Yes, there is leakage in Punycode now and will be for a while. It is not "nice" but I couldnt think of any encoding which wont leak. (UTF-8 will give you gibberish if client are not UTF-8 aware or with the right fonts). Same arguments we have in IDN WG many years ago.
This implies that the conversion must take place at the semantic level rather than the encoding level. This should result in converted representations that can be pronounced, remembered and recognized, at the cost of somewhat large translation tables.