In the beginning of the post you mention that you need to collect updates from all delta layers until you hit an image layer. Does that mean you have to make a O(log n) query once again every time you hit a delta layer? Is that a slowdown?
There was an attempt to use `im` crate instead of `rpds` somewhere in the middle of the Pull Request, but there were some issues with getting wrong results and whatnot. Did you manage to understand whether it's the crate or the calling code?