It might be splitting hairs, but still wrong.
The exporter just marks the collision geoemtry. It does not optimize it. Optimization is achived by the way the data is stored in memory and the way it is accessible by the collision detection algorithm. Optimisation is always tied to the actual collision algorithm(s). One example for optimization is the BVH (Bounding Volume Hierarchy) trees which Bullet internally builds for complex triangle meshes.
If you read the text by drwr which you quoted, then you will see that drwr also talks about the data structures used for storing the geometry.
This is in on way related to file formats. File formats, like egg, hold just the geometry. It is up to the classes managing the information loaded from files to store and provide the data in optimized ways.
@rdb:
Seem like the confusion originates from blurry distinction between “rendering geometry” and “collision geometry”, caused by Panda3D supporting collision detection against visible geoemetry. Which is, imo, a very nice feature for rapid prototyping. Do you have a suggestion on how to best summarize this? I obviously don’t find the right words to explain it. Maybe we can add a paragraph in the manual page panda3d.org/manual/index.php/Physics, since this is not related to Bullet only.