Collision detection systems (radar) are accurately not detecting an impending collision because the lights are not actually on a collision course with the vehicle.
Object recognition systems (computer vision) are working very well, because they recognize the lights and are updating the 3D map accordingly, but the traffic light 3D model is not designed to be a moving object - unlike vehicles, which frequently move. Which is why we see the car "passing through" them.
What we are likely seeing is simply a weird edge case in the output for the user-interface. I'd imagine if an object was actually flying at the car and the car could see it, it would brake accordingly.
Also, the map is two-dimensional. The car frequently drives underneath traffic lights that I'm sure also appear "on top of" the car in normal cases.
Object recognition and collision detection, from what I understand, are two very different systems.