I took a stab at making up some test files. I’m sure we could do better but this was just to try to get something for people to respond to and attempt to get to what would be most useful for people.
These consist of the “CMS” and “cube_faces” images as full cubes and then re-encoded a few different ways for input to PQ transforms at a particular peak luminance, HLG at 1000, or DCDM 48 and 300 nit.
Visually the SDR and Rec709 and even P3 inverses look pretty decent.
The mean absolute difference for the Rec709 is < 6 code values in 16-bit, though there are some values somewhere that are quite large in error (i.e. the max is not so good).
As the peak luminance and color gamut expand, the “inverse > fwd” path gets less and less robust.
Here are the current image renders, which of course can be modified as people suggest what might make these more useful.