Solution

• For this milestone, only one level of boxing is provided per COLLADA file. Each geometry gets a bounding box, regardless of its complexity. To naively generate a hierarchy based on document structure makes risky assumptions about spatial locality. The entire COLLADA file does not get an uber-box because it may contain scattered elements, eg. some geometry that's close to the origin, but distant light sources, or maybe a skybox. Perhaps a top-down algorithm (eg. subdivision) could be implemented for future versions.
• There was no advantage to be gained in pre-instancing geometry during parsing and pulling bounding boxes up to enclose their instancing transformations - SceneJS is optimised to handle handle static transforms by memoising matrices, where there is no matrix computation to be avoided with the bounding box there anyway.