Compare before/after Coval runs using matched cases, stable metric versions and inspected conversation evidence. Use to assess a regression or proposed improvement; does not automatically launch or tune anything.
