Yeah, it's extra frustrating that you never know when clicking a tile that contains just a tiny bit of traffic light, whether the algorithm on the other side will see it too.

I think it is best to click the tile even if it only has a small part of the "object" in it. Other instances may have the same image but the grid is offset by a few pixels. At one point the tile does not contain the object at all and won't get any clicks. That way they could determine boundaries with higher resolution than the grid cell size.

Perhaps that's on purpose? If a tile is being selected only sometimes, you can infer things from that.