I don't have Windows Photoshop in front of me to test; though this might be a large source of error in the library.
Would you be ale to carry this over to an issue on GitHub? If your not familiar with that I can certainly do it for you and cite this thread.
In my preliminary test I have created an ICC Profile with
ociobakelut using the same transforms you have described, and applied them to JPG via
jpgicc provided by the Little-CMS project, and don't see error as dramatic as the one your presenting. Some errors are expected because of the precision loss due to having a Linear -> Display style ICC/3DLUT, which are generally handled by either using a "shaper" LUT or working with Logarithmic images within integer applications like Photoshop.
Long story short, it may be related to the Linear EXR behaviour in Photoshop, though we bake ICC's for Photoshop fairly regularly. It may also be the Windows build of
.