Also the reason to do it bottom-up is because you never understand the problem perfectly, especially if it has anything to do with real-world things. This means that whatever global domain model you created is wrong, and by the time you discover where exactly, you'll be so invested that it's too late to change it. It's easier to adjust the model when going bottom-up - you can change or discard just the top layers, as soon as you realize your new abstraction level doesn't match the problem domain properly.
I hesitate to use the phrase "domain specific language" because it brings baggage, but that's the platonic ideal of a toolkit.
The core skill of building toolkits is picking flexible ways to decompose problems. Top-down planning and decomposition has a way of biting you in the ass at the worst times, when you're almost done except for one corner case. Bottom-up construction tends to be more flexible if dirtier.
Familiarity with Lisp helps understand this concept, because creating a new "DSL" there is so easy that you treat it just like creating a new class or interface, and not as some kind of great endeavour that most people mean when thinking about DSLs today.
A good argument in favor of not breaking them into separate functions is available here[0], basically boiling down to the fact that splitting something into a separate function has the undesirable side effect of hiding the full scope (not in the programming sense) of what it does.
If this function is pure, this isn't a problem, but if it manipulates -- or even just accesses -- global state, you might call it when it's in a state that it doesn't expect, leading to subtle bugs. You also might move the call to it to a new location, not realizing the full scope of what it does (this is doubly true for removing calls).
[0] https://web.archive.org/web/20150926045718/http://number-non... (sad to see this site slip away ...)
This is a very important point, and one that a lot of developers miss, or even believe the opposite of. It's common to approach problems by 1. identifying the abstractions that you'd want to have to solve the problem, 2. implement those abstractions, and 3. use the abstractions to solve the problem.
This is how (by and large) we're taught to program, and it basically attacks the problem in a backwards way. Start specific, implement the whole thing with fairly minimal abstraction, then if the local complexity warrants abstraction, move to something more abstract.
Why? So that you can put off "the point past which you can no longer collectively understand the system". Local complexity is better than global complexity, and defaulting to a globally complex system is a terrible choice.