Accumulator proofs that require updating though are problematic-- I think probably a lot worse than them costing weight even, particularly because they’re quite expensive for anyone who just wants to track all of them (an “index node”?). The idea that coins can be excluded from them is a good one, though if the criteria is too complicated I think people will justifiably worry that their coins will fall into them and then they are even worse off because they didn’t track the proofs. That’s why I was favoring a simple absolute limit: if your coins are above this value they’ll won’t get subjected to the accumulator. Also the non accumulated utxo set size has a fixed upper limit when you do that, – number of coins divided by that limit.
(Depending on how it’s set one could even bite the bullet and preallocate the usage!)
I wonder if it would be at all interesting to hybridize.. E.g. use a txout style proof that never changes for the coins identity and use a utxotree style proof for just the spentness bit. The total proof size would be increased, but nodes that maintain spentness wouldn’t need the spentness proof. (and probably the utxotree proof is smaller than you imagine since many spentness commitments could be stuffed under one leaf).