I’m actually quite proud of having implemented the scoring client-side. It helps keep the load on the server down. Looking up all possible formations is somewhat computationally intensive.

It does create some problems when some clients have high latency and don't detect the formation in time, but I think it's fixable. One way could be to make the computation redundant, e.g. get three pixels to do the detection instead of one, and consider the formation valid if 2 out of 3 of them report success.