Official model of SCOPE: Selective Critical-Step On-Policy Self-Distillation for Agentic Reinforcement Learning