You have the shittiest “convert to grayscale” I’ve ever seen: it doesn’t come close to preserving the value of the colors. Convert the image to Lab, and look at the L channel, and you will see that the “ugly” combinations have very little value contrast, while the “nice” ones have a great deal more.

To see who is “exactly wrong” more graphically: http://img246.imageshack.us/img246/7098/grayfxbetternp7.png

Notice that in the bottom row, where we take the value from ugly and nice combinations and reverse them, the combinations with the “nice” choices of hue are now looking ugly as sin, while those with the “ugly” choices of hue no longer look so bad.

Also, in your second image, the green stripe on the left is quite a bit brighter than red or blue stripes, which is the dominant reason that the combination on the left looks “better”, while the three stripes on the right side are of quite similar value.

Image showing this: http://img503.imageshack.us/img503/2456/vomitev5betteryu5.pn...

I didn’t really have the half hour to spend on this (paper due tomorrow; blargh), so hopefully someone finds it edifying.

(Edit: None of this is to say that there isn’t some real truth to the link’s author’s claims about commonly preferred color combinations based on hue. It’s just that the graphics he uses to demonstrate it [and also the one you made] are horribly misleading, and mostly reflect a completely unrelated effect)

----

Another edit to add: the hang-up may be that the “value” of the HSV model has almost nothing to do with the term “value” as used by artists or color scientists. So that may be the source of your confusion. HSL and HSV were designed for the 1970s and are nearly useless in an age of fast computers.