I started to double-check my assumptions and expectations. So far one of my assumptions was that NxN MIMO works seemlessly, i.e. as long as wifi endpoints don't see conflicts, they just use the maximum available bandwidth and split payload into N pieces and transmit them in parallel. But I doubt that that is the case. After reading into Best practices and tools for measuring wifi performance I tried iperf3 with option -P and suddenly effective bit rate reports as 645Mbps (4 streams). That translates to ~80MBps.
@Lucky1 Did you also use -P in your tests? I also tried using multiple clients and that did not scale similarily to "iperf3 -P 4" and instead just seemed to split the expected throughput (minus a little loss) among them. If not fixable that is a bit disappointing as it renders MIMO to be basically a useless feature.
What still remains to be verified is the unexplained behavior with HT40 on 2.4G and VHT160 on 5G.