Diabetes is among the fastest-growing chronic diseases, so reliable, low-cost risk stratification is a public-health priority. Yet many machine learning models overstate their value: high accuracies, especially on the Pima dataset, are inflated by oversampling-before-split leakage, are validated on a single cohort, are delivered as black boxes, and are rarely benchmarked against strong baselines. This research aims to build a leakage-free, self-explaining framework that fairly benchmarks classical ensembles against modern deep models across cohorts of different clinical modality. X-MAP couples a calibrated soft-voting ensemble (logistic regression, random forest, extremely randomized trees, XGBoost, and LightGBM) with an F1-optimal threshold, and benchmarks it, under one protocol with oversampling confined to each cross-validation fold, against a multi-head self-attention transformer and an inductive patient-similarity graph neural network fused by a cross-validated meta-learner; every prediction carries SHAP, phenotype-manifold, and counterfactual explanations. On the Pima Indians (768 patients, metabolic) and Early-Stage Diabetes (520 patients, symptom) cohorts with bootstrap confidence intervals, X-MAP reaches AUC 0.836 (95% CI 0.809 to 0.863; F1 0.70, recall 0.85) and 0.998 (F1 0.985); neither the transformer (0.798) nor the graph network (0.819) surpassed the ensemble, and their fusion was down-weighted and non-improving, while explanations exposed coherent phenotypes (silhouette 0.74) and actionable counterfactuals. X-MAP delivers fair, transparent decision support and a cautionary benchmark for deep learning on small clinical data.