Position quantization precision problem in some cases

[hu-xd]

gltfpack version > 0.22 (present in gltfpack 1.2)
when quantizes positions with large coordinates + duplicate geometry,
much precision was lost.

repro.zip

Repro: attached repro.gltf.
check the node scale.

suspected cause:

The geometry_duplicate feature (introduced in 0.22):

1. dedupMeshes() marks two meshes geometry_duplicate (same geometry hash, different material).
2. detachMesh() keeps such meshes attached to their node instead of baking the transform to world space:

   if (mesh.nodes.size() == 1 && mesh.geometry_duplicate && !settings.mesh_merge)
       return;

3. So duplicate meshes stay in local space while unique meshes are baked to world space.
4. prepareQuantizationPosition() merges all bounds.

[zeux]

Worth noting is that this would be a problem regardless of deduplication if the source asset also contained two references to one mesh and one reference to another mesh. Currently can be fixed by using -kn (which keeps meshes attached) or in this case -mdd to disable deduplication / -mm to enforce detachment. Or -vpf to switch to floating point quantization which gracefully handles these cases.

This is indeed a general problem with mesh detaching, as it isn't currently restricted by whether the resulting scene will be quantized better or not. I'll take a look at whether better heuristics are possible here.

[zeux]

How was this specific mesh asset created? I am worried it looks quite synthetic; I expected the error to come from the world space transformation by an actual offset, but on further inspection all meshes have a significant offset embedded. This specific case might be better solved by -mm (or -mdd) than any reasonable heuristic I could implement, because for every individual mesh here detaching is not actually a problem - it's only a problem where some meshes are detached and some aren't.