Target feature selection for DSpark

Hi, thanks for great work!

DSpark uses the same target feature layers as the comparison drafters, but does not report a placement ablation. For Qwen3.5-style GDN/full-attention backbones, should features be selected uniformly by depth, by layer type, or at group boundaries? Have you tested whether Markov/confidence training changes the optimal selection?

Thanks again!

The issue names no files, tests, or entry points. Search the repository for target feature selection and comparison drafter implementations, then define experiments comparing depth, layer-type, and group-boundary selection with and without Markov/confidence training; done means a placement ablation and a reported recommendation for the stated backbones.

Written by the indexing model from the issue text.