How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications