I ran a few similar simulations (these are the ones I mentioned before where the woofer was modelled as a rigid boundary below the waveguide and is responsible for most of the diffraction seen in the simulation). When I ran a lower resolution mesh for the fully rounded corners the results had less obvious mesh issues at the highest frequencies than when I upped the resolution in the corners. I put this down to the mesh vertices not being aligned completely together as this has resolved similar differences before. The mitred corners suffered some of the same issue where it becomes difficult to rely on the simulation results as the frequency goes up. I hoped that if you could model them all at high enough resolution it might give a more confident result.