Indeed, and restricting the carve-out to some reduced size would work for lightning, but not general protocols.
But to revisit your example, RBF rule 2 (no other unconfirmed inputs) seems to want to prevent the requirement to consider entire packages, but by playing on top of an unconfirmed tx you work around this? If the RBF bypass rule were only to apply to v3 transactions, I think this could not be done. Otherwise, we could consider replacing Rule 2 with a more sophisticated requirement that the new package outbid the one being replaced.