@iska @hayley the bacon paper for trial deletion basically doesn't track an object until its reference count goes above one and then is decreased again.

when you decrease the refcount and don't destroy the object, it puts that object on the :blobcatpolicedonut: list. so instead of scanning the world, it only scans stuff that has ever looked sussily like a cycle of self-dependence.

apparently its good enough for PHP, Nim and IBM's internal java projects, but the nim page did note that the cycle gc has some inconsistent scanning times.