Deep learning image classifiers are increasingly deployed to resource constrained edge devices through INT8 post-training quantization (PTQ), including plant disease imaging. Although quantization reduces memory footprint and can improve inference efficiency, the impact on model explainability is not well comprehended. This study introduces explainability drift, the systematic change in saliency-based explanations between full-precision (FP32) models and quantized (INT8) models, and suggests a multi-metric evaluation suite, including top-k Intersection-over-Union, top-k Dice coefficient, and Spearman rank correlation to measure it. We apply four post-hoc explainability methods (Grad-CAM, Grad-CAM++, Integrated Gradients, and LIME) to Paddy Disease Classification data (10 classes, 10,407 rice leaf images) using three architectures of different complexity (EfficientNetV2-S, ResNet50, and MobileNetV3-Large) before and after INT8 PTQ. Our findings reveal a significant drift depending on both architecture and method. All of the models have an average top-k IoU of less than 0.17, indicating low overlap between the FP32 and INT8 top-saliency regions. Notably, results show that the depthwise-separable architecture of MobileNetV3 leads to a total collapse of Grad-CAM in INT8 with spatially uniform attribution maps (Spearman r = 0.000). In addition, we demonstrate that XAI-preserving quantization-aware training (QAT) with a teacher–student CAM consistency loss can improve Spearman correlation by up to 49.2%, although the results vary depending on the architecture. The work, to the best of our knowledge, provides a quantization-induced explainability drift multi-architecture benchmark and suggests practical recommendations on how to implement reliable, explainable AI in edge devices.