<p>Zero-shot stance detection aims to train models on labeled data of known targets and enable them to classify stances on unknown targets without any labeled samples in the target domain. However, existing approaches often struggle to generalize effectively due to limited labeled data, weak utilization of sentiment cues, and insufficient cross-domain alignment. To address these limitations, this paper proposes a novel zero-shot stance detection framework that integrates sentiment knowledge into a joint contrastive learning architecture. Unlike prior methods that treat stance and sentiment independently, our approach explicitly models their interaction through a dual dynamic target-aware graph and a dual-channel contrastive learning mechanism that aligns semantic and sentiment representation spaces, enabling more robust transfer of stance semantics to unseen targets. Experimental results on the SEM16 and VAST datasets show that our model achieves strong performance and superior robustness in stance-sentiment conflict scenarios. This work provides a new and effective pathway for improving stance generalization in zero-shot settings.</p>