Interesting. So, out of interest, why is the same not being applied for ɢ? (When I ran it through Python's unidecode I got the roman symbol all the same).

The problem is that the RFCs aren't restrictive enough, partly because the IETF doesn't have much authority over registrars. The domain name rules really ought to be something like "one script, plus numbers, in a domain name part". But this runs into such things as the tendency in Japan to mix kanjii with English words. Then there's the whole right-to-left mark business, which has to coexist with left-to-right TLDs.

So if I mix ASCII with obscure UTF8 characters like the domain in OP's post I can register it then?

Something like www.ｐａｙpal.com --> www.n--pal-n76secrc.com