Type system guard refinement false positive in `cond` with custom guard

[kidq330]

Existing issue

• I have searched existing issues and could not find a duplicate.

Elixir and Erlang/OTP versions

Erlang/OTP 28 [erts-16.4.0.1] [source] [64-bit] [smp:11:11] [ds:11:11:10] [async-threads:1] [jit]

Elixir 1.20.0 (compiled with Erlang/OTP 28)

Operating system

macOS Tahoe 26.5.1

Current behavior

Hi, congrats on your type system :)

I've bumped Elixir and compiled some of our projects and found something that simplified to this:

defmodule MRE do
  defguard g(x) when is_tuple(x) and tuple_size(x) == 3

  def f(x) do
    cond do
      g(x) -> :a
      is_atom(x) -> :b
    end
  end
end

and getting a false positive:

    warning: this clause in cond will never match:

        is_atom(x)

    since it has type:

        dynamic(false)

    where "x" was given the type:

        # type: dynamic({...})
        # from: mre.exs:6
        {arg1} = {x}

    type warning found at:
    │
  7 │       is_atom(x) -> :b
    │                  ~
    │
    └─ mre.exs:7:18: MRE.f/1

It looks like the guard analysis leaks into the next branch. This doesn't happen if the guard is inline or when using case instead of cond.

Expected behavior

No warning

[josevalim]

Following up on #15459, we need to make it so :erlang.orelse, :erlang.andalso, Kernel.or, Kernel.and, Kernel.|| and Kernel.&& can all pass this test:

            assert typecheck!(
              [x],
              cond do
                is_integer(x) or is_boolean(x) -> {:one, x}
                true -> {:two, x}
              end
            ) == opt_union(
              tuple([atom([:one]), opt_union(integer(), boolean())]),
              tuple([atom([:two]), negation(opt_union(integer(), boolean()))])
           )

As far as I know, none of them do it today. And it is tricky because is_integer(x) or is_boolean(y) gets a different result. We already have this machinery in guards, but we need to repurpose it.

[josevalim]

I pushed a fix specific to the v1.20 branch to avoid rocking the boat too much. I will fix main by fixing #15284, and then opening up another issue about refining or properly.