exp: SAK_DUP_CAP + fixed NanGuard (no-labels reforward); sync-tested

exp: decouple trainer seed from data seed; trainer seed back to 42 (the NaN cause, not LR)

exp: SAK_RANK/SAK_DROPOUT/SAK_MAXLEN controls to attack argument accuracy

exp: SAK_LR override; 4e-4 diverges to NaN at 6 epochs

exp script: bench-v2 exclusions, SAK_SEED/SAK_EPOCHS/SAK_TAG controls

Add 0.5B experiment script (masking / full-FT modes); eval moves to the bench repo runner