I have context scratch-pad hashmap object I pass into a top-level function. It can then be decorated with extra scratchpad data all the way down the call-stack and passed into lower functions, for them to make use of. So each function can pass stuff down, but it's not available further up the call stack. It effectively looks like a stack object in terms of its semantics: as you unwind the stack you unwind history, 'undoing' changes. And the stack can take many different paths over execution.

Functions can do pretty much anything they want to the object further down the stack, without affecting other functions' inputs (parents or siblings). If it were mutable, the functions would suddenly be coupled to each other, and could change each other's data inputs. Add concurrency to that and it gets worse.

There are other ways to do this with Clojure. But I like this method, it's obvious and easy to test. It also feels reminiscent of Prolog.

In my example I'm associating new values into a hashmap, not appending to a vector, but it amounts to the same thing.