I just discovered that the default for me is some version of latin1, not ascii, so that utf-8 'ö' is read as 'ö'.  Hence, no UnicodeDecodeError even if 'ö' is being read and decoded by the initial readline call.  I presume the same is true on current Windows buildbots.  Hence, it was a bug to not include an encoding in the test file open calls.  (The widget itself *does* use explicit 'ascii' or 'utf-8' encodings for open().)