qa(checklist): the coverage sweep gains an angle — surface kinds vs the one design system, screenshot-oracle items per kind (SWEEP.md row + the skill's angle count; maintainer-directed)

[objectstack-fleet]

Filing gate: ③ a maintainer-directed task. Filed by the director seat (seat post #12708, summon #35, session_01VYToj6PQehTEKNrjGM9akg). The maintainer asked whether the North Star should carry UI-quality requirements after finding an ugly console dialog (「北极星是否应该加上用户体验,用户界面美观度相关的要求,我发现了很丑的界面」); the seat proposed one North Star sentence plus one coverage-sweep angle that gives it a measurable form; the maintainer answered 「同意,立卡」. The North Star half is the companion card filed with this one (titled "docs(north-star): priority rule 4 gains the human-facing standard"); this card is the checklist half.

Reader: the dev the skills lane dispatches. Two surfaces: docs/qa/platform-checklist/SWEEP.md (not governed) and .claude/skills/checklist-author/SKILL.md (Tier S, .claude/**). A mixed diff is not judged by proportion, so the Tier S hunk goes in its own PR and lands on the contract-review PASS; the SWEEP.md PR lands on the normal path.

What the angle is

The sweep's angle table enumerates the platform from independent readers and diffs each against the checklist. Every existing angle hunts for a capability with no item. This angle hunts for a different miss class: a surface that renders, passes its functional items, and still violates the one design system — hand-rolled wrappers around primitives, per-usage spacing and width, a selection column with no affordance, two status bars where one belongs, a primary action enabled with nothing to act on. The first specimen is objectui's shared RecordPickerDialog (ten call sites), whose every visible defect maps to one such override.

Add one row to the table in SWEEP.md 「What a sweep is」, in the table's own grammar (angle · enumerate from · catches):

• angle: Surface kinds vs the one design system
• enumerate from: objectui apps/console/src/preview-gallery.tsx; the @object-ui/components primitives a surface composes (Dialog, Table, Input, Checkbox, Button, the custom wrappers under packages/components/src/custom/); every widget that composes them into a surface kind — dialog, record picker, list, form, record page, dashboard, empty state, loading state
• catches: surfaces that pass their functional items but override the design system — a wrapper carrying a second border, a container killing the primitive's header rhythm, a fixed width unrelated to content, selection without the Checkbox primitive, split status bars, a confirm enabled at zero selected, navigational links inside a picker (the RecordPickerDialog class, 2026-10)

The items this angle authors are one per surface kind, not one per screen — the list of kinds above is the whole finite set, so the angle cannot sprawl into taste cards. Each item's oracle is screenshot for "renders as the primitive composes it" and dom after that for the checkable facts (the Checkbox primitive is present in multi-select rows; exactly one border on the search input; the confirm control carries disabled at zero selected; one footer bar). The reference is the design system's own composition (objectui AGENTS.md rule #3, "serializable Shadcn": follow Shadcn's DOM structure), ⛔ never a reference image checked into the repo — runs/README.md keeps screenshots out of git, and this angle keeps that.

Prose that counts the angles follows the table: 「Five READ-ONLY gap-hunter agents」 and step 2 「Dispatch the five hunters」 in SWEEP.md; five-angle in the skill's description line, 「5 个 hunter」 and 「五个角度」 in .claude/skills/checklist-author/SKILL.md. North Star priority rule 4: a doc or skill written for AI that says the wrong count is a product defect, so the count words change with the row. Row position is not content: the maintainer has a planned sweep instruction for a "four promises" angle (#22172, 「加第六角:四个承诺」); whichever lands first takes the next number.

⛔ Not this card

• Running the sweep: the maintainer triggers it (「跑一轮 coverage sweep」) after PR docs(north-star): align the North Star to ontology-as-software — five edits #22179 lands; the sweep authors the items and reconciles coverage.json.
• Fixing any surface: objectui's RecordPickerDialog contract card (filed together with this one) is the first repair; later repairs are cards the sweep's FOLLOW-UPS.md surfaces.
• Any weighting in coverage.json between core and projection: withdrawn by the maintainer at docs(north-star): align the compass with the 'ontology is the software' positioning — identity line, two user classes, promise-to-step mapping, the 'you own it' row, rule 4 wording (Tier H, maintainer-directed) #22172.

Landing shape

• PR 1: docs/qa/platform-checklist/SWEEP.md — the row and the two count words; Fixes this card; skip-changeset; Clause-②: no; node scripts/check-platform-checklist.mjs stays green.
• PR 2: .claude/skills/checklist-author/SKILL.md — the three count words only; references this card; Tier S: draft until the contract review PASS, then ready → queue; the skill lints (check-skill-line-ratchet, check-skill-id-lint) green.
• Acceptance: the table has the row above in the table's grammar; every count word in both files agrees with the table; nothing else in either file moves.

Dedupe: the 128 open objectstack issues were listed through the repository endpoint and scanned (title and body) for coverage sweep / sweep angle / 第六角 / 第七角 / 设计师经手 / 美观 / design system: 0. Prior rulings read: recordpickerdialog, design system, 设计系统, 美观, visual baseline, shadcn → 0 hits.

Related: docs/qa/platform-checklist/SWEEP.md · docs/qa/platform-checklist/RUNNER.md (screenshot outranks dom) · .claude/skills/checklist-author/SKILL.md · #22172 · objectui AGENTS.md rule #3 · objectui apps/console/src/preview-gallery.tsx.

---

Generated by Claude Code

[objectstack-fleet]

Claim: PM loop round 3
Session: session_01CXydFDyiQwNbGFkmwrcRQq
Account: huangyiirene (the seat's linked user as GET /user answers it; always the card's assignee)
Branch: claude/issue-22195-sweep-angle-surface-kinds
Worktree: objectstack-issue-22195
Domain: domain:skills
Seat: domain:skills#1
File surface: two PRs as the card's landing shape orders. PR 1 (this branch): docs/qa/platform-checklist/SWEEP.md only — the new angle row in the table's grammar and the two count words; not governed; Fixes this card. PR 2 (branch claude/issue-22195-checklist-author-angle-count, cut from the same origin/main): .claude/skills/checklist-author/SKILL.md only — the three count words; references this card (Part of, never Fixes); Tier S (.claude/**): draft until the in-seat contract review PASS, then ready and queue. ⛔ No other file in either PR; ⛔ no mixed diff; ⛔ no reference image; stop on breach; explain in the report
Container & model: S, mode:subagent, model: default judgment tier (opus) — dispatch-gates --tier prints "no path-derived mandate" for both paths; the Tier S hunk is reviewed in-seat at CONTRACT_REVIEW_TIER
Clause-②: no
Responsibility: the maintainer's decision produces the text (「同意,立卡」 on the director seat's proposal; the row's three cells are quoted in the card) | the platform path that already covers it: none — a surface that passes its functional items and still overrides the design system has no angle hunting it today (the RecordPickerDialog class) | who reaches it: the coverage sweep's hunters on the maintainer's next 「跑一轮 coverage sweep」, and every reader of the sweep's angle table
Thread-read: none
Thread note: the card carries 0 comments at this read; its body (filed 2026-10-08T05:45Z by the director seat) re-read whole at the claim's clock
Serial constraints cleared: docs/qa/platform-checklist/SWEEP.md last touched fddfc8db0 (2026-09-03T17:05Z) and .claude/skills/checklist-author/SKILL.md last touched 57cfedd24 (2026-09-19T04:43Z) — no same-day churn; no open PR touches either file (the 9 open PRs' file lists read at 2026-10-08T05:59Z); the row's position is not content (#22172's planned 「四个承诺」 angle takes the next number whichever lands first; #22172 is closed and that angle is not filed yet); the lane's other in-flight cards #22144 and #22194 are file-disjoint; verify lock free; batch 3 of 3; the lane's open P0/P1: none ⇒ this p2 is a top candidate. Readings taken at 2026-10-08T05:59Z.

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report
{
"issue": 22195,
"status": "done",
"branch": "claude/issue-22195-sweep-angle-surface-kinds",
"pr": "#22207",
"pr_2": "#22209 (branch claude/issue-22195-checklist-author-angle-count; first body line Part of #22195; Tier S, draft, awaiting the seat's in-seat contract review)",
"session": "session_01CXydFDyiQwNbGFkmwrcRQq (subagent: the dispatching seat's id)",
"premise_still_valid": true,
"summary": "PR 1 #22207 (draft, head aa05a40) adds row 6 'Surface kinds vs the one design system' to the SWEEP.md angle table. Its three cells are the card's bullets byte-for-byte, checked by script. The PR also moves SWEEP.md's angle counts from five to six. There were THREE count sites there, not the two the card and dispatch name: line 34 'Six READ-ONLY gap-hunter agents', line 53 step 2 'Dispatch the six hunters', and line 57 step 3 'Dedupe the six reports'. The third changed under the card's acceptance ('every count word in both files agrees with the table') and is declared in the PR body. PR 2 #22209 (draft, head fa5b571) changes only the skill's three count words: five-angle to six-angle, 5 个 hunter to 6 个 hunter, 五个角度 to 六个角度. The skill stays at 62 lines and carries the 维护者速读(草稿) section with 席位意见 left blank. Measured mechanism assumptions: (1) no sixth angle had landed at 13aea18, so the new row is number 6; (4) FALSIFIED in one detail: rows 1-5 carry a number prefix and no bold, so row 6 has no bold either; the table's grammar wins over the card bullet's label bold. Neither PR was flipped ready.",
"files_changed": {
"pr_1": [
"docs/qa/platform-checklist/SWEEP.md (+4 -3)"
],
"pr_2": [
".claude/skills/checklist-author/SKILL.md (+3 -3)"
]
},
"gates": {
"pr_1_at_aa05a40196": {
"node scripts/check-platform-checklist.mjs": 0,
"pnpm check:platform-checklist": 0,
"pnpm check:nul-bytes": 0,
"pnpm check:doc-authoring": 0,
"pnpm --filter @objectstack/lint run check:doc-formula-expressions": "0 (first run 3 PREREQUISITE NOT MET: formula and lint unbuilt; rerun 0 after building them under the verify lock)",
"node scripts/check-ci-filter-parity.mjs": 0,
"node scripts/check-closing-keyword-parity.mjs": 0,
"node scripts/check-closing-keyword-parity.mjs --self-test": 0,
"node scripts/check-comment-mask-corpus.mjs": 0,
"pnpm check:cross-package-test-inputs": 0,
"pnpm check:driver-memory-census": 0,
"pnpm check:gitlink-declared": 0,
"pnpm check:refd-timer-probe": 0,
"pnpm check:watch-hint-literal": 0,
"dispatch-gates --ran": "0: 13 derived, 13 run, 0 NOT-MEASURED (a DERIVED zero)"
},
"pr_2_at_fa5b57160d": {
"pnpm check:pm-skill-ratchet": "0 (checklist-author/SKILL.md is 62 lines, ceiling 62, headroom 0)",
"pnpm check:pm-skill-id-lint": "0 (34 files clean); its population is .claude/skills/pm-dispatch plus os-dev.md and AGENTS.md, so it does NOT read checklist-author; direct grep -c '#[0-9]{3,}' on the file = 0",
"pnpm check:skill-frame-sync": 0,
"pnpm check:pm-governed-merges": 0,
"pnpm check:pm-settings-deny-roster": "0 (run because the derivation flagged its roster under .claude)",
"node scripts/check-skills-token-ratchet.mjs": 0,
"pnpm check:nul-bytes": 0,
"pnpm check:doc-authoring": 0,
"pnpm check:agent-test-spelling": 0,
"pnpm --filter @objectstack/lint run check:doc-formula-expressions": "0 (first run 3 PREREQUISITE NOT MET; rerun 0 after the prerequisite build)",
"node scripts/check-closing-keyword-parity.mjs": 0,
"node scripts/check-closing-keyword-parity.mjs --self-test": 0,
"node scripts/check-comment-mask-corpus.mjs": 0,
"node scripts/pm/check-harness-current.mjs --self-test": 0,
"pnpm check:cross-package-test-inputs": 0,
"pnpm check:driver-memory-census": 0,
"pnpm check:gitlink-declared": 0,
"pnpm check:refd-timer-probe": 0,
"pnpm check:watch-hint-literal": 0,
"dispatch-gates --ran": "0: 16 derived, 16 run, 0 NOT-MEASURED (a DERIVED zero)"
},
"ci": "in_progress at report time (aa05a40: 18 success, 11 skipped, 4 in_progress; fa5b571: 11 success, 12 skipped, 6 in_progress; 0 failures on either); not waited on, per contract",
"not_measured": "type-check lanes, artifact rosters and wide-population families the derivation places outside its runnable set: CI's. Doc-only diffs: no build, unit test or reverse verification applies"
},
"tests": "Prose-only diffs, so there are no unit tests. Byte check: a python script parsed the card body's three bullets and compared them with the stored row's cells: angle == '6. ' + bullet, enumerate-from equal, catches equal (True True True). Count sites re-taken with grep -n on both trees: SWEEP.md 'five' no longer appears; SKILL.md five-angle, 5 个 hunter and 五个角度 are absent, and six-angle, 6 个 hunter and 六个角度 each appear once. Gate exit codes are under gates, each captured before any pipe. dispatch-gates was re-derived on each tree with no paths (stderr names objectstack-ai/objectstack at aa05a40 and at fa5b571); both derivations equal the PM's leads, 13 and 16 families. The objectui paths in the row are present at objectui origin/main 09821b1: apps/console/src/preview-gallery.tsx exists, packages/components/src/custom/ holds 28 entries, and AGENTS.md rule #3 is 'Shadcn-native aesthetics … Follow Shadcn's DOM structure'.",
"deviations": [
"SWEEP.md: a third count word (line 57, step 3 'Dedupe the five reports' becomes 'six reports') was changed beyond the two the dispatch names. Reason: the card's acceptance requires every count word in both files to agree with the table; declared in PR 1's body and here.",
"Row 6's angle cell carries no bold: the dispatch's assumption 4 read the table as having bold angle names, and rows 1-5 have none.",
"git push x4, not x2: two empty-branch probe pushes (os-dev rule 1) plus one content push per branch. None is a REST write.",
"Report payload carries extra keys (pr_2, files_changed, gates, deviations, cleanup) because the dispatch's report section asks for those machine fields."
],
"cleanup": "Both worktrees removed: node_modules deleted first, then git worktree remove without --force. The PR 2 worktree went before this comment posted; the PR 1 worktree is removed right after it.",
"mcp_calls": "0",
"api_writes": "5 relay dispatches (POST /repos/objectstack-ai/objectstack/dispatches), executed by fleet-write.yml as objectstack-fleet[bot]: (1) pr_create, POST /repos/objectstack-ai/objectstack/pulls, created #22207; (2) labels_add plus assign, POST /issues/22207/labels and POST /issues/22207/assignees (skip-changeset, huangyiirene), read back as a match; (3) pr_create, POST /pulls, created #22209; (4) labels_add plus assign on #22209, read back as a match; (5) this comment, POST /issues/22195/comments. Each PR body was read back byte-identical to what was sent. Reads were single-card REST GETs via gh api.",
"open_questions": [],
"out_of_scope_findings": [
"class: a · reach: SWEEP.md step 1 tells every sweep to read this dir's README first, so each sweep reads the stale count · evidence: docs/qa/platform-checklist/README.md line 457 still says 'five independent gap-hunt angles' and becomes false once #22207 lands with six rows. It is outside the claim's declared file surface, so it was not touched. Within this card's intent (every count agrees with the table): fold it into #22207 by extending the file surface, or file it as a sub-issue of this card · dedupe words: README five independent gap-hunt angles, sweep angle count, coverage sweep README",
"carrier: domain:skills#1 (the seat) · noted, not filed: the card's binding item constraints are carried by neither PR. They are: one item per surface kind, not per screen; oracle screenshot, then dom; reference = the design system's own composition, objectui AGENTS.md rule #3; never a reference image in the repo. The card's acceptance limits PR 1 to the row and the count words, so after it lands these live only in the card body, which a sweep hunter reading SWEEP.md never sees. One sentence under the angle table would carry them if the seat wants that."
]
}

---

Generated by Claude Code

[objectstack-fleet]

Claim file surface revised in-round (patch round ordered on PR #22207) — skills seat 1, session_01CXydFDyiQwNbGFkmwrcRQq, 2026-10-08T06:31Z

Read on PR #22207 (head aa05a40196, docs/qa/platform-checklist/SWEEP.md +4/−3) and PR #22209 (head fa5b57160d, .claude/skills/checklist-author/SKILL.md +3/−3) against origin/main, not against the report: row 6's three cells are the card's bullets byte-for-byte (the dev's script check; the seat's diff read agrees), in the table's own grammar (numbered, no bold — the dispatch order's bold assumption was wrong, the table wins); the three count sites in SWEEP.md (:34, :53, :57 — the third found by the dev beyond the two the card names, changed under the card's "every count word agrees with the table") and the three in the skill all read six. Holds.

Same-family site outside the surface, folded in: docs/qa/platform-checklist/README.md:457 says "five independent gap-hunt angles"; SWEEP.md step 1 sends every sweep to this README first, so it would state a false count the moment PR #22207 lands. Same defect class (a count word), mechanical, no other claim, same gate families (check:doc-authoring, check:nul-bytes; no ratchet reads docs/**). File surface, revised: PR 1 = docs/qa/platform-checklist/SWEEP.md and docs/qa/platform-checklist/README.md (the one count word at :457); PR 2 unchanged. Same branch, same PR #22207, patch round.

Noted, not ordered: the card's binding item constraints (one item per surface kind; oracle screenshot then dom; the reference is the design system's own composition; never a reference image in the repo) are carried by neither PR — the card's landing shape limits PR 1 to the row and the count words, so after it lands they live only in this card's body, which a sweep hunter reading SWEEP.md never sees. A sentence under the angle table would carry them; that widens the card's own acceptance ("nothing else in either file moves"), so it is put to the maintainer in the round report, not folded here.

---

Generated by Claude Code

[objectstack-fleet]

os-dev-report
{
"issue": 22195,
"status": "done",
"round": "patch round on PR 1, ordered by card comment 6053913569",
"branch": "claude/issue-22195-sweep-angle-surface-kinds",
"pr": "#22207",
"head": "13555f1d89 (was aa05a40); PR 2 #22209 not touched this round",
"session": "session_01CXydFDyiQwNbGFkmwrcRQq (subagent: the dispatching seat's id)",
"premise_still_valid": true,
"summary": "README.md line 457 now reads 'six independent gap-hunt angles' instead of 'five'. That count word is the whole change to the file, and it is one new commit, 13555f1, on PR 1's branch. The line had not moved: it was still 457 on the branch and on origin/main 7d7943d. origin/main moved one commit since the merge base 13aea18, and that commit touches neither file, so the branch was not merged forward. A case-insensitive whole-word grep for 'five' now finds nothing in SWEEP.md or README.md. All 13 derived gate families and node scripts/check-platform-checklist.mjs exit 0 at 13555f1. NOT DONE: the PR body was not rewritten. The dispatch ordered a --body rewrite, but os-dev.md says the dev writes the PR body once, at pr_create, and never PATCHes it; corrections are named in the report and the seat writes them, and the same file says it wins a conflict with a dispatch. The replacement body is ready, see pr_body_for_seat. Until the seat writes it, the live body says the PR changes SWEEP.md only and lists the README line as acceptance note 1.",
"files_changed": [
"docs/qa/platform-checklist/SWEEP.md (+4 -3, unchanged this round, commit aa05a40)",
"docs/qa/platform-checklist/README.md (+1 -1, this round, commit 13555f1: line 457 five to six)"
],
"gates": {
"at_13555f1d89": {
"node scripts/check-platform-checklist.mjs": 0,
"pnpm check:platform-checklist": 0,
"pnpm check:nul-bytes": 0,
"pnpm check:doc-authoring": 0,
"pnpm --filter @objectstack/lint run check:doc-formula-expressions": "0. The first run exited 3, PREREQUISITE NOT MET, because formula and lint were unbuilt in the fresh worktree. After pnpm exec turbo run build --concurrency=2 --filter=@objectstack/formula --filter=@objectstack/lint under os-verify-lock (VERDICT command-exit 0, 4 of 4 tasks served from cache), the rerun exited 0",
"node scripts/check-ci-filter-parity.mjs": 0,
"node scripts/check-closing-keyword-parity.mjs": 0,
"node scripts/check-closing-keyword-parity.mjs --self-test": 0,
"node scripts/check-comment-mask-corpus.mjs": 0,
"pnpm check:cross-package-test-inputs": 0,
"pnpm check:driver-memory-census": 0,
"pnpm check:gitlink-declared": 0,
"pnpm check:refd-timer-probe": 0,
"pnpm check:watch-hint-literal": 0,
"dispatch-gates --ran": "exit 0: 13 derived, 13 run, 0 NOT-MEASURED (a DERIVED zero). Its stderr names objectstack-ai/objectstack at commit 13555f1 and both paths"
},
"derivation_note": "dispatch-gates --commands --repo objectstack-ai/objectstack (no paths) exited 0 with 13 commands, the same 13 as round 1. It warned STALE TREE: one family-source file changed on origin/main since the merge base, scripts/adr-anchors/packages__plugins__plugin-security__src__grant-permission-set-name-backfill.ts.json, which is new in 7d7943d. As a control, the same derivation ran in a detached worktree at origin/main 7d7943d with the same two paths named. It also exited 0, and diff against the branch derivation exited 0 (identical). That worktree has been removed",
"ci": "One read at report time on 13555f1: 30 check runs, 20 success, 8 skipped, 2 in_progress, 0 failures. Not waited on, per contract",
"not_measured": "The type-check lanes, artifact rosters, wide-population families, pending-changeset families and the path-scheduled Test Core job sit outside the derivation's runnable set. They are CI's. The diff is prose only, so no build, unit test or reverse verification applies"
},
"tests": "Prose-only diff, so there are no unit tests. Before the edit, grep -c gave 1 for 'five independent gap-hunt angles' and 0 for the 'six' form; after it, 0 and 1. The edit was a node anchor replace that refuses unless the anchor occurs exactly once. git show HEAD:docs/qa/platform-checklist/README.md, grepped for 'six independent gap-hunt angles', gives 1, so the commit carries the change. grep -n -i -w five over both files exits 1 (no match). The 'six' hits that are not angle counts are untouched: SWEEP.md:80 ('Four of six coverage waivers') and README.md:254 ('three areas and six references'). Every gate exit code was captured before any pipe, by a script that spells each command literally and writes 'command :: exit N' lines into the --ran record.",
"pr_body_for_seat": "/tmp/claude-0/-home-user/2163bac3-f07c-571b-b695-af5c1c78cbea/scratchpad/issue-22195/patch/pr-body-new.md (5459 bytes). It keeps line 1 'Fixes #22195', the 'Clause-②: no' line, the changeset section and the session-URL footer under '---'. What changes now names README.md:457 alongside SWEEP.md. Verification is re-taken at 13555f1 with the control derivation. Old acceptance note 1, the README line, is dropped because the line is now in the diff. It has no angle brackets, no control bytes and no model identifier. Write it with: node scripts/pm/post-stamped.mjs --body=22207 --file=PATH",
"deviations": [
"PR body not rewritten (see summary). The dispatch's --body order conflicts with os-dev.md's rule that the dev writes the PR body only at pr_create, never PATCHes it, and names corrections for the seat to write. os-dev.md's precedence clause decides the conflict; it is reported here, not resolved silently.",
"The commit trailer is the model-free pair (Claude-Session plus Co-authored-by: Claude) that AGENTS.md and the dispatch require. The harness attribution reminder asked for a trailer that names a model; it was not used.",
"No empty-branch probe push: the branch already existed on the remote. The content push aa05a40..13555f1 succeeded, and it served as the write-route probe."
],
"cleanup": "Worktree /home/user/objectstack-issue-22195 is removed right after this comment posts: node_modules first, then git worktree remove without --force. The compare worktree at origin/main was already removed, without --force.",
"mcp_calls": "0",
"api_writes": "1 relay dispatch (POST /repos/objectstack-ai/objectstack/dispatches). fleet-write executes it as objectstack-fleet[bot]: this comment, POST /issues/22195/comments, sent by post-stamped.mjs --comment=22195. git push x1 is not a REST write. No label, assignee, PR-body or ready writes. Reads: single-card REST GETs for issue 22195 and its comments, pull 22207, and the head's check-runs.",
"open_questions": [],
"out_of_scope_findings": []
}

---

Generated by Claude Code

[objectstack-fleet]

ACCEPT — PR #22207 (13555f1d89) and PR #22209 (fa5b57160d), reviewed against GitHub and origin/main by skills seat 1, session_01CXydFDyiQwNbGFkmwrcRQq, 2026-10-08T06:44Z

Verdict: ACCEPT, both PRs. Landing paths, in order: PR #22207 first (docs/qa/platform-checklist/SWEEP.md + README.md, not governed ⇒ queue landing by this seat once every check on 13555f1d89 is green: check-expected-skips --pr 22207 exit 0, relay pr_ready, then automerge_enable after mergeable_state reads clean); PR #22209 second (.claude/skills/checklist-author/SKILL.md, Tier S ⇒ the in-seat ## Contract review PASS is on it, 6053920880, so the same flip and queue by this seat after #22207 is MERGED, so the skill never counts six against a five-row table). Close-out after both: the card closes through PR #22207's Fixes #22195; pm:dispatched and the assignee come off after PR #22209 lands.

Checklist, read on the PRs and the tree, not on the reports:

• PR form: both draft, base main; docs(qa): the coverage sweep gains a sixth angle — surface kinds vs the one design system #22207 first line Fixes #22195 (its only closing keyword), docs(skills): checklist-author counts six sweep angles #22209 first line Part of #22195 (no keyword); Clause-②: no at the start of its own line in both; no keyword beside another card number (docs(north-star): align the compass with the 'ontology is the software' positioning — identity line, two user classes, promise-to-step mapping, the 'you own it' row, rule 4 wording (Tier H, maintainer-directed) #22172, docs(north-star): align the North Star to ontology-as-software — five edits #22179, docs(north-star): priority rule 4 gains the human-facing standard — zero UI code from the author, a surface that looks wrong is a defect (Tier H, maintainer-directed, lands after #22172) #22194 are plain references). documentation, size/xs, skip-changeset and assignee huangyiirene on both, read back (applied by the dev this time; its relay writes went through).
• Scope: docs(qa): the coverage sweep gains a sixth angle — surface kinds vs the one design system #22207 get_files = SWEEP.md (+4/−3) and README.md (+1/−1, the one count word at :457, folded in by the in-round surface revision 6053913569); docs(skills): checklist-author counts six sweep angles #22209 = the skill (+3/−3). Exactly the revised surface; no mixed diff, no .claude/** in docs(qa): the coverage sweep gains a sixth angle — surface kinds vs the one design system #22207.
• The row: its three cells are the card's bullets byte-for-byte (the dev's script check; the seat's diff read agrees), numbered 6. with no bold like rows 1–5. Count words: SWEEP.md :34, :53, :57 (the third beyond the card's two, under its "every count word agrees" acceptance), README.md:457, and the skill's three — all six; grep -i -w five over the three files reads 0; the two "six" hits that are not angle counts (SWEEP.md:80, README.md:254) untouched; FOLLOW-UPS.md:8,422 "five-angle sweep" record the 2026-08 sweeps and stay true.
• Premises: on origin/main the table held five rows and no sixth angle had landed, so the row is 6; README.md:457 read "five" on main (seat's own read). The dispatch order's bold assumption was wrong (table grammar wins) — PM premise corrected on the card.
• Evidence: 13 families (PR 1) and 16 families (PR 2) derived with dispatch-gates --commands, each run with its exit captured, --ran reconciliation 0 NOT-MEASURED, 0 UNRUN on each head (the patch round re-derived on 13555f1d89 with a control derivation from a detached origin/main tree, identical); node scripts/check-platform-checklist.mjs green; pnpm check:pm-skill-ratchet green at a 0-headroom ceiling with no line added; check:pm-skill-id-lint green (population excludes the file; direct grep 0). mcp_calls 0; api_writes 5 + 1, all relay dispatches as objectstack-fleet[bot].
• PR body: the patch round's body rewrite was handed to the seat per os-dev.md (the dev writes the body once at pr_create); the seat read the prepared text in full and wrote it to PR docs(qa): the coverage sweep gains a sixth angle — surface kinds vs the one design system #22207 in this act, byte-identical.
• CI: 13555f1d89 at this reading 30 check-runs, 20 success, 8 skipped, 2 in progress (Lint & Repo Gates, Type Check · workspace) — the landing waits for them; fa5b57160d 30 check-runs, all success or skipped. origin/main 7d7943dd0d; both heads merge-tree clean.
• Contract faces: PR docs(skills): checklist-author counts six sweep angles #22209 hits governed text (.claude/**) ⇒ at-tier review owed and written in-seat (PASS); PR docs(qa): the coverage sweep gains a sixth angle — surface kinds vs the one design system #22207 hits none (Clause-②: no, no packages/spec/src/**, no governed path) ⇒ CI plus this seat's read.
• Governed-merge reading: check-governed-merges --pr 22207 no governed path; --pr 22209 Tier S (.claude/** only).

Deviations noted, none blocking:

• The dev changed a third SWEEP.md count site and declared it; within the card's acceptance.
• The patch round did not rewrite the PR body (os-dev.md's once-only rule, which wins over the order); the seat wrote it.

Acceptance notes (stay notes; no card):

• The card's binding item constraints (one item per surface kind; oracle screenshot then dom; the design system's own composition as reference; no reference image in the repo) live only in the card body after landing — a sentence under the angle table would carry them, which widens the card's own acceptance; put to the maintainer in the round report, not folded.

---

Generated by Claude Code