Vertex skinning produces no visible effect even with posed joints

[schell]

Vertex skinning produces no visible effect even with posed joints

Repro

1. Load a rigged model (gltf/Fox.glb) with a stage created .with_vertex_skinning(false).
2. Render the bind pose.
3. stage.set_has_vertex_skinning(true).
4. Build an Animator over the scene nodes + the first animation clip, and animator.progress(0.5) to pose the skeleton.
5. Render again.

Result: the render is byte-identical to step 2 — 0 pixels differ.

What is NOT the problem (all verified by instrumentation)

• The animator works: after progress(0.5), 22 of 26 Fox nodes' global translations changed
• The flag is set: set_has_vertex_skinning(true) writes GeometryDescriptor::has_skinning
• The loader wires skins: renderlet.set_skin(skin) (gltf.rs:1152), setting PrimitiveDescriptor::skin_id
• The loader reads per-vertex joints/weights: gltf.rs:419-450
• The shader gates on both conditions: if config.has_skinning && self.skin_id.is_some() (primitive/shader.rs:93)

So every CPU-side piece is in place; the failure is in the GPU skinning path.

Note on the existing test

The rigged_fox crate test does not actually assert a visual difference between the skinned
and unskinned renders: assert_img_eq_cfg is called against the no-skinning baseline with a
different test_name, and no rigged_fox.png baseline exists in the repo. The test only
proves enabling skinning doesn't crash. (Skinning at bind pose is also a no-op by definition -
joint matrix x its inverse bind = identity - which is why posing the skeleton, as above, is
necessary to even attempt a visual check.)

Suspects

• the SkinDescriptor / joints slab data read in the vertex shader
• joints/weights attribute wiring in the loaded vertices (e.g. always zero)
• the skin read path in PrimitiveDescriptor::transform (primitive/shader.rs:93+)

Context

Found while writing the animation manual chapter (worktree
docs-missing-manual-buildout). The chapter documents set_has_vertex_skinning with a
caveat and links this issue; the example omits a skinning demo since the feature currently
has no visible effect.

[schell]

Author's note: vertex skinning is known to have worked at some point - a lot of work went into it. The regression point is unknown, and we should find it with git bisect.

Bisect plan

Anchors from git history:

• c53b8c2 "feat: skinning / skeletal animation (feat: skinning / skeletal animation #35)" - skinning first landed
• 584bad1 "fix: rigging bug (fix: rigging bug #124)" - last commit to touch the fox test baselines and the has_skinning flag; likely the last visually-verified-working point
• 7628b57 "feat: API rework and docs v0.6.0 (feat: API rework and docs v0.6.0 #199)" and 7b393d5 (Renderlet -> GeometryDescriptor rework) - both touched the flag and the shader's skin read after the last verified-good point; prime suspects
• bad: current HEAD

Procedure:

1. First verify the good anchor: check out 584bad1, temporarily add a visual-diff assertion to the existing rigged_fox test (assert the skinned and unskinned renders differ, not just don't-crash), run. If it still fails there, move the good anchor to c53b8c2.
2. git bisect start HEAD <good> and at each step carry the same small assertion edit into the current rigged_fox test.
3. Wrinkle: the big API reworks (feat: API rework and docs v0.6.0 #199 era) change the test's surrounding API, so the edit may need adjustment at some steps; keep the assertion itself identical (byte-compare the two renders, >0 pixels).
4. Once found, diff the culprit commit's changes to primitive/shader.rs, geometry/, and the skin staging path.