We establish a KKT Point Embedding Principle where max-margin solutions (KKT points) in homogeneous networks cleanly embed into wider networks formed by neuron splitting, a property preserved throughout gradient flow dynamics.