It's a geometric assumption like the other two. Consider a flat macrosurface. Its projected area in any direction $v$ is just $v\dot\ \hat N$ times its area (where $\hat N$ is the surface normal). In particular, the case where you're looking at it along the normal is simplest: the projected area is equal to the area of the surface.
Now split the macrosurface into microfacets. The total area of the microfacets is at least as much (assumption 2), but each 'kink' in the surface bends the normals of the separate microfacets away from the original normal. Whatever the shape of the microfacets, the sum of their projected areas doesn't change. In the case where you're looking along the normal, it's easy to see that the total projected area is the same: the surface would have to get larger or smaller for it to change.
For any direction, the microfacet has to cover a portion of the original projected area of the surface. Changing the orientation of the microfacet while still filling that portion doesn't change its projected area.
There's one tricky case, which is where the microfacets overhang each other. In this case, the total area is greater, because some area is covered by more than one microfacet. But in this case, at least one of the microfacets has to end up pointing away from the view direction, back into the surface. In this case, the dot product is negative, so this cancels out the area covered by more than one microfacet. This is why the text is careful to single out that it's the signed projected area.
There's one more tricky case, which is where the microfacets extend past the silhouette of the object. This might happen when you're looking from very glancing angles, or where overhanging facets overhang outside the perimeter of the surface. In this case, the projected area of the microfacets will be greater, violating the third assumption. We don't generally consider this case. Intuitively, it matches up with the fact that techniques like bump-mapping don't change the shape of the silhouette of the object.
