The problem is that without knowing the type in question you don't know the values you're testing with cover all of the possible cases. When interfacing with an external dynamically-typed function this requires attempting to reverse-engineer the types from the definition (and recursively for all the functions it calls).
I really don't know where that canard comes from. My test suites rarely if ever explicitly test compatibility.
They test for specific values: if I do this and that, do I get the value that I expect. The type "test" gets covered implicitly, because if it's not the correct type, it will also not be the correct value.