As long there are encode_color and decode_color driver procedures I cannot see why linear color drawing functions should limit to devices whose color models are linear and separable – especially separable. 

I think default implementations for all such devices would be possible (in the worst case with serious performance penalty) and there are situations when implementing linear driving functions would be even highly desirable despite of the limitation above. 

A typical one would be the situation when the color index itself is a “non-linear and non-separable” POINTER to an otherwise linear and separable color data structure. As Studiorip is exactly the case it would be highly appreciated if this limitation was removed.