We present our third-place submission to the Nakba Image Classification Shared Task at LREC-COLING 2026, which requires binary classification of Instagram images from Gaza into destruction (damaged or destroyed infrastructure) versus not_destruction. Our system fine-tunes a ConvNeXt-Tiny backbone within a five-fold stratified cross-validation framework, combining Focal Loss, weighted random sampling, exponential moving average (EMA) weight stabilization, test-time augmentation (TTA), and out-of-fold (OOF) decision threshold calibration. Our system achieves an official test macro F1 of 0.8893 and 90.05% accuracy, placing third among all participants and within 0.02 F1 of the winning system (0.91), demonstrating that a 28M-parameter convolutional architecture with principled training strategies is highly competitive with much larger models.