Inject failure on demand.

Each switch activates one known regression. Next run captures exact failure evidence.

Judge mode: run clean once for baseline, inject one regression, then verify again and open Compare runs.