Existing generalist detectors, such as mDeBERTa based classifiers, struggle with fine-grained AI-text detection in multilingual settings, particularly when hybrid authorship and morphologically rich languages obscure machine-generated artifacts. We propose a typology-aware Smart Routing Late Fusion Framework for 4-class detection: Human-Written (HW), Machine-Generated (MG), Machine-Humanized (MH), and Human-Polished (HM). The framework consists of two complementary tracks. Branch A targets agglutinative languages, including Korean, Turkish, and Hungarian, using error focused hard-negative refinement and Dependency Distance Minimization(DDM) / Uniform Information Density(UID) based pairwise calibration to recover weak generative signals hidden by dense morphology. Branch B applies RADAR-inspired invariant learning to capture machine-originated artifacts that persist after adversarial rewriting in languages with clearer segmentation cues. Experiments across eight languages show that typology-aware routing improves detection in agglutinative languages while maintaining strong multilingual performance. Our results highlight the importance of modeling both linguistic typology and hybrid authorship for robust multilingual AI-text detection.