Diabetic retinopathy (DR) is a leading cause of vision loss worldwide, yet automated detection systems are hindered by severe class imbalance and limited availability of annotated fundus images, particularly for advanced disease stages. Generative models offer potential for data augmentation, but existing methods often produce anatomically implausible images that violate pathophysiological constraints. This work introduces C2-HDM (Clinical-Constrained Hierarchical Diffusion Model), a generative framework that enforces anatomical validity through vessel topology guidance and clinical constraint losses. Retinal vessel graphs are extracted using morphological skeletonization, encoded with graph neural networks, and used to condition a hierarchical diffusion model on vessel structure, DR severity, and diffusion timesteps. Three clinical constraint losses are proposed: vessel-lesion correlation loss to localize lesions near vascular damage, anatomical structure loss to preserve fundus-specific color distributions, and severity progression loss to maintain disease-appropriate characteristics. On the EyePACS dataset (35,000 images), C2-HDM achieves a Fréchet Inception Distance (FID) of 261.0, outperforming StyleGAN2 (365.3) and vanilla DDPM (300.1). When applied for data augmentation, synthetic images increase F1-score for minority classes by 12.3% (severe DR: 0.78 \(\rightarrow \) 0.87). Ablation studies confirm the essential role of both graph conditioning and clinical losses. These results demonstrate that integrating anatomical and clinical constraints into generative models produces clinically valid synthetic images capable of mitigating data scarcity in medical AI systems.