bench: commit m3 probe manifest and full-pass ledger

[bborbe]

Summary

• Commit bench/prs-probe-m3.json (probe-m3-4 manifest) and bench/probe-m3-results/results.jsonl (m3 curated-1 full-pass ledger, config d5a15f16, 20 rows) — the raw rows the discrimination report page was regenerated from
• Include the pre-existing docs/git-workflow.md addition: repo-admin bypass makes the master ruleset a guard only for non-admins; using the bypass is a separate decision from the edit, prefer branch + PR regardless
• Makes the m3 report page scoring reproducible from committed artifacts (report page itself landed in PR bench: commit curated-1 discrimination report pages (golden-curated-3) #113)

Test plan

• bench/prs-probe-m3.json loads (probe-m3-4, 4 PRs)
• bench/probe-m3-results/results.jsonl has 20 rows, model m3, prs_version curated-1, config d5a15f16
• git status clean

[ben-s-pull-request-reviewer]

Now I have everything needed to complete the review. Let me execute the selector-mode steps.

Step 4b-i — Judgment candidates: Changed files are bench/probe-m3-results/results.jsonl, bench/prs-probe-m3.json, docs/git-workflow.md. Checking triggers:

• git-workflow/never-direct-commit-to-master — trigger @commits (always in scope), but applies_when requires actual commits landing on master. This PR has zero commits in the diff (doc-only edit + new bench data files). applies_when condition not met → skipped.
• git-commit/imperative-mood — trigger @commits, but no commit subject is present in the diff. applies_when not met → skipped.
• No other judgment rules have triggers matching the three changed files.

Step 4a mechanical findings: One finding on CHANGELOG.md (changelog/unreleased-entry-required) is stale — CHANGELOG.md is NOT among the 3 changed files in this PR. Dismissed.

Step 4c-sel short-circuit: Applicable set is empty AND mechanical findings reduce to zero after stale-filter → selector clean — no adjudication needed.

Step 5 — Manual review of the 3 changed files:

1. bench/probe-m3-results/results.jsonl — 20-line JSONL of m3 probe run results. Each record is a valid JSON object with consistent field structure (config_hash: "d5a15f16..." matching concern note, model: "m3", prs_version: "curated-1"). The findings arrays vary by entry (some empty, some populated from the reviewed PRs) — this is expected ledger content, not code.

2. bench/prs-probe-m3.json — 69-line valid JSON manifest version: "probe-m3-4" listing 4 PRs. All fields present and well-structured. No issues.

3. docs/git-workflow.md — One paragraph added correctly documenting that the GitHub ruleset bypass_actors configuration lets repo admins push to master directly. The new text is accurate, appropriately warns that bypass approval is separate from PR approval, and provides a verifiable gh api command to check the ruleset.

Plan concern check: All three concerns are addressed — schema matches config hash, JSON is valid, bypass logic correctly documented.

---

{
"verdict": "approve",
"summary": "PR adds bench ledger data (20-row JSONL result log + 4-PR probe manifest) and a correct clarification to git-workflow.md that repo-admin bypass exists. No code, no violations, no stale findings from the mechanical funnel that survive filtering.",
"comments": [],
"concerns_addressed": [
"correctness: 20-row JSONL ledger matches config_hash d5a15f16 — confirmed, consistent schema across all 20 entries",
"correctness: bench/prs-probe-m3.json is valid JSON, 4-PR manifest version probe-m3-4 — confirmed",
"correctness: git-workflow.md bypass note is accurate — repo-admin bypass_actors is a real GitHub ruleset feature; the added text documents it correctly and advises asking explicitly"
]
}

superseded by new automated review

[ben-s-pull-request-reviewer]

I now have enough information to produce the consolidated report.

Step 4b-i candidates — changed files are:

• bench/probe-m3-results/results.jsonl (new JSONL)
• bench/prs-probe-m3.json (new JSON)
• docs/git-workflow.md (doc)
• rules/index.json (rule index)

Citation check: The mechanical finding's rule_id changelog/unreleased-entry-required does not exist in rules/index.json (verified by grep above) — stale Walker citation, dropped. Additionally, CHANGELOG.md is not in this diff at all (confirmed by diff stat: 4 files, none is CHANGELOG).

Applicable judgment rules:

• git-workflow/never-direct-commit-to-master — always-on (trigger: null), and this PR touches the rule itself.

---

PR Review: bench: commit m3 probe manifest and full-pass ledger

Must Fix (Critical)

None.

Should Fix (Important)

None.

Nice to Have (Optional)

None.

Traceability (selector mode)

• Candidates: 1 rule matched by Step 4b-i glob filter (git-workflow/never-direct-commit-to-master)
• Applicable: 1 (git-workflow/never-direct-commit-to-master)
• Skipped: n/a — no other candidates
• Mechanical findings: 1 (changelog/unreleased-entry-required) — dropped: rule_id not in rules/index.json (stale Walker citation). Additionally, CHANGELOG.md is not in this diff, so the finding is doubly non-applicable.

---

Notes:

The diff commits two new benchmark artifacts (results.jsonl, prs-probe-m3.json) and updates docs/git-workflow.md with a clarification that repo-admin bypass is a real GitHub ruleset behaviour, not a documentation gap. This clarification is accurate and consistent with the rule's scope.

The rules/index.json change removes "trigger": ["@commits"] from the git-workflow/never-direct-commit-to-master rule block. The rule remains (level/owner/anchor unchanged); the doc change explains the bypass caveat. The trigger field's role in rules/index.json is index-level — it is not consumed by any ast-grep YAML (those carry their own trigger globs), so removing it does not break YAML-based enforcement. The ### RULE block in docs/git-workflow.md is the source of truth and it is intact. No issue.

The mechanical funnel finding (changelog/unreleased-entry-required) was a false positive: (a) the rule_id does not exist in the index, and (b) CHANGELOG.md was not modified in this PR — the runner scanned results.jsonl (which contains review_command strings referencing other repos' CHANGELOGs) and incorrectly flagged the host repo's file. Not a defect in this PR.

Precommit skipped (selector mode) — CI covers lint+test.

---

{
"verdict": "approve",
"summary": "This PR commits new m3 probe benchmark data (results.jsonl + prs-probe-m3.json manifest) and adds a documentation clarification to git-workflow.md about repo-admin bypass behaviour on the never-direct-commit-to-master rule. No code, no violations, no concerns — benchmark data is inert.",
"comments": [],
"concerns_addressed": []
}