Accurate and trustworthy disease diagnosis from field imagery requires a framework that balances predictive accuracy with calibrated confidence and auditable reasoning. This work benchmarks a diagnostic system coupling an attention-augmented convolutional network (ResNet-34+CBAM) with post-hoc probability calibration and a rule-aware validator. Agronomic symptom rules are encoded in a lightweight RDF/OWL knowledge graph, enabling a post-hoc check that links model predictions to human-readable explanations for auditability. On a rigorously de-duplicated test split of the public PaddyDoctor corpus, the model achieves 95.13 accuracy (weighted F1: 95.14) with a median latency of 4.6 ms. We analyze the trade-offs of post-hoc calibration: Temperature Scaling, fit on the calibration split (ECE: 1.65→1.35), improves the test-set Brier score (0.0760→0.0758) and NLL (0.1573→0.1566) but results in a slight increase in test-set ECE (0.82→0.94). A robustness analysis using common corruptions identifies critical failure modes: while resilient to JPEG compression (86.15 accuracy at severity 5), the model is highly vulnerable to brightness shifts (47.72) and Gaussian blur (32.13), highlighting the need for domain-specific augmentations. The resulting system provides a comprehensive baseline for combining strong predictive performance with post-hoc calibration and auditable explanations, supporting transparent triage in practical deployments.