The rendering of Gizmos by the last camera in the stack is the expected behavior in most cases since otherwise the base camera/previous overlay cameras widgets will be overdrawn (provided that the cameras have the same transforms). Now this might not be the desired behavior where cameras are non-overlapping and rendering some unique objects but functionally this departs from the initial purpose of the camera stacking. Pre render graph, the gizmos would be rendered for all the cameras but this approach would come with its own set of problems hence reverting to it is not a viable solution at the moment.
This was noted for now as a design limitation and will be partially filed under a feature request.