Reliable brain tumor analysis from limited MRI data requires lightweight models that provides both accurate classification and clinically useful localization. Conventional Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) struggle in low-data medical settings and are typically limited to single-task classification. We present Hierarchically Scaled Vision Transformer (HSViT), a lightweight transformer for brain MRI that integrates a PreEncoder for edge enhancement, Cross-Layer Attention Refinement (CLAR), a Progressive Token Filter, and a Tumor Context Module (TCM) for efficient tumor classification and localization. HSViT jointly outputs tumor class labels and Region of Interest (ROI) maps, offering diagnostic accuracy with spatial localization. These outputs can also support patient specific digital twin pipelines for non-invasive neurotherapy planning with Transcranial Magnetic Stimulation (TMS) and Transcranial Focused Ultrasound Stimulation (tFUS). On the FigShare dataset, HSViT achieves 97.83% accuracy with precision 99.11%, recall 98.66%, and F1-score 98.88%, with class-wise F1 of 98.47% (Glioma), 98.14% (Meningioma), and 99.13% (Pituitary). Ablation results shows that CLAR improves cross-layer cue stability, the Progressive Token Filter suppresses non informative regions, and TCM sharpens lesion boundaries. The lightweight footprint and token efficiency makes HSViT practical for deployment on resource constrained platforms. Code, training configurations, deterministic environment files, and pre-trained weights are available at: https://github.com/ChauhanPratik/HSViT.