Based on Steve's last post, the main challenge is that the IO model assumes a bytes-based streaming API - it isn't really set up to cope with a UTF-16 buffering layer.

However, that's not substantially different from the situation when the standard streams are replaced with StringIO objects, and they don't have an underlying buffer object at all. That may be a suitable model for Windows console IO as well - present it to the user in a way that doesn't expose an underlying bytes-based API at all.

Now, it may not be feasible to implement this until we get the startup code cleaned up, but I'm not going to squash interest in improving the situation when it's one of the major culprits behind the "Unicode is even more broken in Python 3 than it is in Python 2" meme.