> how their behavior so to speak, is defined before conversion to the internal RGB color space
The guides for creating CFAs are based on a trade-off between tolerable metrics ambiguity (including metameric failures) and low noise. To have low noise and better colour response stability in shadows the filters deviate from L-I, and also they are more transparent meaning lower separation. We are now on the third generation of CFAs, first being very strict and thus cameras were more noisy, the second was too relaxed and resulted in mushy and plastic colours, now it is in-between, which is possible due to lower read noise. Incidentally larger pixels allow for stricter CFAs due to "lower" noise, and so are larger sensors (one of the reasons semi-MF sensors are more joy to me). There is a disconnect between filter design and colour transforms design - good colour transforms are based on experiments and not on design data. It is like it used to be with lenses, resolution was determined through tests and not based on calculations, and even now the calculated MTFs can't be trusted much. Internal space is not always RGB; and even if it is an RGB it is often that quite a lot of calculations is performed over the camera RGB.
Scanner filters were dealing with 4 main types of dyes, slide, neg, photopaper cmy, and press cmyk. Those can't be separated well enough out of the box with the same set of filters. But profiling takes a very good care of the diversity making it possible to reproduce all types of originals with a single profile made from a combined target which contains different emulsion types. The situation with Bayer filters is similar, given we normalize the target by a flat field before profiling, and we profile using targets encompassing 10 stops (2 shots of SG 4 stops apart with a black trap /SpyderCUBE/ included in the scene is a good way, but the first shot needs to be exposed so that raw values for the white patch are about 227 /that corresponds to 242 in gamma 2.2/); or have an adequate setup with a monochromator. More, for most of the practical purposes even colour separation we have with a film is quite adequate to cover for a wide variety of light and scenery. On a side note, digital photography introduced a false notion that filters on the lens are not necessary anymore.
Good and bad ways do exist, but there is a lot of information allowing to avoid bad ways. Assumptions currently (since 1999 at least) are not necessary because the math. models and calculation methods we have are quite good. Ironically, not so good for printing.
