A version update meant to steer back towards v4.4's voice,
using top-32 logits to strongly anchor towards the distribution while leaving the remaining ~2% probability mass available for MTP head co-adaptation.
Approximately matches the performance of v4.4.1's head and model.