I had created a script that recursively looped through a boolean difference operation for exactly the reason of being able to identify problematic geometry (see post here). It was pretty straightforward to adapt that script for boolean union. Oddly enough, if the union fails the first time, it seems to work on the second pass. Note these scripts spread branches over threads so that if your building geometry is grouped into branchs, like @inno suggests, you can take advantage of all your cores.
20220430_MultiBooleanUnion.gh (2.2 MB)