The challenging part, which is part of what the research is one, is doing tree transformations in an efficient manner. Because this is a pointerless representation, you have to be careful in how you design the structure to ensure that you maximize locality and parallelism. If you read my "Key" paper in the publications:
https://github.com/arcfide/Co-dfns#publications
You'll see how I manage one of the most tricky elements. By using the techniques I describe in that paper, I can perform arbitrary computation over any group of sub-trees in the AST selected by their parent-child relationships in a data-parallel fashion. This is largely accomplished by converting that depth vector in the Quad-XML format into a "path matrix" which allows for computing and reasoning about the parent child relationships of any two arbitrary nodes in the AST without reference to any other parts of the tree, without pointers. I can then optimize that path matrix representation for either ease of construction or for performance over certain types of common operations.
That's really one of the most significant elements and sort of blows the whole problem wide open and actually makes it possible to do what I'm doing so easily.
Once I replace the standard recursive transformation idioms with this new set of Key/Path Matrix idioms, I use the Nanopass architecture to allow refactoring. Nanopass is a style of compiler construction that builds on the idea of functional programming. So, yes, the compiler itself is very very functionally oriented, and that's a very big part of the refactorability of the code. But also, since I have done so in a way that results in so few variable names, that's also a major component, and it means that code is often highly "independent" and can be removed or deleted easily.
What additionally interests me is the combination of points-free style and the kind of data structures you're processing in an array-biased language, could you give an insight on what that is like to work with?
In particular, I presume from your description, and only a conceptual familiarity with APL, that most or all of this code is "functional", i.e. all data structures exist as values passed between the composed functions, and nowhere else (no globals or similar). I'd love to hear more about the predominant data structures and what shape they take.
Somewhere else you mention Quad-XML, which seems to be a way to represent trees as arrays, with each element pre-fixed with its depth. I presume you use this for the AST? What kinds of operations are simpler on these arrays, and which are harder, compared to tree data structures used in other languages? For example, addressing the Nth child from a parent could be harder, since you have to search past the other children? I could imagine that operations like "set all fields X of the tree to Y" are a lot easier since no tree traversal is required.
Does your ability to quickly refactor rely on this functional nature?