Codify encoding behaviour

[mudge]

While working on #59 last year and trying to plug gaps in test coverage, I realised the behaviour of string encoding when using Latin and UTF-8 re2 patterns is not especially well-documented. This led me to doubt the correctness of its behaviour but, regardless of that, the current behaviour should be better tested and documented.