You are right, I was too quick to paste a quote from lc documentation page, it should have been the last, fall-through case.

Also, kurisuto, my comment was not a solution, rather an attempt at explanation (to myself) of what was happening -- too rarely I deal with extended-ASCII, and yet not-utf8 strings. Proper fix (at least, for anything but one-time scripts) would be to always explicitly decode inputs from all sources, not hoping them to be Latin-1 only, and Perl silently doing "the right thing" in background.