Benchmarking Gradient Boosting and Explainable AI for Credit Default Prediction on Imbalanced Financial Data: A Leakage-Safe Multi-Seed Evaluation with SHAP and LIME