I'm late to the thread, but I've felt the same pain and wrote a library to help deal with the cognitive load of working on "shapes" of data [1].
Totally not saying "you're holding it wrong", but maybe once you're more than a few couple levels deep into a nested map it's time to look at an in memory db like [Datascript](https://github.com/tonsky/datascript)? Actual Datalog queries can replace get-in vectors growing out of hand, and you also get better mutations with transactions.
I do think that a static analysis tool would be of some help. Some sort of tooling to better handle tree structures would be very handy. I often find myself being off-by-one level with get-in calls on tree data (E.g. (get-in m [:a :b :d]) where m is {:a {:b {:c {:d 1}}}}), which is annoying because the NPE gets thrown 3 function calls up the stack.