He means having both implementations use the same infrastructure. This is like taking your Raytracer class, and modify its Trace() method to check some global

He means having both implementations use the same infrastructure. This is like taking your Raytracer class, and modify its Trace() method to check some global flag to determine which algorithm to use, shunting both into the Raytracer. Instead, he argues for making two totally separate implementations: a RaytracerKD and a RaytracerBST (say), each of which have identical interfaces and which you switch between at a higher level. This way your level of coupling is lower and you hopefully have fewer problems, even though it may also be more complex.