previously it backtracked too much, so instead of giving a useful
parse error, it just said "expected end of input" at the beginning of
the problem toplevel. which, if it's a namespace, could be way off.
e.g. "coe [_ ⇒ A] @p @q s" should immediately reduce to "s",
but if the "_ ⇒ A" happened to use an SY it didn't.
this will still happen if a wrong SY sneaks in but the alternative is
re-traversing the term over and over every time whnf runs
