Greedy requests diverge when end-of-sequence termination is suppressed #103

Identical temperature: 0 requests do not produce identical token streams when the
request carries ignore_eos. The stream diverges early, inside the model's reasoning
channel, and because everything after the shift moves by one token, whether the
max_tokens boundary lands on the end-of-sequence token becomes a coin flip. The
visible symptom is an unstable finish_reason; the underlying claim is that greedy
decoding is not reproducible along that path.