Classify every tensor of a qwen3_5 VL trunk (or its MTP companion) into a
FoldOp — pure name analysis, no arrays. R1-ONLY (+γ): the attention output
gate (o_proj(out · σ(gate)), elementwise in head space) does not commute
with a per-head rotation, so R2 is architecturally off for this family.
Corridor map: 02d723a:docs/design/turboquant.md §W1.
Classify every tensor of a qwen3_5 VL trunk (or its MTP companion) into a FoldOp — pure name analysis, no arrays. R1-ONLY (+γ): the attention output gate (
o_proj(out · σ(gate)), elementwise in head space) does not commute with a per-head rotation, so R2 is architecturally off for this family. Corridor map:02d723a:docs/design/turboquant.md§W1.