To Alex
Thanks for sharing your scenario. You are right change tracking for backup is one of the most common reasons people are writing file system minifilters for CSVFS.
From change tracking perspective IMO state of the art is Microsoft DPM minifilter. Unfortunately, there is no public sample for what they are doing. I believe their focus is to track data changes. Their design avoids cross machine synchronization by tracking changes from each cluster node independently and saving these changes to a per node file. Later on when backup is creating volume snapshot, minifilter instances attached to CSVFS will see flush&hold and would seal existing tracking files and move to the new files. When backup application looks at the snapshot it will see same number tracking files as there are nodes in the cluster. It would merge them together and would backup everything that is changes in any of the files.
For metadata tracking (file rename etc) I would suggest to attach an instance of your minifilter to the NTFS/REFS below CSVFS, and do tracking from there to a separate tracking file, that would not be shared with anyone else, and just like above you can rely on backup to merge this tracking file with the rest.
From our experience this design will give you the best scalability since there is no cross-node synchronization/interactions, and would save you time on testing and support.