Rice leaf disease detection systems trained on controlled laboratory datasets often fail when deployed in diverse agricultural environments due to distributional shifts caused by variations in lighting, camera specifications, and background complexity. This paper addresses the critical gap between laboratory performance and field reliability through systematic multisource evaluation of deep learning architectures. We propose a Hybrid RiceLeafCNN that combines depthwise-separable convolutions for local texture analysis with EfficientNet-B0 pathways for global contextual modeling. Evaluated across three dataset configurations representing single-, dual-, and multi-source heterogeneity, the hybrid model achieves the most stable performance profile with only 1.79 percentage point variation (97.47% single-source, 95.68% multi-source), compared to ResNet50's 1.59 points but with 66% fewer parameters (7.97M vs. 23.5M). The architecture maintains about 96% average accuracy across heterogeneous conditions while achieving 97 frames per second on the reported NVIDIA T4 GPU environment, indicating real-time inference potential; mobile-device deployment requires dedicated latency and energy validation. This work contributes both a specialized architecture for agricultural disease detection and an evaluation methodology emphasizing cross-domain robustness over peak single-dataset accuracy.