Skip to content

Human action 42

Jerry Holland edited this page Oct 18, 2026 · 5 revisions

Establish representative host evidence

Canonical work item: #42.

When required: Before claiming representative real-repository or multi-host evidence. The issue carries current status, decisions, and completion evidence.

Recast 2026-10-18 (#289): evidence only. Budgets are no longer ratified here. ADR-0043 stops the performance program for Git-backed repositories, and ADR-0044 reports timings without gating them. The evidence informs the native phase ADRs as Gate B design evidence.

Owner role: Qualification-machine owners and repository maintainer.

Prerequisites: Select actual qualification machines, stable labels, representative real repositories/workloads, and quiet measurement windows. #22 is complete; the identified Windows reference is hosts.lab-windows-a. Do not re-record it merely to pass.

Real-repository Windows results now cover the public vcs-lab history and its actual notes. The catalog contains no resolutions, so that phase is explicitly an empty fast path. Inherited launcher configuration and a controlled local-read profile are recorded separately. These observations do not complete broader workload or POSIX evidence.

  1. Choose the representative workloads and document why they represent the intended use. Record host identity and environment without exposing credentials or private repository content.
  2. On each qualification machine, run npm run test:benchmark -- --host <assigned-label> first. Investigate deterministic regressions. For an initial identified baseline, deliberately run npm run benchmark:record -- --host <assigned-label>, then recheck with --json.
  3. Confirm each check selects the intended identified reference and does not skip latency. Review and commit the intended baseline changes with their provenance and reason.
  4. Use the maintained real-repository measurement procedure. From the source checkout with an optional native build, run node scripts/measure-real-repository.mjs --repo <clean-disposable-clone> --host <assigned-label> --output <new-file-outside-clone.json>. First clone with git clone --no-local <source> <clone>, then fetch notes and cst refs as that procedure specifies. Expected result: complete: true, unchanged repository checks, identical domain/CLI JSON, raw samples and explicit native/fallback counts. The raw-Git resolution floor is acquisition-only; no metadata-validation floor is claimed. Retain original configuration runs even if native falls back.
  5. Review the evidence and record on #42 what it shows and what it cannot show. No budget is ratified (recast 2026-10-18); timings are reported, not gated.

Verification: Identified synthetic baselines qualify those host checks only. Representative evidence additionally requires reviewed real-workload/raw-Git measurements; a deterministic-only pass does not qualify latency.

Failure handling and cleanup: On a mismatch or noisy run, stop and preserve evidence; arrange a quiet rerun rather than relaxing limits. Keep private workload details and run artifacts in the designated private destinations.

Completion: attach the evidence/decision to issue #42, resolve its human-action label when satisfied, update the private board, and point the Human TODO entry to the completed outcome.

Publish only evidence suitable for the public issue/wiki. Redact private paths, repository content, and credentials from shared output. Sensitive vulnerability findings belong in a draft security advisory.

Back to Human TODO · Benchmark host baselines

Clone this wiki locally