I did look at QBVH and flattening it to a single dimension, but my understanding is that most specialized BVHs like this prefer mostly static geometry because updates are so expensive. I'd be happy to be wrong on this though - QBVH looked complicated enough that I didn't want to experiment with it without some strong signals that it would be the right direction.

From my ray-tracing days, I recall that the majority of the time spent in the acceleration structure was due to cache misses when traversing nodes.

It might be you want to use a binary partitioning algorithm or similar for just a few levels, and then have the leaf nodes be N spans in a (sorted) list, where N is somewhat large. Then you can have some fast loop to mow through the leaf spans.

Fair point. They mention they do a regular construction and then reduce it to the QBVH, so was wondering if the BIH approximate sort trick could be used effectively.

Again not put a lot of though into it.

Sounds like a fun challenge though.