Accurately diagnosing student errors in programming education is both pedagogically critical and technically challenging, particularly under highly imbalanced data distributions that arise from long-tail error patterns. To address this challenge, we propose a two-level imbalance mitigation (TLIM) framework that integrates Borderline-SMOTE oversampling at the data level with class-weighted loss functions at the model level. Furthermore, we reclassify a multitude of fine-grained error labels into three pedagogically salient error categories—syntax, runtime, and logical—to alleviate extreme class imbalance while preserving essential instructional distinctions. Experiments on two large-scale, real-world coding datasets show that our approach consistently outperforms traditional pipelines by achieving higher macro F1 and AUC across four tree-based classifiers. Notably, CatBoost coupled with TLIM yields the best performance, underscoring the synergistic effects of informed error-type consolidation and advanced imbalance remediation. Our findings highlight the importance of targeted error features and robust multi-class strategies for enhancing automated feedback mechanisms in intelligent tutoring systems for programming education.