I'm still not perfectly happy with the audio response, tbh. But it mostly suffices.

I experimented with both approaches, and landed on mixing the band energy normalized by global energy vs. per-bin max energy, with a very arbitrary coefficient that seems to work. I also keep a few copies of the levels per band with different rates of decay, so I can add features in the sketches which should be sensitive to transients, vs. features which can rely on slower rates of change.

Not at all an expert in any of this and it was a lot of trial and error, still lots of room for improvement.