Wondering if this would be more feasible. If the user could specify three points for the shape then the normals could be calculated more easily trigonometrically as well as the area for mass/inertia calculations. Then, models could possibly be more easily made to collide as well if the model tris could be used to create triangle collision shapes automatically. Although this would probably not perform well, so some ability to create a model collision mask with lower polys could be provided.
EDIT: Thinking about it more, in conjunction with the supported shapes, there would have to be a depth for the triangle in 3D. If a front and back triangle were provided, then this would entail having to calculate collisions to a rhombus shape for the sides of the 3D object, so it would be more complex. Not sure if this would be the best way to thicken a triangle in 3D or not though, maybe the 3D experts can weigh in here Otherwise the other obvious way would just be to provide a 4th point and construct a pyramid style object with 4 triangle surfaces.