At the very least because of my point-- any overhead costs (from the bandwidth, yes) don’t apply when the output is not actually spent because its actually unspendable or just uneconomic to spend. So for at least these outputs the tradeoff seems easier.
Aside, I was just ignoring the spentness encoding because it can probably be represented much more compactly. (Though the first scheme I had in mind required having the block handy to decode it, so no good. 
)
Alas, that is problematic given that a single input can require a megabyte txout proof. 
But it’s not fatal, you can just remember with every block header an additional hash for a root for some output tree, and proofs are against that. Wallets can remember their static fragments in that tree. Less reuse of existent datastructures, but at least it can be efficient and not have to present potentially a 1mb witness stripped transaction. One could even trim up the tree a bit by omitting provably unspendable outputs.