Rethinking Language Model Scaling Under Transferable Hypersphere Optimization