RSpec allow/expect vs just expect/and_return

In RSpec, specifically version >= 3, is there any difference between:

• Using allow to set up message expectations with parameters that return test doubles, and then using expect to make an assertion on the returned test doubles
• Just using expect to set up the expectation with parameters and return the test double

or is it all just semantics? I know that providing/specifying a return value with expect was the syntax in RSpec mocks 2.13, but as far as I can see, the syntax changed in RSpec mocks 3 to use allow.

However, in the (passing) sample code below, using either allow/expect or just expect/and_return seems to generate the same result. If one syntax was favoured over another, perhaps I would have expected there to be some kind of deprecation notice, but since there isn't, it would seem that both syntaxes are considered valid:

class Foo
  def self.bar(baz)
    # not important what happens to baz parameter
    # only important that it is passed in
    new
  end

  def qux
    # perform some action
  end
end

class SomethingThatCallsFoo
  def some_long_process(baz)
    # do some processing
    Foo.bar(baz).qux
    # do other processing
  end
end

describe SomethingThatCallsFoo do
  let(:foo_caller) { SomethingThatCallsFoo.new }

  describe '#some_long_process' do
    let(:foobar_result) { double('foobar_result') }
    let(:baz) { double('baz') }

    context 'using allow/expect' do
      before do
        allow(Foo).to receive(:bar).with(baz).and_return(foobar_result)
      end

      it 'calls qux method on result of Foo.bar(baz)' do
        expect(foobar_result).to receive(:qux)
        foo_caller.some_long_process(baz)
      end
    end

    context 'using expect/and_return' do
      it 'calls qux method on result of Foo.bar(baz)' do
        expect(Foo).to receive(:bar).with(baz).and_return(foobar_result)
        expect(foobar_result).to receive(:qux)
        foo_caller.some_long_process(baz)
      end
    end
  end
end

If I deliberately make the tests fail by changing the passed-in baz parameter in the expectation to a different test double, the errors are pretty much the same:

  1) SomethingThatCallsFoo#some_long_process using allow/expect calls quux method on result of Foo.bar(baz)
     Failure/Error: Foo.bar(baz).qux
       <Foo (class)> received :bar with unexpected arguments
         expected: (#<RSpec::Mocks::Double:0x3fe97a0127fc @name="baz">)
              got: (#<RSpec::Mocks::Double:0x3fe97998540c @name=nil>)
        Please stub a default value first if message might be received with other args as well.
     # ./foo_test.rb:16:in `some_long_process'
     # ./foo_test.rb:35:in `block (4 levels) in <top (required)>'

  2) SomethingThatCallsFoo#some_long_process using expect/and_return calls quux method on result of Foo.bar(baz)
     Failure/Error: Foo.bar(baz).qux
       <Foo (class)> received :bar with unexpected arguments
         expected: (#<RSpec::Mocks::Double:0x3fe979935fd8 @name="baz">)
              got: (#<RSpec::Mocks::Double:0x3fe979cc5c0c @name=nil>)
     # ./foo_test.rb:16:in `some_long_process'
     # ./foo_test.rb:43:in `block (4 levels) in <top (required)>'

So, are there any real differences between these two tests, either in result or expressed intent, or is it just semantics and/or personal preference? Should allow/expect be used over expect/and_return in general as it seems like it's the replacement syntax, or are each of them meant to be used in specific test scenarios?

Update

After reading Mori's answer's, I commented out the Foo.bar(baz).qux line from the example code above, and got the following errors:

  1) SomethingThatCallsFoo#some_long_process using allow/expect calls qux method on result of Foo.bar(baz)
     Failure/Error: expect(foobar_result).to receive(:qux)
       (Double "foobar_result").qux(any args)
           expected: 1 time with any arguments
           received: 0 times with any arguments
     # ./foo_test.rb:34:in `block (4 levels) in <top (required)>'

  2) SomethingThatCallsFoo#some_long_process using expect/and_return calls qux method on result of Foo.bar(baz)
     Failure/Error: expect(Foo).to receive(:bar).with(baz).and_return(foobar_result)
       (<Foo (class)>).bar(#<RSpec::Mocks::Double:0x3fc211944fa4 @name="baz">)
           expected: 1 time with arguments: (#<RSpec::Mocks::Double:0x3fc211944fa4 @name="baz">)
           received: 0 times
     # ./foo_test.rb:41:in `block (4 levels) in <top (required)>'

• The allow spec fails because the foobar_result double never gets to stand in for the result of Foo.bar(baz), and hence never has #qux called on it
• The expect spec fails at the point of Foo never receiving .bar(baz) so we don't even get to the point of interrogating the foobar_result double

Makes sense: it's not just a syntax change, and that expect/and_return does have a purpose different to allow/expect. I really should have checked the most obvious place: the RSpec Mocks README, specifically the following sections:

• Mock Objects and Test Stubs
• Test-Specific Extension
• Setting Responses

See the classic article Mocks Aren't Stubs. allow makes a stub while expect makes a mock. That is allow allows an object to return X instead of whatever it would return unstubbed, and expect is an allow plus an expectation of some state or event. When you write

allow(Foo).to receive(:bar).with(baz).and_return(foobar_result)

... you're telling the spec environment to modify Foo to return foobar_result when it receives :bar with baz. But when you write

expect(Foo).to receive(:bar).with(baz).and_return(foobar_result) 

... you're doing the same, plus telling the spec to fail unless Foo receives :bar with baz.

To see the difference, try both in examples where Foo does not receive :bar with baz.