Grape leaf diseases threaten viticulture by reducing yield and quality, while manual scouting remains inconsistent and can drive unnecessary chemical use. Although deep learning enables automated diagnosis, standard models trained on curated datasets are computationally heavy and prone to overconfident errors on out-of-distribution (OOD) vineyard imagery. We present VitiNet, a lightweight, OOD-aware grape leaf disease classifier built on a ResNet18 backbone and optimized through staged fine-tuning, an engineered ‘Other’ class for explicit OOD rejection, and a risk-mitigating dual-threshold deployment protocol. Evaluated on a challenging, OOD-skewed test set, VitiNet achieved 98.2% accuracy and a 0.982 weighted F1-score, with exceptional OOD performance (99.6% precision for the ‘Other’ class). The primary limitation of this model was found to be the confusion between Bacterial_rot and OOD images (Bacterial_rot F1 = 0.807). Grad-CAM analyses showed that the model focuses on symptomatic regions but can be overly sensitive to necrotic textures that resemble Bacterial_rot in non-disease images. To ensure safe and reliable operation, we suggested a dual-threshold policy for deployment: T-disease = 0.96 (chosen to achieve approximately 95% recall for disease detection) and T-other = 0.99, directing ambiguous predictions to human review. VitiNet delivers fast, accurate, and non-invasive decision support for grape growers and provides a robust foundation for future domain adaptation with real-world field data, including eventual on-device deployment for in-field use.