The process of generating diffs requires different types of strings at

different stages. The differ itself can technically work with either

byte or Unicode strings and doesn't really care, but when bringing the

"interesting lines" regexes into the process, the string types suddenly

matter. Our code expects the strings to be normalized to Unicode at this

stage, so that there's a consistent format to diff (without worrying

about mismatched encodings). However, we were passing byte strings in

some unit tests, which wasn't consistent with normal usage and caused

problems on Python 3. Those have been fixed to be Unicode.
The diff chunk generator, on the other hand, expects byte strings. It

takes these, normalizes them, converts to Unicode, and then hands them

off to the differ. To ensure it's getting the format it requires, it now

checks the types coming in during construction so that there's no

accidental cases of Unicode strings coming in.

Unit tests pass on Python 2.7 and 3.7 (with other in-progress changes).
Tested viewing a handful of diffs with Emoji and other non-ASCII content,

with and without a primed cache.