Two Heads Are Better Than One: Dual-Model Verbal Reflection at Inference-Time