Multimodal Federated Learning (MFL) faces significant challenges from statistical heterogeneity and modality gaps, particularly in unimodal client settings. Prevailing methods often fall short by inadequately modeling complex inter-client relationships for global aggregation and failing to guide local clients toward learning globally-aligned features. To address these limitations, we propose FedHyperClass, a novel framework featuring two core innovations. First, a server-side Hypergraph-Enhanced Global Aggregation mechanism constructs a hypergraph over local class prototypes and leverages a Hypergraph Neural Network (HNN) to generate robust, cross-modal global prototypes. Second, a client-side Adaptive Feature Enhancement module uses these refined global prototypes to guide local feature learning, fostering discriminability and alignment with the global consensus. Experiments on CIFAR-10 and CIFAR-100 demonstrate that FedHyperClass consistently and significantly outperforms state-of-the-art baselines across various heterogeneity levels. Ablation studies further validate the indispensable and synergistic roles of both proposed components.