The accurate histological classification of diffuse gliomas is challenging due to tumor heterogeneity and severe class imbalance in medical datasets. This paper presents our methodology for the pathology task of the BraTS-Lighthouse 2025 Challenge, which introduces a novel hybrid ensemble approach combining deep learning with classical machine learning. We found extreme class imbalance and used a two-stage data balancing strategy: using StyleGAN2-ADA for image augmentation, followed by random oversampling of the minority classes in the feature space. We then extracted a rich 4096-dimensional feature vector, by concatenating feature embeddings extracted independently from two pre-trained Vision Transformer based foundation models, UNI2-h and Virchow-2. Our classification system consists of two arms: a deep learning arm featuring a Gated Attention Multi-Layer Perceptron (GA-MLP) which contains stacked attention blocks, and a classical arm comprising an ensemble of K-Nearest Neighbors, a calibrated Random Forest, and an XGBoost classifier. The GA-MLP is trained using a composite loss function, comprising of Focal Loss and F1 Loss. The final predictions are generated by a weighted soft-voting mechanism between the two arms. This strategy utilizes the distinct inherent strengths of each arm to improve diagnostic performance and consistency for glioblastoma multiforme, with the overarching objective of enhancing patient outcomes.