Accurate patch-level classification of glioblastoma sub-regions is essential for diagnosis but challenged by tumor heterogeneity and extreme class imbalance. We propose an ensemble framework that uses feature embeddings from four foundation models: UNI, Virchow, Gigapath and Midnight. Each embedding is partitioned into fixed-size chunks to preserve morphological detail and enable localized modeling. Dedicated XGBoost classifiers are trained on each chunk using balanced sample weights to mitigate class imbalance. A shared chunk selection mechanism ensures consistency between training and validation sets and prevents data leakage. Chunk-level predictions are fused via simple averaging into a unified probability vector. To enhance detection of rare sub-regions we apply rare-class boosting by scaling their predicted probabilities before re-normalization. Per-class decision thresholds are optimized on the validation set to maximize macro-F1 score improving sensitivity to underrepresented patterns. The pipeline achieves a global-averaged F1 score of 0.76 and a Matthews correlation coefficient of 0.70 demonstrating robustness under domain shift and severe label imbalance. Our approach highlights the value of feature chunking ensemble fusion and adaptive post-processing in integrating foundation models for digital pathology. The method provides a reproducible and effective solution for sub-region analysis in glioblastoma histology.