Improve kNN for splat visibility
The current search does not find k-nearest exactly in the case where k is exceeded after validity test. Options include changing the search order to incrementally expand radius of search until k have matched, ensuring that the initial selection will be closer than later selections that may then get cut off by k being reached. Another option is to go back to the original camera sources and do a tighter search there. This last option has the advantage that points previously lost in the initial depth merge are now also included.
Edited by Nicolas Pope