Record: NO-LOAD-BLOCK
Status: in force
Built: partly built
Decided: 2026-10-15 by the user

Nothing is ever blocked by CPU load

Waiting for a busy machine to calm down before running, merging or judging anything means work stalls for reasons that have nothing to do with it, and on a shared host the wait can be endless. So no task ever waits on, or fails because of, how busy the machine is: what is judged is the work itself.

The user’s standing rule of 2026-10-15: “i never want anything to be blocked no matter what cpu usage we have” … “no matter what task”.

It applies to EVERY task: merges, validations, tests, builds, lab rounds and agent work.
Every judged measurement in the harness and the labs is load-independent: CPU work, not wall time (Q-TX2, UI-CPU). Wall values are recorded signals only.
No quiet waits or load thresholds anywhere: chains, gates, runners, labs.
Landings are never queued behind lab runs.
The one global instrument lock (INSTRUMENT-LOCK) gives way to concurrent runs under per-run caps once measurement is load-independent.
A “flake under load” needs no quiet host to reproduce: a test that fails under load is itself a load-dependence defect to fix (CPU work or event waits), not a flake to excuse (FLAKE-REPRO, amended).
Retroactive: a whole-tree and all-labs audit lists every load-dependent gate and wait, and each gets an owned fix.

Built (2026-10-22): heavy runs are admitted at once under per-run caps and never wait for each other (engine/src/process/instrument-lock.ts). Admission moves the admitting process itself into the run’s own cgroup v2 scope (memory ceiling, CPU weight) and under the harness’s CPU-time limit, so everything the run starts — a supervised command or a child spawned directly — is born under them, and a process borrows a run only from inside its scope: there is no uncapped path. A host that cannot cap a run refuses it. The harness’s script subprocesses are bounded by CPU work and one hang backstop, never a wall timeout (engine/src/process/subprocess-bounds.ts).

Built (2026-10-22), the measurement half: no reading is stamped “held” or refused “unslotted” — the stamp, its refusals in both suite reporters and the held filters of the rolling baseline, the measuring tool and the workspace’s budget judgement are deleted (audit A-04, A-05, A-07, A-12); a profile report is judgeable whatever ran beside it. The ambient envelope is deleted (A-08): no cap moves with load, and the load legs are printed as recorded signals. The browser profile’s wall-time budgets (LCP, long tasks, switch, typing, response and render times) are recorded signals only, never a verdict (WEB_PERFORMANCE_WALL_SIGNALS), and the profile takes no exclusivity.

What is missing: the browser’s work judged in CPU work (UI-CPU), which turns those signals back into gates, and a residual measured past the calibrated metric’s validity bound (validThroughLoad5PerCpu): past it a suite budget verdict is NOT JUDGED (said, never a pass, never a wait).