> I can't imagine how this could ever work, since a non C-collation
> string cannot be atomized into tokens like that. Considering each
> token successively during a comparison does not always produce the
> same answer as a straight comparison against the original string
> would.
I think the only fix is to remove support for those operators from the
operator class. The logic on match_special_index_operator() doesn't
seem to need any change. Patch attached. As far as I understand, we
cannot back-patch catalog changes. In this case, we can leave it
broken and apply the documentation part.