speaking of this step, I’m wondering if it would make sense to consider same-txid-different-wtxid as an RBF conflict? Today we simply drop the new tx and report the existing wtxid in the mempool.
Post-cluster mempool if the conflicting tx is larger, dedup it with smaller. If it is smaller than existing and is in a new chunk that satisfies the package RBF rules, I think wtxid-replacement works naturally without introducing free relay. As a side benefit it “prefers” smaller witnesses for relay, reducing benefits of witness stuffing, e.g., inscriptions during a coinjoin.