A central bottleneck in unsupervised cross-modal hashing is that both similarity guidance and objective weighting are typically static, while the semantic structure induced during training is inherently dynamic. We present DSSL-Hash, which turns semantic structure learning into an evolving process rather than a one-shot preprocessing step. Specifically, DSSL-Hash distills semantic cues from a vision–language foundation model and constructs a dynamic perception graph via an adaptive similarity matrix, enabling graph propagation to capture higher-order relations beyond pairwise matching. We further reshape Hamming-space learning with a semantic-channel constraint that explicitly regulates distances for partially similar pairs, reducing semantic distortion after binarization. Finally, we develop a Retrieval-Aware Adaptive Scheduler (RAAS) that leverages retrieval feedback to co-adjust modality interactions and objective weights, achieving robust optimization without extensive manual hyper-parameter search. Extensive experiments on MS COCO, NUS-WIDE, and MIRFLICKR-25K across multiple code lengths demonstrate that DSSL-Hash consistently outperforms recent state-of-the-art unsupervised cross-modal hashing methods.