Remove one decoder layer at a time — an early cliff, not a U

Teacher-forced on each prompt's own greedy continuation. Only L0–L1 are load-bearing; every layer from L2 on is remarkably survivable. Toggle the metric and the sample size.