Here’s the answer as proposed by @DMGregory.

The key’s to make use of the collider itself for the .Cast as an alternative of the rigidbody

Old code – irrelevant elements omitted

non-public Rigidbody2D rb;

void Start()
{
  rb = GetComponent<Rigidbody2D>();
  collider = GetComponent<Collider2D>();
}

non-public bool HasCollision(Vector2 course)
{
  int rely = rb.Cast(
    course,
    movementFilter,
    collisions,
    collisionOffset
  );

  return rely > 0;
}

New code – irrelevant elements omitted:

non-public Rigidbody2D rb;
non-public Collider2D collider;

void Start()
{
  rb = GetComponent<Rigidbody2D>();
  collider = GetComponent<Collider2D>();
}

non-public bool HasCollision(Vector2 course)
{
  // Here's the important thing change!
  int rely = collider.Cast(
    course,
    movementFilter,
    collisions,
    collisionOffset
  );

  return rely > 0;
}

This works, nevertheless it’d be nonetheless nice to listen to a couple of solition solely based mostly on layers. The usecase could be to permit for composite shapes with out resorting to one thing like a foreach the place each step within the iteration does a .Cast.