And this is despite the fact that Emacs Lisp routinely flouts every single software engineering "best practice". The language is dynamically scoped by default! It simply doesn't have namespaces! Static types? Hah! (And I, an inveterate Haskeller, don't even miss them.) You can—and people routinely do—hook directly into all sorts of implementation details from other parts of the codebase.
And yet it just works. And it works remarkably well.
My theory: what matters isn't "best practices", it's have a coherent conceptual design and code that reflects that design. Emacs is designed around a small but expressive set of core concepts that it uses in a consistent manner. Text with properties, buffers, modes, commands, customization variables... Almost everything more complex in Emacs is structured out of these (+ a handful more), and, once you've internalized them, it's surprisingly easy to both learn new higher-level tools and to write your own.
The design of both the user interface and the code directly reflect these concepts which gives us a naturally close connection between the UI and the code (it's almost trivial to jump from an interaction to the code that powers it), makes both UI and code units effortlessly composable and generally makes it easier to understand what's going on and how we can change it.