I would always do the versioning code, otherwise data can not get converted when using builds.

Best to add the source (image) ocio role not the target display space. The proposed API can only we used for a strict setup.

Use OCIO (IMB_colormanagement)

I was also looking into this part and my solution was let the caller do the color management setup (set correct shader) and use the GPU package for only the low level stuff.

I would expect that the API is always added. It could be that some functions will have a dummy implementation when XR is turned off.

Add to top of the header

This presumably is specific to XR and does not need to be taken into account into other cases (and if it does, our code is incorrect because this is not used anywhere). Please update the description and function name to make that clear.

Same comment as above.

Make sure to run make format before committing, this seems like it should remain on one line.

Is this a temporary workaround? An option only for debugging/testing but not officially supported?

There should at least be a comment here explaining why it works this way, rather than being integrated in a more user friendly way.

Don't use the b prefix, we should get rid of that at some point for consistency rather than introduce it in more places.

It's not clear to me why there is Windows/DirectX specific logic here. I would expect it's possible to hide that in GHOST entirely, so that the secondary context is automatically created and used by GHOST without the WM being aware? Not a big deal but could be a good cleanup (possibly after commit).

Remove "Attempt to", no need to hedge words like this in descriptions.

I don't trust the use of context here, we have no idea what will be in there when this error handler is called.

Why is a popup menu needed here at all, rather than just showing them in the status bar? It seems unpredictable how many errors an OpenXR implementation might send or what their severity is. Maybe there are cases where it sends them continuously which would prevent the user from doing anything.

I suggest to just remove this line and evil_C.

Remove ROT_MODE_QUAT, already handled below.

But really, this should extract rotation and translation from obmat so constraints are taken into account. Camera rigs like that are pretty common.

Use CLOG_ERROR.

It's how the buffers interact, maybe some comments could clarify that.

I don't quite understand why ED_view3d_draw_offscreen_simple takes both the viewport and offscreen, but then at the end nothing was written to the offscreen buffer? Is it only user as an intermediate buffer, or is it not needed at all?

This does not appear to be used, so remove it from the header again?