Fold a qwen3_5 VL trunk (or text-only model when it has no vision tower). Handles both tied (clones an untied folded lm_head) and untied heads.
Reader: fold the last (input) axis; γ (by tensor name) premultiplied.
Writer: R1ᵀ on the output axis of a 2-D [out, in] weight.
Writer bias: R1ᵀ on a 1-D vector.
MTP fc [H, 2H]: per-block input folds (embed γ, hidden γ) + output fold.
Fold a qwen3_5 VL trunk (or text-only model when it has no vision tower). Handles both tied (clones an untied folded lm_head) and untied heads.