No rerun forgiveness, no ambient-load allowance
The user’s audit decisions of 2026-10-07 (PLAN.md, “USER DECISIONS (2026-10-07, audit)”):
- Flaky-rerun forgiveness is DELETED. Each flaky test is fixed at its cause, and the checks are deterministic. A new failure that the flaky memory already holds is a regression like any other: nothing is re-run alone to pass. A suspect leaves the reds only through FLAKE-REPRO, or is owned by a shrink-only burn-down entry: nothing ages out (there is no window, and a failing row the run history trims moves to its never-trimmed suspect archive, round 13).
- The budget-level forgiveness is DELETED too (round 12, the user’s
decision #2 of 2026-10-08). A CPU budget breach once passed until three
consecutive runs breached (
overCapRunsBeforeRed) and inside a “recorded, not red” spread band; both are deleted: one judged run over the cap fails (UR-24-BUDGETS). - The ~12% ambient-load allowance is REMOVED. It is a slope of 5% per
unit of recorded 5-minute load per CPU, capped at 12%. It is STILL APPLIED
in code (engine/src/shared/budget-verdict.ts
ambientInflationFraction, configwebUiSuiteCpuBudget.ambientInflation*) until AMBIENT-ALLOWANCE-REMOVE lands. The removal is written, regenerated in round 13 against the landing (with the floor re-derived from raw readings by the measuring tool, D-17), checked on every run to still apply (engine/test/workflow/parked-patches.test.ts) and kept at docs/trackers/patches/AMBIENT-ALLOWANCE-REMOVE.patch; it spans two leases (UR-23), so it lands ATOMICALLY in the web writer’s rebased workspace right after the engine merge, as its FIRST item. Until then every budget verdict states the allowance in force numerically, and what it added (D-17, UR-24-BUDGETS); and a reading counts only when it was measured inside the host’s one instrument lock.