I think users quickly respond to standardized index points and the concern you have over items which match the labels, while valid, is less important than it seems at first. Given the two choices, I'll take standardized divisions of the alphabet within the site. This is especially true since it means there's no batching up the output, but I'd prefer it anyway.

As for the splitting specifically on the vowels, I tend to agree that there are probably statistically much better places to divide. One of my nodes in the thread already gets into that. However, that's one choice that's pretty easily changed by tweaking one array in my example. It could be tweaked as the data is updated and measured to fit hacker's specific site, it could be done by frequency analysis of an existing dictionary, it could be done ad-hoc, or there are probably existing works expounding on such matters. It has more to do with the data than the code, and in advance we really don't know how well any particular section breaks will work for hacker, although I have my suspicions that something other than simply the vowels would work better. I'm just not sure what.