How to reproduce:

Type these two domains in omnibox and try to navigate to it. 

   한.google.com   : '한' is displayed as punycode.
   한.notgoodle.com  : '한' is displayed as Unicode (regular character)


The reason for this inconsistency is that we use punycode whenever the eTLD+1 portion of a hostname matches one of top domains in terms of 'similarity skeleton'.   To save space, we don't store the original names of top 10k domains, but only skeleton. So, the original name (google.com) matches itself and IDN-subcomponent is displayed as punycode. 

There's a (bit hackish) way to avoid that assuming that all top domains are non-IDN (which is the current assumption anyway).