To answer this we probably need more details, otherwise it's gonna be an XY Problem. What is it that I'm trying to do? How would I type this function in, say, SML, which isn't going to allow incorrect types but also doesn't allow these kinds of type gymnastics?
There's nothing I can do about the standard JavaScript library, but in terms of code I have influence over, I very simply would not write a difficult-to-type method like Array.prototype.flat(), if I could help it. That's what I mean by an XY Problem - why are we writing this difficult-to-type method in the first place and what can we do instead?
Let's suppose Array.prototype.flat() wasn't in the standard library, which is why I'm reviewing a PR with this gnarly type in it. If I went and asked you why you needed this, I guess you'd say the answer is: "because JavaScript lets me make heterogenous arrays, which lets me freely intermix elements and arrays and arrays of arrays and... in my arrays, and I'm doing that for something tree-like but also need to get an array of each element in the structure". To which I'd say something like "stop doing that, this isn't Lisp, define an actual data type for these things". Suddenly this typing problem goes away, because the type of your "flatten" method is just "MyStructure -> [MyElements]".
Sure, if you're living fully in your own application code, and you don't need to consume things from an API you don't control, it's easy to live in a walled garden of type purity.
I can recognize that most people are going to go for inaccurate types when fancier semantics are necessary to consume things from the network.
But we also have the real world where libraries are used by both JS devs and TS devs, and if we want to offer semantics that idiomatic for JS users (such as Array.prototype.flat()) while also providing a first-class experience to TS consumers, it is often valuable to have this higher-level aptitude with the TS type system.
As mentioned earlier, I believe 90% of TS devs are never in this position, or it's infrequent enough that they're not motivated to learn higher-level type mechanics. But I also disagree with the suggestion that such types should be avoided because you can always refactor your interface to provide structure that allows you to avoid them; You don't always control the shape of objects which permeate software boundaries, and when providing library-level code, the developer experience of the consumer is often prioritized, which often means providing a more flexible API that can only be properly typed with more complex types.
We don't have to deal in hypotheticals - we have a concrete example here. There's a method, array.flat() that does a thing that we can correctly describe in TypeScript's type system.
You say you would reject those correct types, but for what alternative?
It's hugely beneficial to library users to automatically get correctly type return values from functions without having to do error-prone casts. I would always take on the burden of correct types on the library side to improve the dev experience and reduce the risk of bugs on the library-consumption side.