Both Box2D (Box2D Lite, along with old and current versions of Box2D) and in Dirk Gregorius's most recent GDC lecture, the axis of most penetration seems to be used to identify the reference face.
I created an image to try to help explain what I'm thinking:
![Image](http://i.imgur.com/9yCNt7I.png)
I'm imagining the red arrow to point to a support point intersecting the plane with normal n2. However, I imagine the blue arrow pointing towards the support point for the plane with a normal of n1.
Here's a screenshot from Dirk's slides:
![Image](http://i.imgur.com/IzcFkua.png)
This algorithm would seem to me to end with a query resulting in the vertex the blue arrow points to, which is definitely not the vertex I'd be looking for. I'd want to end up finding the vertex the red arrow points to. It's my understanding that the > operator I underlined should be flipped (which I know it shouldn't be, so my understanding is flawed).
I'm imagining the vertex that represents the deepest penetration would be the support point that intersects it's corresponding plane, but has an intersection depth lower than all other possible intersection depths. This would properly represent the axis of least penetration, and penetration depth. So in other words, the intersecting support point with the lowest penetration.