The problem was that an exception can occur in the text.get() call or
in the write() call, when the text buffer contains non-ASCII
characters.  This causes the previous contents of the file to be lost.

The provisional fix is to call str(self.text.get(...)) *before*
opening the file, so that if the exception occurs, we never open the
file.

Two orthogonal better solutions have to wait for policy decisions:

1. We could try to encode the data as Latin-1 or as UTF-8; but that
   would require IDLE to grow a notion of file encoding which requires
   more thought.

2. We could make backups before overwriting a file.  This requires
   more thought because it needs to be fast and cross-platform and
   configurable.