many gaps that look more like processing errors

Hi, I've spent some time perusing these data. And they nowhere near cover the intended areas without supplementing with old v2 building datasets.

Here's a very curious example. Take the two quadkeys covering La Paz, Bolivia.

Quadkey: 210031332 LngLatBbox(west=-68.90625, south=-16.636191878397657, east=-68.203125, north=-15.961329081596647)
Quadkey: 210031333 LngLatBbox(west=-68.203125, south=-16.636191878397657, east=-67.5, north=-15.961329081596647)
DEBUG: 761 building polygons found in quadkey 210031332
DEBUG: 224,632 building polygons found in quadkey 210031333

Notice the vast difference in polygons between quadkeys covering large urban area. There are many buildings obviously missing from the 210031332 quadkey.

Loading the buildings into GIS software reveals massive holes of missing buildings.

Also notice the building polygons mysteriously disappear on the boundary line between those quadkeys! This gives me an indication that the missing building polygon's aren't due to some error on the ML, but instead some error in the data processing code used to build the quadkey files from the ML results.

I went looking for the data-processing source code, but it appears to be missing from this repo. Where is the source code for this?

No files or tests are named. Start by locating the repository entry point that builds quadkey files from ML results, then compare processing for quadkeys 210031332 and 210031333, especially at their boundary. Done means identifying the data-processing source and determining whether it explains the missing building polygons.