Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It