Accurate segmentation of dental caries on panoramic radiographs is challenging due to subtle lesion appearance, overlapping anatomy, and domain variability across imaging sites. While task-specific models such as CariesNet and U-Net3+ have been proposed, their generalization to unseen data is rarely examined. We present a controlled, multi-site benchmark for caries segmentation on single-tooth images from panoramic radiographs and compare specialized networks to DINOv3, a vision foundation model. Using an expert-annotated dataset of single-tooth images from the University Medical Center Schleswig-Holstein (UKSH), we evaluate model generalization on an out-of-distribution (OOD) test set from the Federal University of Bahia (UFBA). All modelswere trained under identical conditions with standardized preprocessing, loss functions, and hyperparameter optimization. Two DINOv3 variants, one with a linear head and one with a U-Net head, were compared to CariesNet and U-Net3+. While CariesNet reached the highest in-distribution (ID) Dice score on carious images (0.4503), it dropped notablyOOD(0.2656). DINOv3 with a U-Net head showed more stable performance (0.3724 ID vs. 0.3505 OOD), suggesting stronger robustness to domain shifts. Overall, pretrained foundation models demonstrated competitive and more consistent results without dental specific pretraining. The proposed benchmark provides a reproducible framework for future evaluation of segmentation models in dental radiography.