Retraining policy for forward test: one-shot only (no retraining inside the 20% window). Freeze model, hyperparameters, and thresholds before touching the 20% block. Run a single final evaluation on ...