The last commits consisted of cleanups, adding the same functionality for I3DM, and creating/adding the corresponding test data as specs.

The specs use the "golden" approach: The upgradeTileset/upAxisHandling directory contains the input directory, with the various test cases, and the golden directory with the expected output. When the TilesetUpgraderAxesSpec is run, then it will write the results into an output directory and compare them to the golden one.

The directory also contains a short README and a Sandcastle for comparing the input and output.

The case of gltfUpAxis+I3DM indeed brought up some issue, but ... one can argue whether this is something that should be handled: When the glTF 1.0 in an I3DM uses Z- or X-up, and this is not declared via the gltfUpAxis in the tileset JSON, then the upgrade does not generate a proper (unmodified) GLB. I'd have to spend much more time (with creating test data, and probably pen+paper) to figure out whether this can sensibly be addressed, and how: The conversion of the I3DM into EXT_mesh_gpu_instancing inevietably has to make assumptions about the meaning of the translations in the I3DM, and how they should be transferred into the glTF extension.

The following shows the Sandcastle for comparing the inputs and outputs, pausing for a moment in this case:

I think that this can be justified: The upgrade cannot magically "know" how the data was supposed to be interpreted. And when the gltfUpAxis is specified, then the upgrade works as expected.

Beyond that, there may still be a million cases that don't work. Maybe some tileset with gltfUpAxis=Z that refers to external tilesets with gltfUpAxis=X? Or obscure combinations of the CESIUM_RTC extension and an RTC_CENTER in the B3DM? ...

What I did test, though, was the current state with a user-provided real-world data set that used gltfUpAxis=Z, and the CESIUM_RTC extension, and I made sure that the result is the expected one. So even when there may be many cases where ~"something that may appear in legacy data" does not work, at least one real-world configuration of legacy data can be properly upgraded now.