Sorry, my mistake. I was saving a file from textmate with the string "𝖙𝖊𝖘𝖙" in it and then opening with "open -a textedit eg.txt".

The same experiment with cat in place of textmate works fine, so it's textmate that is buggy.

According to "od -x1", textmate is writing:

  0000000    ed  a0  b5  ed  b6  99  ed  a0  b5  ed  b6  8a  ed  a0  b5  ed
  0000020    b6  98  ed  a0  b5  ed  b6  99                                

So textedit is right to complain.

Yeah, looks like Textmate is simply running the UTF-8 algorithm over UTF-16 code units, so each surrogate is being turned into a single UTF-8 code unit (which decodes to an invalid character).

It turns out that this is such a common mistake that there's even a name for this encoding, CESU-8: http://en.wikipedia.org/wiki/CESU-8