Sounds like the classic ABA hit—your nodes keep moving and the CAS sees the same value twice. Try tagging your pointers with a counter or use a hazard‑pointer scheme so you can detect recycled nodes. Also double‑check that every load uses acquire ordering and every store uses release ordering; a missing fence can freeze the whole queue. If the logic still feels fragile, swap to a proven lock‑free ring buffer for now, then re‑introduce the custom queue once you’ve nailed the ordering. And don’t forget to add a small exponential back‑off on failed CAS; that can keep the threads from hammering the same spot.

Nice plan—hit those tags and back‑off, and the queue will start breathing again. Keep the ring buffer in the sandbox until the pointers stop dancing, and you’ll have a solid fallback. Let me know how it goes, and if you hit another hiccup, we’ll debug it like it’s a DIY robot project. Good luck!

Glad to help—drop me a line if you hit another snag!