Thank you for the pull request, @keyboardspecialist!

✅ We can confirm we have a CLA on file for you.

1. Is there any sandcastle or anything I can use to verify the fix?
2. Reading through the issue, I'm a bit confused on why this set up causes splats to be transformed on the CPU rather than the GPU. Can you elaborate?
3. Did the issue also happen if the root tile has a non-identity transform and the child tiles have their own world transforms? Or does the root tile have to have an identity transform.

The issue should contain a test data set. The _rootTransform should probably no longer be called _rootTransform, because it isn't necessarily the "root transform". Although... it wasn't the root transform to begin with - it was the "computed transform", which includes the model matrix. In fact, the different configurations of where a certain transform takes place (in the data itself, in the root tile transform, in the child tile transform, in a glTF node matrix) would warrant ~5 (or maybe 5x5) test data sets. But creating test data where the transform happens in another glTF node than 'node 0' is moot, because this will break anyhow.

However, here's a test data set that probably demonstrates what the issue was probably about: With the transform in the root tile, it looks fine. With the transform in the child tile, it looks wrong and wiggly

The tilesets with a sandcastle:

Cesium Splat Tileset Transform PR 12926.zip

(It still uses the "old" extension format. I've prepared supporting the new extension format, but ... this was before the new extension format was ""final"", so it will have to be straightened out and reviewed...)

Thanks @keyboardspecialist. Once those last comments are addressed this should be good to go. I confirmed that this fixes the tilesets that were having the precision issues.