Space invasion/AnyRGB: Decompose to RGB doesn't work
Here is a sample 32f linear gamma tiff in a well-behaved custom camera input profile:
As shown by this screenshot, decomposing this file to RGB works just fine in GIMP-2.10 (the two files shown on the right), but fails in GIMP-2.99 (the two files shown on the left):
This particular tiff, while still in the "camera space", has nice camera-captured detail in the Blue channel - very distinctive, including a water drop, which detail is missing from the Red and Green channels even in the "camera space".
GIMP-2.10 produces exactly the right channels from decomposing to RGB. But GIMP-2.99 produces the same channels as one would get from first converting to sRGB, and then decomposing. So instead of being nicely detailed, in GIMP-2.99 the blue channel is mostly black.
What is different in GIMP-2.99, that suddenly something so basic as decomposing to RGB channels suddenly produces wrong results by means of an inappropriate conversion to sRGB?
In GIMP-2.10, it doesn't matter what RGB color space is assigned to the image when decomposing to RGB - even assigning the wrong color space still produces the right RGB channels. This is true of many editing operations.
So why, post-space-invasion, is there an obvious use (misuse) of built-in sRGB when decomposing to RGB? I have a suspicion that this odd use of built-in sRGB is also being done other places in the code where in reality the color space information simply isn't needed for the requested operataion. Unless I'm totally forgetting something, the only places "space" (as in the actual primaries and TRC) information is needed is when there's a necessary conversion to "Y" or "XYZ" (eg xyY, Lab, LCh and Luminance operations), or if the window is being color-managed so there's a conversion from the image color space to the monitor profile, or some other actual color space conversion (such as when the user requests a conversion from one color space to another) is required.
So these sort of conversions to sRGB "post space invasion" are very puzzling. Even in GIMP with the babl TRC transforms, decomposing to RGB shouldn't need "space" information, of course apart from getting the TRC encoding correct.