138M, threads sharing one searcher (concurrency.py, 2,000 queries after 300), run again after stage6b,
nind and Lucene alternated (24 callers: numactl --physcpubind=0-23; 48: unpinned), queries/s:

             run 1   run 2
nind 24        479     533
nind 48        634     642
lucene 24      591     479
lucene 48      770     753

stage6b's nind with 48 threads (519) was a low run; the 24-caller runs vary by 10-20% either way.
nind's profile (perf, 12 and 48 threads) has the same shape: no lock, the scan and decoding.
