I'm opening this issue in order to leave this idea here for future reference.
In #274 we started to consider a subset of the collidable points available in the model during simulation to reduce the computational burden for kinematic/dynamic quantities (transforms, Jacobians, Jacobian derivatives, contact forces, etc.). This kind of filter is static, and a potential direction to investigate is how to apply algorithms that can automatically choose the most appropriate collidable points (like as it is done in the fcl library for collision detection).
There several different considerations to do before starting to address this issue, for some of them see the comments in #247, like #274 (comment) and #274 (comment).