How does Ruby Enumerators chaining work exactly?

Todd's answer is excellent, but I feel like seeing some more Ruby code might be beneficial. Specifically, let's try to write each and map on Array ourselves.

I won't use any Enumerable or Enumerator methods directly, so we see how it's all working under the hood (I'll still use for loops, and those technically call #each under the hood, but that's only cheating a little)

First, there's each. each is easy. It iterates over the array and applies a function to each element, before returning the original array.

def my_each(arr, &block)
  for i in 0..arr.length-1
    block[arr[i]]
  end
  arr
end

Simple enough. Now what if we don't pass a block. Let's change it up a bit to support that. We effectively want to delay the act of doing the each to allow the Enumerator to do its thing

def my_each(arr, &block)
  if block
    for i in 0..arr.length-1
      block[arr[i]]
    end
    arr
  else
    Enumerator.new do |y|
      my_each(arr) { |*x| y.yield(*x) }
    end
  end
end

So if we don't pass a block, we make an Enumerator that, when consumed, calls my_each, using the enumerator yield object as a block. The y object is a funny thing but you can just think of it as basically being the block you'll eventually pass in. So, in

my_each([1, 2, 3]).with_index { |x, i| x * i }

Think of y as being like the { |x, i| x * i } bit. It's a bit more complicated than that, but that's the idea.

Incidentally, on Ruby 2.7 and later, the Enumerator::Yielder object got its own #to_proc, so if you're on a recent Ruby version, you can just do

Enumerator.new do |y|
  my_each(arr, &y)
end

rather than

Enumerator.new do |y|
  my_each(arr) { |*x| y.yield(*x) }
end

Now let's extend this approach to map. Writing map with a block is easy. It's just like each but we accumulate the results.

def my_map(arr, &block)
  result = []
  for i in 0..arr.length-1
    result << block[arr[i]]
  end
  result
end

Simple enough. Now what if we don't pass a block? Let's do the exact same thing we did for my_each. That is, we're just going to make an Enumerator and, inside that Enumerator, we call my_map.

def my_map(arr, &block)
  if block
    result = []
    for i in 0..arr.length-1
      result << block[arr[i]]
    end
    result
  else
    Enumerator.new do |y|
      my_map(arr) { |*x| y.yield(*x) }
    end
  end
end

Now, the Enumerator knows that, whenever it eventually gets a block, it's going to use my_map on that block at the end. We can see that these two functions actually behave, on arrays, like map and each do

my_each([1, 2, 3]).with_index { |x, i| x * i } # [1, 2, 3]
my_map ([1, 2, 3]).with_index { |x, i| x * i } # [0, 2, 6]

So your intuition was spot on

map seems to carry the information that a function has to be applied, on top of carrying the data to iterate over. How does that work?

That's exactly what it does. map creates an Enumerator whose block knows to call map at the end, whereas each does the same but with each. Of course, in reality, all of this is implemented in C for efficiency and bootstrapping reasons, but the fundamental idea is still there.

Using Array#map without a block simply returns an Enumerator, where each element is then fed to Enumerator#with_index and the results of the block are returned as a collection. It's not complicated, and is similar to (but perhaps cleaner than) the following code. Using Ruby 3.0.1:

results = []
[1, 2, 3].each_with_index { results << _1 * _2 }
results
#=> [0, 2, 6]

Using Array#each doesn't return a collection from the block. It just returns self or another enumerator, so the expected behavior is different by design.