Diff a candidate version against a baseline over the last N sealed runs (zero side effects)

Pre-publish confidence. Replays both versions over the SAME frozen input packages of the app's most recent sealed runs and reports which subjects change verdict and how each rule's pass/fail footprint moves. Reads only: nothing is persisted, no actions fire, and the candidate is never activated. The LLM adjudicator is withheld so the diff is free and repeatable. Supply the candidate as inline content or candidateVersionId; with neither, the app's current draft is used.