TestTheLoadTest failed on the post-merge main run's windows-latest -race leg
(job 95067154071) with "timed out after 5s waiting for churned clients to
fully unregister" — every worker had finished, the hub was still running,
and the goleak dump that followed was only the anchor drainers the t.Fatal
skipped stopping. The runner was simply slow: the ws package took 276s
against 159s on the identical tree an hour earlier, db and service ran
13-24% slower too, and Unregister drains asynchronously behind the hub
loop's remaining broadcast work.
Bound the settle wait by the test's own overallTimeout (90s), the same
"only a genuine hang takes this long" limit the workers use. waitFor
returns as soon as ClientCount matches, so a healthy run pays nothing —
locally under -race the whole test still finishes in ~9s.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>