Self-Policy Distillation via Capability-Selective Subspace Projection