Covariant type variable error in nested function

I believe the below code should be allowed:

from typing import *

T_co = TypeVar('T_co', covariant=True)

class Col(Generic[T_co]):
  def __init__(self, items: Iterable[T_co]) -> None:
    self.items = items

  def any(self, pred: Callable[[T_co], bool]) -> bool:
    return any(map(pred, self.items))

  def any_zero(self, fn: Callable[[T_co], int]) -> bool:
    def pred(x: T_co) -> bool: # error: Cannot use type variable as a parameter
      return fn(x) == 0
    return self.any(pred)

I believe the purpose of this error is to disallow breaking variance, e.g.:

  def unsafe(self) -> Callable[[T_co], int]:
    def fn(t: T_co) -> int:
      return id(t)
    return fn

However, restricting variance on nested functions has both false positives (first example), and false negatives:

class Unsafe(Generic[T_co]):
  def __init__(self, fn: Callable[[T_co], int]) -> None:
    self.fn = fn

  def get_fn(self) -> Callable[[T_co], int]:
    return self.fn

The above code has no error, but is not type safe. I believe the proper way to check for this is to look at the signatures of each exposed method and verify that the type parameter is not used with the wrong parity.

mypy 0.740, python 3.7.5

Start by reproducing the nested-function examples with mypy 0.740 and compare the reported variance error with the unsafe exposed-method example. Trace the variance checks for nested functions and exposed method signatures. Done means the safe example is accepted while the unsafe cases are rejected without introducing false negatives.