Operate on characters rather than bytes in the dialog so that it actually works with Kanji. Rewrite the auto-matcher to handle more cases and add unit tests for it.
