Previously we'd stash a null pointer in a sorted vector of CUs - the
next time around, we'd try to do a binary search in that vector (sorting
on a key inside the objects pointed to by the elements of the vector)
which would deref null if we'd stashed a null in there previously.
As a reasonable, but not ideal, workaround - don't stash any result in
the vector - this means every query will produce a new warning
(resulting in duplicate warnings) but better than a crash.
Stashing null in the list could be workable if we also stashed the
offset in a pair - but then all the clients would need to be fixed up
(maybe using a filtering iterator) which seems like overkill for this
uncommon error case.
35fd37177b