Several sessions' make preflight runs overlap on one machine with nothing to space them out, and the overlap fails gates on tests no lane touched. On 2026-10-10 evening, two other sessions' full test suites were running at once (two go-build trees, a dozen .test binaries at 40-80% CPU each) and the one-minute load reached 124 on 16 cores. Four drain lanes in a row then lost a preflight run to load-only timeouts: TestAnIssueLaneLandsOnePullRequestThatResolvesItsIssue and TestAClosedPullRequestIsRefusedAndNothingIsCleanedUp (internal/core/implement/loop), and two internal/core/interview tests ('WaitDelay expired'). Each passed alone. The preflight load check (implement load --site preflight) warns above four times the core count but never waits, so every lane pays a re-run of about 20 minutes, and a lane whose re-run also lands in a peak has its receipt refused.

Several sessions' make preflight runs overlap on one machine with nothing to space them out, and the overlap fails gates on tests no lane touched. On 2026-10-10 evening, two other sessions' full test suites were running at once (two go-build trees, a dozen .test binaries at 40-80% CPU each) and the one-minute load reached 124 on 16 cores. Four drain lanes in a row then lost a preflight run to load-only timeouts: TestAnIssueLaneLandsOnePullRequestThatResolvesItsIssue and TestAClosedPullRequestIsRefusedAndNothingIsCleanedUp (internal/core/implement/loop), and two internal/core/interview tests ('WaitDelay expired'). Each passed alone. The preflight load check (implement load --site preflight) warns above four times the core count but never waits, so every lane pays a re-run of about 20 minutes, and a lane whose re-run also lands in a peak has its receipt refused.

Evidence 2026-10-11: a hand-written gate deadlocks, then releases together

Without the slot, sessions wrote their own waits, and the obvious one fails. A waiter written as bash -c '... pgrep -f "go test -race" ...' finds its own command line, which contains the pattern, so the gate never opens. Several sessions' waiters, and the briefs given to their sub-agents, used that form on 2026-10-11 and sat for hours. When they were stopped or edited they released within minutes of each other: four race runs started together (run-2610111657111142 lane-1, run-2610111657473488 lane-1, run-2610110908176735 lane-2, and a security lane), and the one-minute load reached about 500 to 920 on 16 cores at about 18:45 local time.

The anchored form, pgrep -f '(^|/)go test -race', does not match its own shell. Even so, a load gate alone let five preflights start in one quiet minute, and one timed out. Since then the sessions on this machine have used a hand-agreed rule: one preflight at a time, started only when the one-minute load is under 48, no (^|/)go test -race and no (^|/)make preflight is running, then a random wait of 15 to 60 seconds and a re-check of all three. That rule is the slot this record asks for, kept by hand. The remedy should ship it as one primitive that every session and every sub-agent brief reaches, and the LOAD rule should name the anchored pattern until it does.