I think the concerns about "perfect" behavior in corner cases are in general irrelevant here.

In the scenarios where this optimization matters, there is no quantitative change that occurs at 100% coverage. Preventing 99% of CoW is 99% as good as preventing 100% :) So the fact that a few objects here and there in special cases could still trigger CoW just doesn't matter; it's still a massive improvement over the status quo. (That said, I wouldn't _mind_ improving the coverage, e.g. if you can suggest a better way to find all heap objects instead of using the GC.)

And similarly, gps is right that the concern that immortal objects can keep other objects alive (even via references added after immortalization) is a non-issue in practice. There really is no other behavior one could prefer or expect instead.

> if said objects (isolated and untracked before and now tracked) acquire strong references to immortal objects, those objects will be visited when the gc starts calculating the isolated cycles and that requires a balanced reference count to work.

I'm not sure what you mean here by "balanced ref count" or by "work" :) What will happen anytime an immortal object gets into the GC, for any reason, is that the GC will "subtract" cyclic references and see that the immortal object still has a large refcount even after that adjustment, and so it will keep the immortal object and any cycle it is part of alive. This behavior is correct and should be fully expected; nothing breaks. It doesn't matter at all to the GC that this large refcount is "fictional," and it doesn't break the GC algorithm, it results only in the desired behavior of maintaining immortality of immortal objects.

It is perhaps slightly weird that this behavior falls out of the immortal bit being a high bit rather than being more explicit. I did do some experimentation with trying to explicitly prevent immortal instances from ever entering GC, but it turned out to be hard to do that in an efficient way. And motivation to do it is low, because there's nothing wrong with the behavior in the existing PR.