Label-free steering: Compressing test-time reinforcement learning into bias-only subspaces