regression-gatekeeperlisted

# Regression Gatekeeper No run, no verdict. The golden set exists to be run before shipping — a ship opinion formed without results is the failure this skill exists to prevent. ## Verification gates (defined first; output is blocked until all pass) - **G1 — Run before verdict:** a ship verdict (SHIP/HOLD/INVESTIGATE) appears only alongside the golden-run results that justify it. With no run yet, the output is a run plan and `VERDICT: PENDING` — "the change is small, low risk" is not a verdict input. - **G2 — Verdict rules pre-committed:** SHIP/HOLD/INVESTIGATE criteria are written before results arrive (zero gate regressions; any captured failure passing through = HOLD; drift bounds for INVESTIGATE) — so the results can't be renegotiated into a ship. - **G3 — Per-case results + scope honesty:** results are shown as a per-case table (id, class, baseline, new, delta), never only an aggregate; the plan states what the golden set does NOT cover, and changes introducing new requirements with zero golden coverage get flagged for case collection before they can be gated. ## Steps 1. **Bank the change and the set:** what changed (prompt diff, model version), the golden set's size and class split (pass-class / fail-class), and whether baseline outputs are stored — if not, the plan runs both sides on the same cases. 2. **Write the run plan:** every golden runs on the changed configuration; fail-class cases must still catch their failure (the captured incident's assertion holds); p