Automated EEG-based seizure detection shows strong promise, with deep learning models often reporting validation accuracies above 95%. However, such results typically rely on controlled datasets, limiting clinical reliability. This creates a critical evaluation-deployment gap: high validation performance does not ensure effective detection when seizures are extremely rare. We address this through dual-protocol evaluation contrasting balanced validation with deployment-realistic testing under natural class distributions. A CNN-Attention-BiLSTM baseline achieves 98.7% accuracy under controlled validation. We propose an Enhanced Multiscale Fusion CNN (EMF-CNN) with four parallel convolutional branches spanning 31-323 ms receptive fields, integrating time- and frequency-domain representations through dual-stream design and attention-based fusion. Under controlled validation, EMF-CNN achieves 96.10% accuracy, 0.9867 ROC-AUC, and 94.66% recall. Deploymentrealistic testing reveals sharp collapse to 3% seizure recall, exposing a 91% evaluation-deployment gap. Ablation and SHAP analyses confirm the contribution of ultra-large temporal kernels and neurophysiologically meaningful attention. These results demonstrate that architecture-level improvements alone do not solve deployment-level class imbalance, and that accuracy-centric validation is insufficient for clinical translation. Although the implementation supports real-time inference (10-20 ms per segment), clinical readiness requires substantial recall recovery through cost-sensitive learning, calibration, and