Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe