The geometry and bbox in these records aren't the same shape, which matters if you're indexing them.

I ran the shoelace formula over the Nepal Pelican footprint from the post. The bounding box around it comes out about 40% larger in area, so somewhere over a quarter of that box is ground with no pixels behind it. That scene is only 3.8 degrees off-nadir as well. The summarize table further down has some going up to 28.9, and those would be further off.

Indexing the bbox is the obvious move because it's four numbers and any database will put an R-tree on it. But coverage queries then hand you scenes for an AOI and the file turns up empty over that part of it. Took me longer than it should have to work out why there was a hole in my mosaic.

Separately, eo:cloud_cover is one number for the whole scene. Sorting on it to find your clearest option over a small area doesn't hold up. 85% cloud can still be clear over the part you want. The UDM2s give you it per pixel, but that's a second raster open per candidate before you can rank anything.