Interesting! Foundry USA mined a full-rbf double-spend, with a few minutes between tx #1 and tx #2: https://web.archive.org/web/20221221022035/https://fullrbf.mempool.observer/
My best guess is this has something to do with unconfirmed inputs. My stock v24.0 node saw both tx #1 and tx #2 at the same time initially, but rejected the latter for spending an unconfirmed input. Then a few minutes later, it accepted tx #2 once the input confirmed.
When you say “double spend” you’re just saying that the outputs on the replacement tx (that was included in a block) were distinct from the outputs in the original?
