One Request, End to End: Tokenize, Prefill, Decode, Stream

Corrections (updated Oct 29, 2026): the tiny engine in the first version of this post had two hidden slowdowns. One line (/ np.sqrt(self.hd)) quietly switched the math from 32-bit to slower 64-bit num