The burst of canceled waits went through Fetch, where net/http races the
failed first write against its read loop on the torn-down connection, so
the error returned was sometimes EOF or "use of closed network
connection" instead of context.DeadlineExceeded (release/2026-10-06c
pull_request CI, run 468 / job 6104).
Drive the burst through waitTurn, the wait turnConn's first Write makes,
which returns the wait's error directly. Advance the fake clock during
the burst and check the host's slot is still the first send, so a slot
leaked by a canceled wait now fails the test; the request after the
burst still goes out exactly HostInterval after the first send with no
wait. Production code is unchanged.
Maestro: 73e54394-0c57-444d-bd5a-a14ca76dd9e8
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dev_agentandClaude Opus 5.5