First, thanks for splitting this out, because indeed it was getting a like sidetracked.
There are other reasons too, such as bandwidth wasting. In probably 99.9% of the time the hints won’t be required, so deciding when and what to send is an open question.
I don’t believe that’s correct; a better linearization could cause it to avoid eviction, or cause it to be mined faster. Perhaps the issues aren’t as severe as with RBF, but they still exist.
In general we cannot assume a “wallet” knows anything about mempool contents, and that a specific mempool is remotely consistent with another mempool.
You don’t have to “know” all the possible variants, you just have to relay a package with linearization information which ends up dominating in the diagram check.
The trick is finding a fix that isn’t “send entire clusters redundantly” and is still useful for the <<0.1% of the time it’s actually needed.