GPU-accelerated PPO (4,096 parallel envs) for Snake using CoordConv + zero-downsampling CNN. Trained on single T4.