I mean, my point was that a few dozen light sources times a few hundred sprites isn't large amounts of math, unless they're doing something really fancy with each pair. Language speed definitely isn't key here. It's got to be something else.

Finding the nearest three light sources is the main bottleneck, and in scenarios with many lights, is highly dependent on having a broadphase algorithm to filter them out early. If they did it the naive way, the algorithm becomes O(number of sources * number of lit objects).

But that doesn't seem to be their concern here, either. Optimizing broadphase collision is a well-understood problem, and the scenario of finding the nearest is only slightly different from that.

In more recent versions it is actually IL2CPP for most targets, and with the addition of former Insomiac devs, their C# subset (HPC#) has been taking a greater role on the engine.