This fixes the issues I saw with generic/168 over NFS. Turns out it wasn't anything specific to that particular test cases, we just needed a sufficiently aged refcount btree to trigger it.