In theory it's simple: we just run cheaper type analysis on lower tiers.
In practice it's more complicated, since the different tiers have to
cooperate on the types that they send back and forth.

I fixed up some of the code that didn't expect to receive a type worse
than what it knows to be true.  Ran into an issue where we have to
treat potentially-undefined symbols specially, since we can't do a class
check on them because they probably point to garbage.

Stopped there for now, since this section of the code is horrible and we
need to rewrite the tiering framework anyway.