Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning | AIChainDay