nx_nxa_retarget.nx

nx_nxa_retarget.nx -- DONOR RETARGET: drive a rigged NXA that carries NO authored clips from a SOURCE rig's POSE, preserving the TARGET's OWN bone lengths BY CONSTRUCTION. WHY (measured 2026-08-23): all five donors report poses_in_library=0 -- they carry rigs but no authored motion, so the visiting cast on /world/beach stand as statues while every other piece of the character stack ships. This organ is the missing leg. THE NAIVE RETARGET IS PROVABLY WRONG, AND THAT PROOF IS WHY THIS FILE EXISTS. In the NXA normative form M x = D(x-b) + b + dt, applying joint j's transform to its own bind position gives M_j b_j = b_j + dt_j. So a posed joint sits at b_j + dt_j and POSED BONE LENGTH IS A PURE FUNCTION OF dt. Copying source dt -- or scaling it by a height ratio -- therefore cannot preserve the target's bone lengths unless the two skeletons are proportionally identical, and ours are emphatically not (104 joints at 1.715 m against 370 joints on a differently-proportioned donor). A dt-transfer retarget stretches and shears limbs on every donor: the exact defect the operator named as awful motion. It is not shipped here. THE ALGORITHM (four lines, and limb length is preserved BY CONSTRUCTION, not by tolerance): A_j = conj(D_parent) . D_j source: world delta -> local articulation W_j = W_parent . A_j target: parent-first composition P_j = P_parent + W_parent (x) (b'_j - b'_p) the TARGET's OWN bone vector, ROTATED ONLY dt'_j = P_j - b'_j The target bone vector is only ever rotated -- never scaled, never replaced -- so no pose this organ can emit is able to change a target bone's length. The limb-length tooth in the gate is a REGRESSION GUARD on that property, not the thing that establishes it. The general form needs a similarity conjugation by each joint's bind orientation. It collapses to the four lines above because NXA bind orientation is IDENTITY BY CONSTRUCTION -- nx_nxa_skin emits SKEL quats as identity and states the reason in its own header (delta-LBS convention: animation is applied RELATIVE to bind). VERIFY THAT BEFORE TRUSTING IT: the `bindquats` verb measures it per asset and is the reason this organ has three verbs instead of one. JOINT MAPPING IS BY TOPOLOGY AND NORMALIZED BIND POSITION, NEVER BY NAME. SKEL carries no names by design, and the donors come from different authoring tools -- a name map would be the classifier-keys-on-a-string defect. The normalized space is the same one nx_nxa_joints uses: every axis divided by the rig's OWN stature extent, so a taller or differently-scaled rig lands in the same space. The stature axis is DERIVED as the largest bind extent, never assumed z-up: the FBX/VRM donors arrive Y-up, which is the axis-convention defect a sibling lane measured in the renderer on the same day. usage: nx_nxa_retarget bindquats <file.nxa>