review-checklist: add the git-first spot-check row (#11365 half ③) — checklist sits at 82/82, pay by in-place compression

[os-zhuang]

Follow-up split out of #11365 (ruled option A on its report's open question, ACCEPT comment 5402161645) — the one half of that card its claimed file surface deliberately did not cover.

The row to add

.claude/skills/pm-dispatch/references/review-checklist.md gains one spot-check: when an os-dev report's tests/evidence shows API reads where git would have answered (issue-list sweeps re-run instead of treated as PM-supplied, MCP GraphQL list/search calls for things git log/local grep answer), flag it at ACCEPT time. The instruction side of this rule landed in PR #11870 (os-dev.md rule 3 + dispatch-runbook.md); this row is what makes a silent dev regression visible at review — the reading itself does not surface it, which is why "the PM already reads the evidence" was rejected as an argument for dropping the row.

The constraint that makes this its own card

review-checklist.md is at line-ratchet ceiling 82/82 (headroom 0) — the row must be paid for by in-place compression of that file, which is its own small editing task and was outside #11365's claimed surface (widening a claim mid-flight is the collision the claim protocol exists to prevent).

Refs: #11365 (the ruling and the ACCEPT) · PR #11870 (the instruction half) · #11364/#11366 (the sibling quota rulings, landed via PR #11749).

[claude]

Triage state repair (concentrated round, mechanical): finding means awaiting first grading, and this card is already graded into another pm-state — removing the stale finding label so the ungraded count stays honest. No other change.

---

Generated by Claude Code

[os-zhuang]

Claim — dispatching now (skills lane, PM seat session_01RMTpSRF5CjMmQBFfPtPCwJ).

• Branch: claude/issue-11872-review-checklist-gitfirst
• Worktree: objectstack-11872-gitfirst
• File surface: .claude/skills/pm-dispatch/references/review-checklist.md — single file, at its line-ratchet ceiling (82/82), so the new row is paid by compression, not a ceiling bump. File contention with [finding] The pm-skill ratchet counts LINES while its header prices a per-session token read — a green 134/134 run grew platform-readings.md ~33% in bytes #11106/PR pm-ratchet: cap ceilinged lines at 120 bytes, re-wrap the legacy corpus, re-pin the map #11948 is cleared (that PR merged).
• Basis: ruling A recorded in os-dev quota shaping: PM-supplied dedup reading rides the dispatch prompt; dev search-before-file goes through the REST channel — one dev measured ~5,658 GraphQL points in ~15 minutes #11365's ACCEPT comment (5402161645) — the checklist half of the git-first rule; the instruction half (os-dev agent + runbook) already landed in PR skills(pm): PM-supplied dedup reading rides the dispatch prompt; dev search-before-file goes REST #11870.
• Governed surface ⇒ draft PR, human merge. Clause-②: no.

---

Generated by Claude Code

[claude]

{
 "issue": 11872,
 "status": "done",
 "branch": "claude/issue-11872-review-checklist-gitfirst",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/11987",
 "pr_url": "https://github.com/objectstack-ai/objectstack/pull/11987",
 "premise_still_valid": true,
 "files_changed": [".claude/skills/pm-dispatch/references/review-checklist.md"],
 "ceiling": "84/84 (headroom 0, ceiling untouched) — the card says 82/82; that number went stale when the 120-byte max-line re-pin re-wrapped the corpus wholesale. Constraint identical: shrink-only, maintainer-only to raise.",
 "compression_summary": "Net 0 lines, two payments, both in place. (1) The two closing-keyword rows — the pre-ready body scan and the post-merge closed_by_pull_requests read — merge into one chronological row, 8 -> 5 lines: both carried the parser facts (ignores negation; body and commit message parsed separately) and the same pointer at the platform-readings table, now stated once. Every check survives; the PR body carries the row-by-row substance table. (2) The CI-convergence row re-wrapped at the corpus's own 120-byte width, 8 -> 7 lines, no word changed — proved byte-identical after whitespace normalization (688 vs 688 normalized bytes) with inline code spans identical in sequence. Ledger: -3 -1 +4 = 0.",
 "summary": "Added the git-first spot-check row to review-checklist.md: when an os-dev report's tests/evidence shows GitHub-API reads a local git command against the fetched branch would have answered (MCP list_issues/search_issues-class GraphQL reads, or re-running the dedup reading the dispatch prompt already supplied), the PM flags it at ACCEPT time. Explicitly not a REWORK — the ACCEPT still goes out, the flag is recorded in the ACCEPT comment so the pattern is visible and correctable, matching the instruction half already in os-dev.md + dispatch-runbook.md. Funded entirely by in-place compression; the ratchet ceiling in check-skill-line-ratchet.mjs is untouched.",
 "checks_run": [
 { "name": "check:pm-skill-ratchet", "result": "green — 'review-checklist.md is 84 lines (ceiling 84; headroom 0)' + 'self-test: 71 cases pass'" },
 { "name": "check:pm-skill-id-lint", "result": "green — 'check-skill-id-lint: 22 file(s) clean (pattern /#[0-9]{3,}/g)'" },
 { "name": "check:skill-frame-sync", "result": "green — 'self-test: 12 cases pass, plus 5 dispatch-gates declaration cases'" },
 { "name": "check:doc-authoring", "result": "green — '389 files clean' + '38 published skill files clean'" },
 { "name": "check:required-contexts", "result": "green — 'self-test: 150 assertions'; both legs exit 0" },
 { "name": "check:pm-governed-merges", "result": "green — 'self-test: 129 assertions'" },
 { "name": "check:agent-test-spelling", "result": "green — 'self-test: all cases pass'" },
 { "name": "check:doc-formula-expressions", "result": "green — '22 record-scoped formula example(s) across 421 files / 1448 TS blocks judged clean'" },
 { "name": "check:nul-bytes", "result": "green — 'OK (scanned 6658 text file(s) ... no raw ASCII control bytes)' + self-scan of the edited file, no match" },
 { "name": "pnpm lint (repo-wide eslint)", "result": "DECLARED NARROWING, not run — empty by construction: none of the 7 config blocks in eslint.config.mjs carries a files pattern matching markdown, so the one path in this diff is outside eslint's governed population; with no type-aware linting reaching a .md file the diff cannot move a verdict on an untouched file either. CI runs the farm regardless." },
 { "name": "CI gate jobs", "result": "in_progress — reported at draft-PR time per contract; convergence reading belongs to the PM" }
 ],
 "tests": "Gate family derived from the real change set, never recalled: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (1 path, 9 families, --repo assertion held against this checkout's origin). Union re-run AFTER the final commit at 6f7dfe58d — all 9 green plus check:nul-bytes, each exit code captured before any pipe (redirect-then-tail), each result quoted from the gate's own verdict line. Compression proof for the re-wrapped CI row: 'CI row byte-identical after whitespace normalization: true | normalized bytes: 688 vs 688' and 'inline code spans identical in sequence: true', 8 -> 7 lines. Every new and edited line passes the 120-byte cap — the replacement text was produced by the ratchet's OWN wrapLine, so the wrapping is the canonical form the gate would compute. No ablation applies to this card.",
 "open_questions": [],
 "out_of_scope_findings": [],
 "notes": "Card-number drift, worth a triage line rather than a card: the issue body pins the ceiling at 82/82 and the file on origin/main is at 84/84 — the 120-byte max-line rule's wholesale re-pin moved eighteen ceilings after the card was written. The instruction (pay by compression, never raise) is unaffected and was followed literally; only the arithmetic in the card is stale. PR is draft on purpose — .claude/skills/** is a governed surface, human merge. skip-changeset label added via the additive endpoint and READ BACK after the size-labeler settled: labels are ['documentation', 'size/s', 'skip-changeset'] (the first two are other actors' writes, left alone). Worktree removed without --force; deps installed and the two prerequisite package builds (@objectstack/formula, @objectstack/lint, needed by check:doc-formula-expressions) ran through scripts/pm/os-verify-lock.sh, each printing its own VERDICT line."
}

Generated by Claude Code

---

Generated by Claude Code

[os-zhuang]

ACCEPT — PR #11987 (draft; governed surface ⇒ human merge, joins the maintainer's merge list).

Review of record — verified against the fetched branch (git diff origin/main..., line counts re-derived locally), not the report's self-description:

• The row is the ruled one. Flag-at-ACCEPT, explicitly not a REWORK ground; criteria = MCP GraphQL-class reads (list_issues/search_issues) or re-running dispatch-supplied dedup readings where a post-fetch local git answers file contents, diffs, history, branch state. This is the checklist half of ruling A recorded in os-dev quota shaping: PM-supplied dedup reading rides the dispatch prompt; dev search-before-file goes through the REST channel — one dev measured ~5,658 GraphQL points in ~15 minutes #11365's ACCEPT (5402161645); the instruction half landed in PR skills(pm): PM-supplied dedup reading rides the dispatch prompt; dev search-before-file goes REST #11870. The two halves now reference the same read order (git → REST → MCP/GraphQL).
• Compression audited row by row, nothing lost. The two closing-keyword rows merged into one chronological row (8→5): all nine substance items survive — pre-ready body scan, the ⛔ never-beside-an-open-card rule, negation-sentence-is-the-misclose, patroller-only-walks-open-PRs, post-merge closed_by_pull_requests read, right-card-and-no-other-card confirmation, parser-ignores-negation, body/commit-parsed-separately + readings-table pointer, misclosed-hides-as-completed. The CI-convergence row re-wrapped 8→7 with whitespace-normalized byte identity (688=688) via the ratchet's own wrapLine. Ledger −3 −1 +4 = 0; my independent count: 84 → 84 lines, one file, ratchet registry untouched.
• Ceiling-number drift noted, correctly handled: the card said 82/82; pm-ratchet: cap ceilinged lines at 120 bytes, re-wrap the legacy corpus, re-pin the map #11948's wholesale re-pin moved it to 84/84 after the card was written. The constraint (pay by compression, never raise) was followed literally — this is the right reading of a stale number, no card amendment needed.
• Git-first self-check on this very report (the new row, applied to its own delivery): evidence shows local git/gates for all repo reads; the only API reads are label read-back after the size-labeler settled — state git cannot answer. Clean, nothing to flag.
• CI: the filter job on head 6f7dfe58d died environmentally (other side closed while paths-filter fetched the changed-file list, before any path judgement); a single rerun is scheduled once the run completes. Convergence reading stays on the review side per the checklist's own rule; ready-flip and merge remain the maintainer's.

Landing todo recorded: on merge, this card closes via the PR's Closes #11872; pm:dispatched comes off in the same stroke.

---

Generated by Claude Code