After a discussion of this on the perl-i18n mailing list, I've come to
understand the problem a bit more. In Japanese, text ordering is based
on phonetization. But as in english, there are multiple pronunciations
of a particular piece of text. Moreover, the "more correct"
pronunciation among the possibilities is influenced by the context of
the text. So in other words, the problem is intractable if all you have
is the text alone and inefficient even if you have more information
(because of the myriad factors that influence pronunciation).

The solution that I am using then is to store with each piece of
kana/kanji text, a kana-only phonetization of that text. I then rely on
the content editors to know the context of the text and supply an
accurate phonetization in kana. (In other words, I'm putting the
responsibility on someone else!) There does exist a determinate
ordering of the kana-only text and so this becomes a tractable problem.

Mike