Its possible that geometry has a part to play. As to inverse kinematics, it is what is used instead of coding long scripts to calculate the position of where the feet should be on the geometry. So it may actually end up being both.
But when it comes to collision checking, If it was that, I would understand why its not completely functioning just yet because constantly checking for collisions gets pretty heavy, especially if many things are checking at the same time.
Interesting insight. Thanks for sharing that! 