Graph Neural Networks (GNNs) have demonstrated impressive performance on homophilic graphs, largely due to their inherent message-passing mechanism. However, their effectiveness decreases significantly on heterophilic graphs. The core of this issue lies in information loss caused by heterophily mixing during aggregation: semantic messages from neighbors of different classes become entangled, leading to diminished discriminability. In this work, we introduce XMan-GNN, a novel message-passing paradigm as the remedy. The underlying idea is intuitive: constraining information to propagate along class-specific tracks, thereby preventing semantic interference. This is accomplished through the mixed-curvature product manifold, where each class is mapped to a distinct submanifold based on the assumption that nodes from different classes follow separate label distributions. Enabled by the disentangled aggregation strategy, XMan-GNN achieves state-of-the-art performance across a wide range of both heterophilic and homophilic graph benchmarks, with accuracy improvements of up to 4.18%. The source code is available at https://github.com/liulizhi1996/XMan-GNN .