LGTM. Some comments, not blockers for merging:

• are we quite confident allocating reachable_state on the stack is okay (it's 8 KB)? When GC is called from the eval breaker, it's probably okay. What about calling gc.collect() explicitly? Perhaps we should just heap allocate this. If that fails, you could fallback to a zero size stack (or maybe a small C stack allocated one). If heap allocated, you could make it a bit larger than 1024 slots, assuming benchmarking justifies that.
• for the two-ended sweep, a comment about why the backward step can't read an already spliced node would be good. The code is correct but it's suble.
• gc.get_objects() generally returns them "in that order" but only if the stack is large enough. Should perhaps be careful about making it should like something that can be relied on.

Closing this PR. I’ve learned that @brandtbucher's mentee has been working on something very close to this for a while and was planning to open source it once a few things cleared up. Rather than duplicate that effort, I’d prefer to focus on their work and help get it upstream, so I’ll be following up on that instead :)