We must continue to support at least ascii strings, for backward compatibility reasons.  We can certainly improve the error messages, but the goal of this issue is to add support for bytes passwords.  I lean toward continuing to only support ascii strings, and making it the responsibility of the program to do the encoding to bytes when dealing with non-ascii.  However, I'd like to also be able to recommend in the docs what encoding is most likely to work, if someone can find out what encoding Thunderbird uses...however, it occurs to me that it may be using whatever encoding the OS is using (LC_LANG, oem codepage, etc), and that David's experiments worked because the same encoding was used for the same reason when the password was set.  I'm not sure how browsers/webmail works in that regard, honestly.

That's less important than just adding support for bytes passwords, though.