This gives us a much cleaner word list than other approaches I've
tried.

Ideally voc.txt should include common word forms and exclude very rare
ones.  There's no explicit word frequency information taken into account
in building this word list.  Someone having taken the time to add a
Māori translation to wiktionary means we'll favour more common words
(which is good) but the distribution of suffixed forms may not be
anything like that in real text.