewer spawns. A script that spawns the reviewers: a subagent cannot spawn one, so the coordinator drives and the script only prepares worktrees and mandates. Linking fixes to origins by git blame at fix time only: it is a one-off harvest (used for the seed), not a link the next reader can follow.

Consequences. pr-reviewer and security-reviewer take an optional base (default origin/main) because a case is a past PR. The seed set is small and dated; it is extended by escapes, not curated. A score is a count over a handful of cases: read it as a regression alarm, not a measurement.