Am I missing something? Done right, mark and sweep shouldn't have any POLA violation problems.

Objects may not be erased, but references can be dropped, resulting in
unreferenced records not getting marked, and thereby swept the next time through. As all exported references go through a membrane, one may effect an apparent deletion by dropping the wet reference. The dry user still has a reference to their now-useless lightweight access proxy, but has no way to get in any further than that. As useless references still count against the user's reference allowance, users are motivated to drop them from their lexicons.

how does mark and sweep not solve this space recovery problem?

The frequently encountered arguments (aside from ain't-broke-don't-fix) for reference counting over mark-and-sweep as I understand them are

   (1) reference counted systems never have to pause for GC
   (2) reference counted systems use slightly less memory / are faster, because they don't have to waste resources on marking and sweeping
   (3) M&S is broken because it does not guarantee time or ordering of destruction. 

M&S answers are

   (1) with robust threads,  GC can happen in a continuing thread, thereby avoiding jitter
   (2) that's a lie and even if was true it wouldn't matter; besides, M&S systems don't have to waste resources counting references.
   (3) you don't really care about time and order of space recovery, and if you think you do you have simply become spoiled by the immediacy of destruction following going-out-of-scope provided by reference counting . In a M&S system, when you really need to know when a resource is no longer accessible, like to support the "resource acquisition is locking" pattern, you must instead either explicitly close or implicitly trap goes-out-of-scope instead of destruction,  and have the shared resouce manage its own reference count concerned with the lock, instead of abusing the GC reference count for non-GC purposes. In theory, this approach to RAIL may release locks very slightly earlier.

Hmm.Should the referent get informed (if it cares) when a reference to it is dropped?

Also, using lightweight access proxies solves the identified flaw, regardless of GC model. Without that crucial piece, M&S would have it worse.
 

-- 
"Oblige a man to rise at four in the morning, and it is more than probable he will go willingly to bed at eight in the evening; and, having had eight hours sleep, he will rise more willingly at four in the morning following." -- Benjamin Franklin

Am I missing something? Done right, mark and sweep shouldn't have any POLA violation problems.