This article presents a Hierarchical Pathway-Masked Attention Autoencoder (H-PAAE), a biologically inspired representation-learning framework that enables explainable AI-guided cancer diagnosis. The model directly integrates the curated MSigDB Hallmark pathways, introducing pathway-constrained information flow and mechanistic interpretability through multi-level attention mechanisms. Based on TCGA RNA-seq data from 33 tumor types, H-PAAE compresses approximately 20,000 genes into a 128-dimensional latent space while preserving biologically meaningful structure. When used with XGBoost classification, H-PAAE delivers 92.37% test accuracy and 99.38% macro-AUROC with robust cross-validation results (92.5 $\pm$ 0.6%). SHAP analysis identifies a small number of key latent features, corresponding to conserved oncogenic processes, and pathway enrichment analysis shows strong overlap with cancer hallmarks. H-PAAE provides a clear and interpretable biological foundation for pan-cancer classification, with well-calibrated posterior probabilities that can be used for clinical decision-making, and can be easily integrated into multimodal diagnostic workflows.