Add trained contextual action candidate with browser evaluation and explicit real-web limits

Optimize CPU feature encoding and add bounded batch inference with measured parity

Upload experimental BAIM code, research checkpoints and measured evaluations