Well using Copra shows a tiny difference between V2 and V4 everything else being equal (these were CMYK profiles FWIW). I think Doug is using i1Profile which I also use. But I decided to try Copra to see if it too made a difference and it does by 1 device value in K and 2 in Yellow reported by Photoshop. I can pull out the old Macbook and run a test on MonacoPROFILER but I suspect we'll see the same thing. RelCol produces identical values with both V2 and V4 generated profiles.
V4 profiles are today, rather worthless anyway.
That's pretty much what I get. One of the curious aspects of the V4 profiles is the flat area below L=3 in graph 3 (V4 Perceptual). Drilling down into the profile the flat area is made by the "B Curve" of the BtoA0 tag. It cuts off exactly at L=3.1 which is the presumed black point of the ICC "reference medium." Per the spec, this only applies to Perceptual in V4 profiles in the BtoA0 tag. V4 profiles are rendered identically in the Relative tag (BtoA1). V2 tables have something similar named "input channels" and these are, in the ones I've looked at, always just a unity mapping.
Both V4 and V2 profiles implement a shallow "S curve" to enhance contrast in Perceptual which is appropriate for the assumed, standard reference medium.
Oddly, Photoshop either doesn't implement the B Curves in the BtoA0 tag or adds a bias to the PCSLAB values that obviates the BtoA0 clipping,
probably the former. This can be seen by converting an L ramp from 0 to 10 using a V4 printer profile. Proper conversion using the V4 algorithms should incorporate the B Curve and clip the device outputs below L=3.1 for Perceptual Intent. Little CMS does implement this per the V4 spec. Many programs use LCMS so there can be some compatibility issues interchanging between these. Which gets back to your point about V4 profiles being iffy. V2 profiles suffer from a more ambiguous specification but it appears that profile vendors are now making fairly consistent V2 profiles and the extant software handles them more consistently. Also, we are a ways away from being able to use the new floating point and parametric conversions that V4 promises which should have better profiles with a smaller footprint. But when you dance with a gorilla you have to follow the big guy's steps or you get trod on.
EtoA: Turns out it is the latter (see above) . I modified a profile's B curve on the BtoA0 tag to test this. Adobe adds a bias on V4 profiles to the PCSLAB values which produces somewhat similar results as the V2 profiles and then processes the V4 profiles while LCMS does not add a bias. Interesting.
Perhaps the most important takeaway is that Relative Colorimetric (both V2 and V4) is rendered quite accurately within gamut (Nice straight line) but the way it is defined is that it scales from the White Point to L=0,0,0 and not the paper's black point. As a result, in the proximity of the print's Dmax, there is an algorithm shift where an attempt is made to map to the closest gamut point and that is the paper's black point, not the darkest neutral color. This results not only in deep shadow blocking but a sudden color shift as well. In the case of the Baryta paper this is an increase of just under 2 in the a* channel. I've seen other papers where the shift was as much as 4. While this is arguably the correct approach for Relative Colorimetric Intent it can produce unexpected results in deep shadows, particularly on matte papers with a low (poor) Dmax.