[rmodels] Fix glTF skinning when joints have non-joint parent nodes

[cseelhoff]

[rmodels] Fix glTF skinning when joints have non-joint parent nodes

Summary

Two surgical fixes in src/rmodels.c that correct glTF skinning for files
where the skin's joints are nested under intermediate non-joint transform
nodes that carry part of the bind-pose offset. Common in exports from
wow.export, and seen in some Blender/Maya/Houdini pipelines that wrap an
armature in a parent "rig" or "offset" empty.

Both spec-compliant files (every skeletal .glb shipped with raylib's
examples) and previously broken files now render correctly.

The bug

A glTF skin lists which nodes are joints. The skeleton hierarchy that
raylib reconstructs from that list assumes each joint's parent in
skin.joints[] equals its node-tree parent. Many real-world files
violate this: a joint can have a non-joint ancestor (a plain transform
node) sitting between it and the next joint up the chain. That
intermediate node's TRS is part of the model's bind pose and must be
applied to the joint's animation frames; otherwise the per-frame world
transform diverges from the bind world transform, the inverse-bind
multiplication in UpdateModelAnimationBoneMatrices produces
non-identity skin matrices at rest, and the mesh explodes.

LoadGLTF already handles this correctly when building bindPose
because it calls cgltf_node_transform_world(joint) per joint, which
walks the full ancestor chain. The two animation-side functions did not.

Repro file: babyoctopus.gltf from a wow.export of a World of Warcraft
model. Stock raylib renders a stretched/inverted mesh; f3d, the Khronos
sample viewer, and three.js all render it correctly. Screenshots
attached.

The fix

1. LoadBoneInfoGLTF

Replace the single-step parent comparison with a loop that walks
node.parent->parent->parent... until it either hits a node listed in
skin.joints[] or runs off the scene root. Joints whose direct parent
is a non-joint were previously assigned parent = -1 (treated as
skeleton roots), corrupting the chain used by BuildPoseFromParentJoints.

2. LoadModelAnimationsGLTF

Precompute once per skin a Matrix extOffset[k] per joint that bakes in
the composed TRS of every intermediate non-joint node between joint k
and its nearest joint ancestor. Then, in the per-frame loop, compose the
joint's local TRS into a matrix, multiply by extOffset[k], and
MatrixDecompose back to TRS before storing in keyframePoses. The
result is the joint's transform expressed relative to its skeleton
parent, which is what BuildPoseFromParentJoints expects.

This also subsumes the previous root-only worldTransform adjustment
block — the new extOffset[0] covers it for the root joint and
correctly extends the same logic to every joint. The dead variables and
the root patch-up are removed.

Diff size

+49 / -29 lines in one file. No new files, no API changes, no behaviour
changes for files that don't have intermediate non-joint nodes.

Validation

Built on Linux with cmake + gcc, no warnings.

Regression sweep across every skeletal .glb in raylib's examples
(robot.glb, greenman.glb, raylib_logo_3d.glb, old_car_new.glb,
plane.glb, shaders/robot.glb): pixel-identical output before/after
under fixed-camera screenshot comparison. Any noise was sub-0.0002 and
traced to HUD framerate-counter antialiasing, not 3D output.

Visual verification of the broken file (babyoctopus.gltf): now matches
f3d, the Khronos sample viewer, and three.js.

Known limitation (pre-existing, not addressed by this PR)

If a non-joint ancestor uses a 4×4 matrix instead of TRS
(node->has_matrix == 1 in cgltf), this patch's extOffset builder
reads n->scale/n->rotation/n->translation directly and ignores
matrix. This matches the rest of rmodels.c (no has_matrix
references; joint TRS reader behaves the same way). Properly handling
matrix-form nodes would be a separate, file-wide change.

Checklist

• Fixes a real bug observed in third-party files
• Two existing functions changed; no new APIs
• No behavioural change for files that previously rendered correctly
• Builds clean (Linux, gcc, no new warnings)
• Code style matches surrounding rmodels.c

[raysan5]

@cseelhoff Please, could you provide some affected model for testing?

[cseelhoff]

cow.zip

[kimkulling]

Just started to test it.

[kimkulling]

The fix worked for the cow. I will try the rest.

[kimkulling]

Works nice, great work.

[kimkulling]

Formatting

[kimkulling]

Formatting or try to write this more readable.

[raysan5]

@kimkulling Thanks for the review! Just approveed the CodeQL checks running.
@cseelhoff Please, could you reviewed the mentioned open points?

[cseelhoff]

• NULL check: validate the extOffset RL_MALLOC result before use, which
  was previously missing.

• Fail-fast handling: detect the allocation failure at its source and
  abort animation loading (log a warning, free resources, return NULL)
  instead of propagating a NULL pointer into the per-frame loop.

• Brace formatting: expand the collapsed one-line joint-match check
  (if (...) { isJoint = true; break; }) into raylib's standard Allman
  brace style.

• Readability: break the deeply nested MatrixMultiply(MatrixMultiply(...))
  into named nodeScale/nodeRotation/nodeTranslation/nodeTransform locals,
  mirroring the existing S/R/T composition later in the function.

• Spacing/line breaks: add blank lines within the precompute block to
  match the surrounding code style.

[raysan5]

@cseelhoff thanks for the improvement and thanks @kimkulling for the review! Merging!