The widespread occurrence of missing data in traffic sensor networks critically undermines the performance of intelligent transportation systems. Existing data-driven models often fail to accurately capture dynamic spatiotemporal dependencies, particularly in highly heterogeneous road environments. To address this issue, this paper presents a diffusion-based framework with physically consistent priors for modeling spatiotemporal traffic data (DCDM), which integrates physical constraints to improve the robustness and interpretability of the imputation and forecasting processes. Specifically, we develop a spatio-temporal feature extractor grounded in the infiltration principle of Fick's law to model directional flow dynamics between adjacent road nodes. To mitigate the impact of missing data at critical nodes, we introduce a global information compensation mechanism that enhances the denoising process by capturing long range spatiotemporal dependencies. The proposed model jointly optimizes data imputation and prediction tasks within a unified diffusion framework. Extensive experiments on three real-world datasets demonstrate that DCDM consistently outperforms state-of-theart methods under various missing data conditions, achieving superior reconstruction accuracy and predictive stability. Moreover, statistical tests and visualization results further confirm the model's robustness, interpretability, and strong generalization ability in generating high-quality data that closely matches the true distribution.