GEOALIGN Geometric Filtering Framework: Solving the instability problem in LLM reinforcement learning training