Re: Surely the answer is
right - OCR "homonyms" should all translate to the appropriate charset before name lookups happen, or at least before registrars accept them as non-duplicates.
And doing periodic name cleanup might be a good idea, requiring takedowns of any domain that's a lookalike (and assuming they're being used for fraud).
So basically construct a map of UTF-8 chars to ISO8859-1 lookalike chars, then run every domain name through that matrix, see if duplicates show up.
I assume other-than-english lingos might need something similar.