Chapters one and two of the 1.63 arc: the bench exists, and the gate it feeds is a ratio, never a wall clock.
Added#
The ratio gate — chapter two of the 1.63 arc. Every bench workload is now also timed against a fixed calibration loop run in the same process, right after it — an integer loop no release has a reason to touch, deliberately not the product's own code, so a ratio moves only when the workload does.
wallMs / calibrationMsis the number a gate can hold: a CI runner lies about wall time to the workload and the yardstick by the same amount, and the lie cancels out.--record <file>writes the measured ratios as a committed baseline (schemaVersion: 1, covered by the versioning freeze the waytrazum.baseline.jsonis — an unknown version is a loud error naming--record, never a best-effort read);--against <file> --max-ratio <n>exits 1 when any measured workload is past its recorded ratio times the stated factor. The factor is required, not defaulted — how much regression is too much is a policy, the same rule as pulse's threshold. The JSON shape never changes with the gate flags: the verdict is the exit code and sentences on stderr, the waycheckhas always gated. A workload measured but absent from the baseline fails rather than passing silently, because a silent skip reads as coverage. Proved by breaking it: the suite gates a real run against a baseline whose recorded ratio is absurdly small and watches the build go red with the workload named.trazum bench— this machine, measured. Chapter one of the 1.63 arc: the standard workloads — a 1MB prompt at both levels, a 200,000-line profile, a 10,000-file walk, a 20,000-line roll-up — one shot each, wall time and peak RSS, as a table and as--json. No comparison and no judgement: the number is for a person with a change in hand, before and after; the CI gate is the next chapter's ratio, not a wall clock smuggled in here. Each workload runs in its own child process (this same CLI with--workload), because a peak is a fact about a process — five workloads sharing one heap would each report the high-water mark of whichever ran biggest before them, and the test proves the isolation by watching a peak sit below its predecessor, which a shared process cannot produce. The inputs are generated with the hostile-input suite's own LCG against a fixed pricing date — deterministic, and never written to the project, which the test proves by running the bench inside an empty directory and looking. "Peak heap" from the plan is reported as peak RSS and named as such: a heap high-water mark is not observable from inside a synchronous run without instrumentation that would itself move the number.
Fixed#
- Landing the bench collected four debts, each from a guard doing its job.
Three fired on the first push: the bench document had no CLAIMED harvest,
docs/format.mdhad no row (its count and the README's moved to seventeen of eighteen, and the two count guards learned the indefinite article on the way — "a eighteenth" was the sentence they would have demanded), and the--jsoncommand partition had bench in neither list. The fourth was the security suite:child_processis allowed in exactly one file, and the bench's spawn had landed in the CLI instead — it moved intogit.tsasrunSelf(), under the same stated rules, with nothing loosened.
Release mechanics#
- Five manifests,
@trazum/corein three places, lockfile at 1.62.1; README Action pins advanced tob55038e(1.62.0).