I refactored some parts of CJK codecs for performances, after the PEP 393
was implemented. A blocker point was that these codecs have very few tests.
Not for valid data but for invalid data. It may be a little bit better. I
tried to write a test for each path in if/else, to test all cases, in the
codecs that I modified.

By error prone, it mean that it's easy to introduce a bug or a regressio,
since the code is complex and almost nobody maintains it.

I'm not stongly opposed to any change. I'm just trying to understand the
context.