Hello Will,
I think you pushed me in the right direction with your idea about the pixel shader.
But my current solution is a little bit simpler that what you described:
I now use a shader that (depending on a parameter) only draws the front or back-halve of the model.
Since the whole affair is from an orthographic top-down view along the z-axis, the shader only needs the z-position to recognize which part is front- and which part is back-side. So there is no need for directional normals and dot-operations.
I also don’t need a separate render target (which is good because of performance). I can now create the desired effect by drawing all the elements (back-model. front–model, sprite) in the right order.
So, thank you for your valuable input^^