Boundaries of Stationary Feature Learning: A Minimax Barrier for Scaling Laws