One subtle aspect is the replacement of _find_prefix with
_notmuch_database_prefix, which understands user headers. Otherwise
the code mainly consists of creating a fake prefix record (since the
user prefixes are not in the prefix table) and error handling.