RECAP: Transparent Inference-Time Emotion Alignment for Medical Dialogue Systems