Creating a new datatype would pretty seriously go against Underscore's implicit philosophy of being a set of functions you bring to bear on ordinary data structures. In other words, you can try, but I suspect your pull request be voted down fairly decisively.
EDIT: apparently you are the creator of Underscore. Which makes me feel sort of silly for saying at least the second part. I think the first part stands, however.
So in order to save one call to chain() you'd use a non-standard object over a native array? What would underscore methods return, then? Perhaps it could do type detection and return what it gets, but then you have the issue that the executing code needs to be aware of that too - which means a type check at the caller before invoking the underscore expression. Needless to say, I think that's a lot of added complexity for not a lot of benefit.
Heck, if you really wanted to adapt something from Array.js you could do the naive (but I think correct) thing which is to add a flag to all underscore methods indicating whether or not the method should return an "Underscored" object or the basic data type. This could even be an initialization parameter for Underscore.
What does this buy you? I can only think of two things. First is they get a more object-oriented feel to their code.
var foo = _.filter([1,2,3], func, true);
// later on...
foo.filter(func2);
I guess the first question I have is, what does line 3 do? Are we mutating foo? Or returning something new? In which case we've managed to save precisely one character
_.filter(foo, func2);
Even worse, we're saving the wrong character. One undersung quality of having one global object per major library is that, in the code, you get a big contextual clue about what's going on from the way it's invoked. Contrast these two lines:
foo.filter(func2);
_.filter(foo, func2);
Try to imagine that you're encountering these lines deep in the bowels of the code. Which one is easier to understand?
The second thing you get is a (what I think is) esoteric benefit of being able to "ship" underscore functionality with the data to far away places. This is only an issue in full require()-controlled programs where there might be some communication between closures, and one side wants to use underscore in that context, but not add a dependency. Personally, I think being able to do that is just asking for trouble, as, again, your calling code maintainers are going to be totally befuddled by running into objects doing things that no other objects in that side are doing - or worse, doing something that looks familiar, but isn't.
makes it clearer that the thing you're getting back is like the thing you had previously.
One place this could be useful is in Backbone.Collection where the array-returning mixins from underscore all hand back bare arrays, on which devs must then switch to _ prefixed functions (although you're right that this does keep a clear conceptual separation between collections and other arrays).
I get a warm and fuzzy when I know what that method does, that it's not some random filter that could mean something totally different. (And yes, this does mean that my code is not very object oriented at all!)
It has one important downside tho - because you can't chain calls, your code ends up much less readable: It needs to be read from the inside-out, and the functions you're applying on the array are in reverse order compared to the order of execution. Its also harder to to see which arguments go to which function.
> The devil's in the details of interoperability with other libraries, as other folks note in this thread.
Indeed. I should have also mentioned that editing angular.js to work is fairly straightforward (~5-10 LOC), and this is probably fairly typical. So if you are willing to patch this isn't necessarily a blocker, and done right could probably be submitted as a reasonable merge request.
EDIT: apparently you are the creator of Underscore. Which makes me feel sort of silly for saying at least the second part. I think the first part stands, however.