First full million-vector run through the new harness. Cohere-768, 1M vectors, cosine, k=100. Recall came back 0.9929, serial p99 135ms -- genuinely good. Then the throughput number: 48.8 QPS. The leaderboards I measure against do thousands. 49 looks like a typo.
Spent most of a day convinced it was a bug, because I wanted it to be one. It wasn't. Queries go through the coordinator one at a time -- a mutex I put around query dispatch months ago, back when the epoch system was new and I didn't trust concurrent queries during a transition. "Remove it later," said the comment. Later never came, because worker compute was the bottleneck and the mutex never seemed to matter. At 1M it matters: one query in flight, appx 20ms of service, and you have a hard ceiling around 50 QPS no matter how many workers sit idle behind it.
The number was honest and the cause was mine, sitting in the code the whole time with a comment telling me to remove it. Wrote it down as a real ceiling with a real fix. Moved on.