There is an argument (from a number of expert sources) that the 18 colours in the CC24 are very useful but "low hanging fruit"; so I am taking your statement as the proposition to be tested rather than a confirmed statement of fact. If the tests show it to be correct, that would confirm your statement at least in my mind and make me happy.
The relevance of this issue to the O/Ps problem of course relates to the unconfirmed proposition that hue shifts may be a function of high colour saturation that the profile is not handling well either because the blue is OOG and not being appropriately compressed into the printable gamut, or the blue is say barely in-gamut which the profile is also not handling well. Or what Andrew more directly calls a "goofy" profile. Lending credibility to the profile being the problem is that Printer Manages Color renders the blues more accurately according to the visual perception of the O/P, but distracting from the profile being the problem is the O/Ps observation that another profile has reproduced the problem, so two goofy profiles? Maybe, but maybe not. This is a tough nut to crack, so far.
For what it's worth, I just made two prints of Andrew's image on Epson Prem. Luster 250 gsm using Rel Col on my 10+ y/o 9800. One used the canned Epson profile for that paper and one using a new profile made using my Isis. There is no visible difference outside of a very slight difference in rendering of the rainbow gradients used at the top and in some of the lettering. Big difference when viewing soft proofs but even there, both soft proofs are reasonably close. The woman, colorchecker, and image with the hand/jewelry looks the same on both prints.
I then ran a Matlab script on all the image pixels and compared the Lab values reported by the device space image RGB pixels comparing the canned profile and my custom profile.
Average dE00 1.2
Average dE00 w/o the OOG colors, 1.0.
Max dE00 (including OOG rainbow portions) 6.0
Average dE76 1.8
Average dE76 w/o the OOG colors, 1.5.
Max dE76 (including OOG rainbow portions) 8.1
Note that the test image doesn't have significant amounts of very low L* areas which are where the defects in Epson's canned profiles show up "bigly" (see Andrew's stress test image with Bill's Balls).
I've always found Epson's canned 9800 (and earlier 2400) profiles to be quite good but I didn't expect this close a match. Typically enough variation just between spectro brands to produce errors in this range and this was a 10 y/o profile and printer. I've only used OEM inks but haven't been particularly kind to my printer otherwise.
Note to the OP: If you PM me a mailing address I can mail you a print so you can physically compare with what you are getting.
EtoA: The blue jewelry shows no sign of purple/violet and, as noted, looks the same in both prints. The hue is really a long, long way from anything purple. It's closer to a cyanish blue if anything.