Trust Region Masking for Long-Horizon LLM Reinforcement Learning

First non-vacuous monotonic improvement guarantees for long-horizon LLM-RL