The current pretty conservative approach is to avoid merging parse states which
would cause a pair tokens to co-exist for the first time in any parse state,
where the two tokens can start with the same character and at least one of the
tokens can contain a character which is part of the grammar's separators.