I now vaguely recall writing an optimisation where the first attempt is skipped under some circumstances, and that could be the problem here (apparently that does not take into account possible negative values of \hyphenpenalty). I had forgotten about that optimisation, but it seems that that is what is going on.
Oh, ok! What will you do about this? Will you document the current behaviour as the LuaTeX behaviour, or will there be an option to change it? Thank you! -- Elie
