RL for LLM Reasoning Is Sparse Policy Selection, Not Capability Learning