A recording stand-in for a generated script's Ctl.World, so an app's tests drive every <app>-ctl command without docker on the box and assert each command's exact docker argv.

It ships in lib/ rather than test/support/ because a dependency's test support is not compiled into the app that depends on it: this way every app's tests use the one recorder that matches the generated Ctl.World, instead of copying it. It has no dependencies and does nothing until install/2 names it in the calling process.

# test/test_helper.exs: once per VM, before any test runs, since the
# script's modules (`Ctl`, `Ctl.Env` …) are global.
Macaw.Host.Ctl.Recorder.load!("container/my_app-ctl")

# a test, with `@moduletag :tmp_dir`
alias Macaw.Host.Ctl.Recorder

test "start runs the container", %{tmp_dir: tmp_dir} do
  {script, _home} = Recorder.home!(tmp_dir, "my_app-ctl")
  Recorder.install([{["docker"], {"", 0}}])
  Ctl.dispatch(["start"], %{}, script)
  run = Enum.find(Recorder.docker_calls(), &match?(["run", "-d" | _], &1))
  assert ["run", "-d", "--name", "my_app" | _] = run
end

install/2 takes the canned answers: a list of {matcher, reply}, the matcher a list the command's [exe | args] must start with, or a function of (exe, args), and the reply {output, status}. The first matching entry answers; anything unmatched is {"", 1}, a docker that says no. Every call is recorded, with its options (where a secret's value travels, in :env). Everything lives in the calling process's dictionary, so async tests do not see each other's calls.