Excuse me, maybe I hadn’t been expressive enough on my question.
I thought on making a ZKP without considering consensus rules (like script executions) at all, but only as a compact way to prove a certain utreexo proof is bound to a tip, and that a tip has N PoW. I mean, taking exactly your trust assumptions, only counting UTXOs in and out of each block, and updating utreexo in the inside logic. If that was possible (and computationally feasable), so the ZKP substitutes the necessity of an utreexo proof commitment in the block.
If this ZKP is recursive (takes a past ZKP and proves upon it the addition on new blocks), so we could have a network of light clients that actually decentralize the network, capable of bootstraping third parties with (minor) trust. Instead of a competition of chains, we have a competition of ZKPs generated by those nodes, and someone just takes the most recent, with the most PoW ZKP.