Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers

Connections are not prepared for this record yet, or its source metadata has changed. The original record remains available while background snapshots are rebuilt.