Rigidbody stuck in corners of Mesh Collider

My plan is to have a sort of vertical "tilt the board to get the marble to the bottom" game. The conditions so far are:

• I have a single mesh collider that doesn't move and a Rigidbody sphere collider.

• The only thing I ever change is the direction of gravity (which rotates based on mouse position).

• Also, this is restricted to a XY-plane (2D game).

Mostly, things are smooth, and the physics are great. But if you tilt the board and get the sphere in a corner for too long, it'll be stuck there, no matter where the gravity is rotated afterwards. It doesn't get stuck if the sphere stays in the corner for only a brief moment, though.

My first guess would be to do some kind of "sphere is in the wall" check, but I don't know how I would approach applying the correct force to get it out. As of right now, I do not apply any forces manually on the sphere; PhysX influences the sphere's velocity by itself.

The rotate gravity script is as follows:

 private var baseGravity : Vector3 = Vector3(0.0,-48.0,0.0);
 private var mousePos : Vector2;
 private var rot : float;
 private var rotSpeed : float = 250.0;
 
 function Update ()
 {
     // Get mouse position in screen in range of -1.0/1.0
     mousePos = Vector2(Mathf.Clamp(((Input.mousePosition.x / Screen.width) * 2.0) - 1.0,-1.0,1.0),
                     Mathf.Clamp(((Input.mousePosition.y / Screen.height) * 2.0) - 1.0,-1.0,1.0));

     // Rotate rot by mouse's x
     rot += mousePos.x*rotSpeed*Time.deltaTime;
     
     if (rot > 360.0)
         rot -= 360.0;
     else if (rot < 0.0)
         rot += 360.0;

     // Rotate baseGravity and change physics
     var v = Quaternion.AngleAxis(rot,-Vector3.forward) * baseGravity;
     Physics.gravity = Vector3(v.x,v.y,0.0);
 }

I've tried thus far:

• Checking/unchecking the "Smooth Sphere Collisions" in the mesh collider.

• Using standard primitive colliders in replacement of the mesh collider.

• Every combination of interpolation/collision detection mode on the sphere's rigidbody.

• Changing the scale of the objects in question (bigger/smaller), gravity speed (less/greater than -9.81), and increasing the speed of physics timesteps

• Rotating the mesh collider instead of gravity, which leads to a lot more problems than simply rotating gravity (and doesn't actually solve the problem in question).

...amongst other things I probably can't remember. Any ideas would be greatly appreciated!