Predicting fees is Too Hard, and over-paying increasingly Too Expensive. So you really need to BYO fees at spend time (rather than construction/signature exchange time), and if you can do that you can ride out fee spikes without needing to renegotiate with the other party, too, making your protocol more robust.
With current tech, we can only do this for SIGHASH-SINGLE/ANYONE_CAN_PAY (limiting your protocol to single in and out, which doesn’t usually work for multi-party stuff), OR using CPFP which requires one output which is spendable immediately.
i.e. CPFP is the only game in town today. v3 helps here today!
But I think the end game of Bitcoin is high fees (which will justify a great deal of engineering and effort to mitigate), and introspection which allows us more room to optimize. The medium case is bringing your own fee input/output at spend time, but really optimal is paying someone (over lightning or anything but onchain Bitcoin) to gather multiple unrelated txs together from everyone, with one fee input and output covering the entire thing.
Perhaps we would end up using introspection to limit the tx size anyway, so we just need the no-unconfirmed-children rule. Or perhaps the diminishing returns from bundling mean the tx size doesn’t have to be very large.
Now I write this out, I think I understand @instagibbs point about the generality of the v3 proposal, which would still serve here?
(End of a long day here, I will digest the rest of your post before the weekend, I hope!)