This paper present our submission for SemEval-2026 Task 13 Subtask B, which requires the multi-class attribution of code snippets across 10 distinct AI generator families and a human baseline. Our proposed system utilizes a three-stage ensemble architecture specifically designed to navigate extreme class imbalance and capture subtle stylometric fingerprints. Initially, we employ Supervised Contrastive Learning to fine-tune a UniXcoder and ModernBERT backbone. Resulting embeddings are then processed by five heterogeneous shallow experts, each utilizing a multiclass decomposition to master specific generator lineages through specialized architectures. A Human Shield acts as a hierarchical safety auditor as an aggressive binary layer of human vs machine. Finally, a Context-Aware Gated Meta-Learner dynamically aggregates these expert opinions into a final predictions. Our experiments reveal that streamlining the system to a pure UniXcoder backbone fine-tuned with supervised contrastive learning improves performance, outclassing the official CodeBERT baseline with a final Macro-F1 score of 0.31389, ranking 26th overall.