In a sufficiently complex game every build is on the pareto front as it optimizes some specific cost function.
However, if your one and only goal is winning I suspect that the high dimensional vector space will end up not looking so high dimensional once you account for the correlations between the different features you use. This is already clear from the very strong correlation between speed and accelaration.
[]I myself have played MK64 a lot and sometimes the goal was simply to see the world burn, standing on a corner with a shell waiting for the what would've been the winner of the race. Fond memories.
Even if you're only optimizing for race times, in the case of Mario Kart, the choice of track will have a huge impact on the optimal kart selection. Tool-assisted speedruns pick different karts for different tracks.