The caller waits for every chunk of a query, so a pool thread the
scheduler sets aside for a 15 ms time slice while holding one stalled the
drag. On a desktop busy with other work (12 busy processes on 16 cores) a
sweep of 16,000 Summit queries had p99 6.6 ms and a 163 ms worst case;
with the calling thread raised for the query and the workers raised for
good (they sleep when idle), p99 is 2.6 ms and the worst 5.5 ms. A quiet
machine is unchanged (p99 1.9 ms).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>