Humanoid Goalkeeper: Learning from Position Conditioned Task-Motion Constraints