docs(pm-skills): adopt eight measured platform readings into the governed fact table, each paid in place

[os-justin]

Part of #18426 — the carrier of a seven-card batch, all landing rows in one governed file.

Clause-②: no

One file: .claude/skills/pm-dispatch/references/platform-readings.md, held at 466 / 466 (headroom 0,
widest line 120 B). No ceiling was raised. Every adopted row is paid by a compression or a dedup in the same
file, listed beside the sentence it shortened. Rows :10–:12, :212 and :431 are untouched — verified by content,
and none of them appears in any hunk of the diff.

Each reading was adopted or rejected on its own, as the cards ask.

Verdicts

card	reading	verdict	row line	bytes	paid by
#18426 ①	zod v4 defines the parse methods per INSTANCE, so prototype spying throws; the seam is the barrel population	ADOPTED	:277 + :278	97 B + 116 B	:172 — the conditional-request row absorbs the payload-precedence row (its 「只有 body 精确」 is restated at :178–:179) · :185 — the payload-staleness pair
#18426 ②	the REST write identity is per-TOKEN, not per-channel; read back your own write	REJECTED	already at :126–:132	—	—
#18426 ③	os-verify-lock.sh is objectstack's and absent in objectui	ADOPTED	:393	112 B	:207 — the created-order pagination pair
#18426 ④ (comment)	ref deletion is refused while every other write on the token succeeds	REJECTED	already at :374–:377 and :430	—	—
#18426 ⑤ (comment)	a suspended author's comment is counted but not served	REJECTED	already at :250 (cause at :164–:167)	—	—
#18446	a label strip re-triggers the label-sensitive checks ⇒ pre-check ③ is read after the last state write	ADOPTED	:85	118 B	:289 — the missing-command pair
#18458	gh-readonly-queue/* ref names embed the stacked-on sha ⇒ queue ORDER from ls-remote	REJECTED	already at :70–:73	—	—
#18467 (main)	a full rule-plus-footer BLOCK to an issue body is re-anchored, not doubled	REJECTED	already at :410 + :357	—	—
#18467 (operating half)	that re-anchor is what post-stamped's body mode reports as mutated	ADOPTED	:357 (new row under it)	119 B	:318 — the MCP-truncation pair
#18467 (addendum)	bare REST POST /issues with a footerless body synthesises exactly one, +58	ADOPTED	:410, refined in place	120 B	zero — folded into the existing create row
#18498 row 1	grep -z with a literal newline is an OR — a false POSITIVE; grep -Pz with the backslash-n escape discriminates	ADOPTED	:291	118 B	:451 — the usage-wall pair
#18498 row 2	(withdrawn on the card as a duplicate)	not adopted	—	—	—
#18521	unstable can come from a commit STATUS the /check-runs enumeration cannot see	ADOPTED	:28	115 B	:303 — the two CI-log reading rows
#18522	the footer has two spellings, so a bare-only strip strands the session-URL one mid-body	ADOPTED	:347	106 B	:331 — the fence caveat folds into the trap row beside it

8 readings adopted, 9 lines, 8 payments (the create-channel cell costs nothing — it refines a row that was
already there). Net line change: 0.

Every payment, and what it shortened

payment	before	after	fact lost
:172	「文档载明未实测:条件请求答 304 不计 core 池。」 + 「公开仓零配额读法两档,payload 档优先 —— 只有 body 精确。」	one row	「只有 body 精确」 — restated in full at :178–:179 (assignees stale, claim-bearing fields never read from this channel)
:185	「⇒ 前 15 项那条判别式…」 + 「失效方向是据它对别人的工作下没有认领的判词。」	one row	none
:207	「created 序承重…」 + 「以两趟的重叠证覆盖…」	one row	none
:289	「⛔ 一般规则:任何可能不存在的命令…」 + 「PR 与查重类核验改走 MCP GitHub 工具或 git。」	one row	none
:303	「get_check_run 对本仓 CI job 回空 output.text。」 + 「get_job_logs 无论 tail_lines 只回占满日志尾部的 post-step service-container teardown。」	one row	「post-step」 and 「占满」 — the position is still given by 「尾部的」
:318	「GitHub 存储字节完好,raw REST 取回一字不差。」 + 「⇒ ⛔ 永不单凭 MCP 读判截断…」	one row	none
:331	「⛔ 围栏不防护:照删、留空围栏。」 + 「另一坑:感叹号紧跟左方括号即触发(TS 非空断言下标),无需尖括号。」	one row	「TS」 and 「下标」 from the parenthetical — the mechanism is already spelled out by 「感叹号紧跟左方括号」
:451	「盲区:窗口起点是本地推断的近似值。」 + 「撞墙报文形如 limit reached, resets at HH:MM…」	one row	none

Why the five rejections — the four axes

All five reject for one reason: the reading is already in the fact table, measured and phrased. That makes
this a single class, so the axes are argued once and the per-card evidence follows.

• 实际业务需求 — the need each card names is real, and it is already served. A seat asking 「who authors my
  REST writes」 lands on :129 (「类只认每次写回读的 user.login/user.type」); a seat asking 「where is my PR in
  the queue」 lands on :70–:73, which carries the parent-tip chain recipe and a caution the card does not
  have (:72: the embedded parent is a presumed base, measured to reach main ten-plus minutes later). Adding a
  second row would not add a reader path; it would add a second place to disagree.
• 项目长远合理性 — a fact table with two rows for one fact is the workaround shape this repo refuses. When one
  copy is later corrected, the other stays, and the next seat reads whichever it reaches first. One row, one
  reader path.
• 防 AI 写代码犯错 — this is the axis that decides it. A duplicated row is a lenient surface: it lets a
  future edit narrow one copy while the other keeps the old claim, and nothing goes red. The strict form is one
  row, which the ratchet already prices. Note the direction of the rejection: none of the five loses a fact —
  each is being kept, in the line that already carries it.
• 创业阶段不扩散需求 — at 466 / 466 every row is paid by deleting or compressing another. Paying real
  compression for a fact that is already written down spends the budget twice and shrinks the table's coverage by
  exactly the sentence that was compressed to make room.

Per-card evidence:

• finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 ② — :126 names both token classes (claude[bot] and the account), :127 says the class is per
  Claude Code account and can flip mid-session, :129 says the class is only known by reading back each write,
  :132 says never assume claude[bot]. :128 already lists two measured flips. The card's third measurement
  (os-sam vs hotlong) confirms a rule the table states; it does not change it.
• finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 ④ — :375 records the git push --delete disconnect with the misleading tail, :374 and :376
  record the consequence (probe branches accumulate), and :430 records the second form, the proxy 403, with the
  same disposition and an explicit 「⛔ 不重试」.
• finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 ⑤ — :250 is the reading: 「comments 计数会多读:实测 1 而列表与 timeline 双 0 ⇒ 线程空否读那两条,
  ⛔ 不读计数」. The cause is at :164–:167. ⚠️ Worth the seat's eye: the triage grading on finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 asked
  specifically for this one to be written in, and it already was.
• platform-readings: gh-readonly-queue ref NAMES embed the stacked-on sha, so git ls-remote alone reconstructs the merge-queue order (dropped vs merely waiting) #18458 — :70 gives the name form and says the embedded sha is the previous entry's tip; :71 gives the
  walk and says lexical order is meaningless; :73 says it is zero-quota and answers the position that the
  timeline cannot. :61 keeps the positive-only caution the card asks to preserve.
• [finding] a full --- + footer BLOCK sent to an issue body through bare REST is RE-ANCHORED, not doubled — one newline moves, the net is zero bytes, and post-stamped reads the whitespace as mutated #18467 (main) — :410: 「issue 正文 PATCH 识别按整块:送全块或不送页脚都存回恰一条,已有页脚归一末尾不
  复制」 is the not-doubled half; :357 is the blank-line re-anchor and the instruction to compare by first
  differing offset rather than by length. What was missing is the operating consequence, and that is the row
  adopted at :357.

Measurements taken for this PR

• platform-readings: grep -z with a literal newline is an OR, not a wrapped-phrase match — the discipline line manufactures false confirmations #18498's restatement sweep. The dispatch asks where else the -z advice is restated, so the same edit can
  ride the same PR. Measured repo-wide, exit code captured before any pipe: git grep -F 'grep -z' → exit 1,
  git grep -F -- '-Pz' → exit 1, git grep -F '行向' → exit 1. Lit control in the same batch:
  git grep -F '拼接后' → exit 0, hitting platform-readings.md:290, and grep -c grep on that file → 13.
  ⇒ the -z advice is restated nowhere in this repo — it lives in the dispatching seat's own prose, handed
  down per round. Nothing in SKILL.md or core-rules needed the co-edit, so nothing is reported for re-scoping.
  The row is placed under pit ⑥ (:290), which is where this repo's wrapped-phrase discipline actually lives.
• [finding] stripping the needs:contract-review carrier re-triggers Check Changeset — so landing pre-check 3 must be read AFTER the strip, never before #18446's second half is NOT line-neutral in contract-review.md. That file is at 60 / 60. Its landing
  pre-check block :40–:46 runs 99–120 B per line, and the cheapest adjacent merge inside it is 213 B —
  no insertion fits without a raise. ⇒ per the dispatch's own instruction for that case, the row at :85 carries
  the sequence: 「③ 在入队前末次状态写之后读」. contract-review.md is untouched, still 60 / 60.
• [finding] stripping the needs:contract-review carrier re-triggers Check Changeset — so landing pre-check 3 must be read AFTER the strip, never before #18446 is recorded in its generalised form, per the card's own addendum: a label strip and a ready flip are
  both state writes. :84 already recorded the ready-flip re-run with a different consequence (enqueue timing);
  the new row states the consequence for pre-check ③.

Gates

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived the list from the tree
(no hand-fed paths); --ran reconciles: 18 derived, 18 run, 0 NOT-MEASURED, 0 UNRUN — a DERIVED zero, all 18
carry an exit code and none of them is 3. Every code captured by redirect-then-$?, never through a pipe.

command	exit
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-governed-queue-guard.mjs --self-test	0
node scripts/pm/check-harness-current.mjs --self-test	0
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0
pnpm check:agent-test-spelling	0
pnpm check:doc-authoring	0
pnpm check:driver-memory-census	0
pnpm check:nul-bytes	0
pnpm check:pm-governed-merges	0
pnpm check:pm-half-states	0
pnpm check:pm-skill-id-lint	0
pnpm check:pm-skill-ratchet	0
pnpm check:refd-timer-probe	0
pnpm check:required-contexts	0
pnpm check:skill-frame-sync	0
pnpm check:watch-hint-literal	0

Named explicitly by the dispatch, both inside that list:

• pnpm check:pm-skill-ratchet → exit 0; its own verdict lines:
  platform-readings.md is 466 lines (ceiling 466; headroom 0) and
  platform-readings.md: widest table row is 0 bytes (pin 0; headroom 0);
  contract-review.md is 60 lines (ceiling 60; headroom 0).
• pnpm check:skill-frame-sync → exit 0: the one declared copy of the decision frame is internally coherent,
  4 axes, 74 markdown files scanned for undeclared copies.

Governed-surface predicate, as the dispatch asks it be recorded:

• node scripts/pm/check-governed-merges.mjs --test .claude/skills/pm-dispatch/references/platform-readings.md
  → exit 3 = GOVERNED (not a failure): 1 of 1 path(s) hit the register, .claude/** ×1.

First eight of those ran at exit 3 = PREREQUISITE NOT MET on a fresh worktree with no node_modules; that is
not a measurement, so pnpm install ran and every one was re-run. check:doc-formula-expressions then still
answered 3 (@objectstack/formula and @objectstack/lint unbuilt); its own prescribed fix was run through the
shared verify lock — os-verify-lock: VERDICT command-exit 0 — and the gate re-ran at 0.

Manual byte discipline beyond the gates: grep -naP for control characters over the edited file → exit 1
(none).

Acceptance notes

Noted while measuring, not filed, and not touched here:

• platform-readings.md:412 states the issue-body PATCH cell has two empties (MCP with a bare footer; a
  single-field issue_write on title/labels). Still true after this PR — the adopted create-channel cell is
  a different verb. Successor: the next seat that writes an issue body through MCP.
• scripts/pm/post-stamped.mjs's docblock says 「whether the platform synthesises a footer for a footer-less ISSUE
  BODY is unmeasured」, which [finding] a full --- + footer BLOCK sent to an issue body through bare REST is RE-ANCHORED, not doubled — one newline moves, the net is zero bytes, and post-stamped reads the whitespace as mutated #18467 shows is stale prose — the table has carried that cell since PR docs(pm): eleven platform readings into the fact table — 23 rows, 3 in-place corrections, ceiling 425 to 448 #17746, and
  this PR adds the create half. The tool and the table now disagree in writing. ⛔ Out of scope here: the file is
  not on this PR's declared surface, and the card itself declines to judge whether one observation is enough to
  extend the tool's declared normalisation set. Successor: whoever next touches post-stamped.mjs.
• [finding] platform-readings: the list_issues labels filter is OR and truncates (can read as an empty lane), and auto_merge's stored merge_method is not always merge #18461 is deliberately absent from this batch (its row ② contradicts [finding] the seat's arming channel forces merge_method: merge and silently drops a requested squash — the readback says "merge" while the queue lands a 1-parent squash, and nothing tells a seat that #18421's reading; that decision is the
  maintainer's). Nothing here touches it.

维护者速读(草稿)

改了什么 —— 只改一个文件:PM 车道的平台读数事实表。八条今天实测出来的 GitHub/工具行为写进去,五条因为
表里已经有了而没写。文件行数一行没涨(466 / 466),每加一行都在同一个文件里删掉或压掉一行来付账。

为什么改 —— 这些读数原本只存在于各个 issue 的讨论里。没有载体的读数,下一个席位会从零重新推一遍;这次
的批次本身就是证据:五条里有五条已经在表里了,而填报席没查表就填了卡。

风险与代价(含回滚) —— 风险在八处压缩:两行合成一行,可能丢掉措辞。逐条列在正文的「Every payment」表里,
三处明确写了丢掉哪几个字以及那个事实在哪一行还能读到,其余五处无损。回滚是 git revert,单文件、无生成物、
无下游消费者。

席位意见 —— (留空,由席位在复核时填)

你要做的 —— 不需要做什么。这是事实层(references/**),按裁决 #17950 由技能席在自己的契约档复核后经队列
落地;⛔ 没有席位把它翻 ready、入队或挂 auto-merge。

Closing lines

Part of #18426 — readings ① and ③ adopted; readings ②, ④ and ⑤ rejected as already recorded, so the card stays
open for the seat to confirm those three duplication claims before it is retired.
Fixes #18446
Part of #18458 — the entire reading is already recorded at the queue-ref rows; no row was added, and the card
stays open for the seat to confirm that before it is retired.
Part of #18467 — the operating consequence and the create-channel cell are adopted; the re-anchor itself is
rejected as already recorded, so the card stays open for the seat to confirm that one.
Fixes #18498
Fixes #18521
Fixes #18522

---

Generated by Claude Code

[os-justin]

Contract review

Served-tier: CONTRACT_REVIEW_TIER
Head-sha: 5f809c1817b67f31dca0438b5677e0d35bfcb9aa

Review of record on the PR thread — the same record the seat posted on the carrier card as 5716694216 at 2026-09-17T15:10:43Z; the merge-group leg of the Governed Surface Guard reads the PR thread, not the card, and dequeued this PR at 2026-09-17T15:36:18Z for want of it. In-seat review of PR #18689 (the seven-card platform-readings batch carried by #18426) by the dispatching domain:skills seat, 2026-09-17T15:38Z, read against the PR's diff, files and head on GitHub and a scratch worktree at the head — ⛔ not against the dev report (5716488532). Clause-②: no on the PR body and on the claim (5716076148); check-clause2-carriers --pair 18689 answers 0 for all seven cards. Fact-layer landing: the one file is .claude/skills/pm-dispatch/references/platform-readings.md, so check-governed-merges --test answers GOVERNED (exit 3, .claude/** ×1) and the seat lands it on this review of record, not on a maintainer approval — the references/** carve-out the SKILL names.

① Derived judgments

• Scope = the claim: one file, +17 / −17, 466 / 466 (check:pm-skill-ratchet exit 0 on the head: ceiling 466, headroom 0, widest row 120 B; check:skill-frame-sync exit 0), rows :10–:12, :212 and :431 absent from every hunk. No ceiling raised; every adopted row paid by a compression in the same file.
• The nine adopted lines read on the head, each a measurement with its ⇒: :28 (unstable from a commit STATUS ⇒ read /commits/{sha}/status), :85 (label strip and ready flip are state writes ⇒ pre-check ③ after the last one), :277–:278 (zod v4 per-instance parse ⇒ wrap the barrel exports), :291 (-z literal newline is an OR ⇒ grep -Pz with the escape), :347 (two footer spellings), :357 (the re-anchor is what post-stamped --body reports as mutated, net zero), :393 (os-verify-lock.sh lives only in objectstack), :410 refined in place (bare POST /issues synthesises exactly one footer, +58). Placements sit under the pit each belongs to.
• The five rejections confirmed against the head, line by line: finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 ② at :126–:132 (token class per account, flips mid-session, read back user.login/user.type on every write, ⛔ never assume claude[bot]) — holds; finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 ④ at :374–:377 and :430 (probe branches undeletable, proxy 403, ⛔ 不重试) — holds; finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426 ⑤ — the row is at :249 on the head (:250 on the base; the PR body cites base numbering and the :207 payment above it shifted it by one): 「comments 计数会多读:实测 1 而列表与 timeline 双 0 ⇒ 线程空否读那两条,⛔ 不读计数。」 with the cause at :166–:167 (suspension destroys the identity's comments) — holds; platform-readings: gh-readonly-queue ref NAMES embed the stacked-on sha, so git ls-remote alone reconstructs the merge-queue order (dropped vs merely waiting) #18458 at :70–:73 (name form, parent-tip walk, lexical order meaningless, the :73 caution the card lacks) — holds; [finding] a full --- + footer BLOCK sent to an issue body through bare REST is RE-ANCHORED, not doubled — one newline moves, the net is zero bytes, and post-stamped reads the whitespace as mutated #18467 main at :410 + :357 — holds. None of the five loses a fact; each is kept in the line that already carries it.
• Payments spot-read: :172 folded (the payload-precedence sentence's 「只有 body 精确」 is carried in substance by :178–:179 — assignees stale, claim-bearing fields never read from that channel); :303 (the two log-reading rows into one, 「尾部的」 keeps the position); :331 (the fence caveat and the ![ trap on one row, the mechanism still spelled). The PR body's table lists all eight with what each lost; nothing operating was dropped.
• Q1 (closing lines) — ruled A: Part of stays on finding(platform-readings): three readings owed to the governed fact table — zod v4 parse-spying, a per-TOKEN (not per-channel) REST write identity, and a depth-wait command that is not repo-local #18426, platform-readings: gh-readonly-queue ref NAMES embed the stacked-on sha, so git ls-remote alone reconstructs the merge-queue order (dropped vs merely waiting) #18458 and [finding] a full --- + footer BLOCK sent to an issue body through bare REST is RE-ANCHORED, not doubled — one newline moves, the net is zero bytes, and post-stamped reads the whitespace as mutated #18467; Fixes stays on [finding] stripping the needs:contract-review carrier re-triggers Check Changeset — so landing pre-check 3 must be read AFTER the strip, never before #18446, platform-readings: grep -z with a literal newline is an OR, not a wrapped-phrase match — the discipline line manufactures false confirmations #18498, platform-readings: mergeable_state: unstable can come from a commit STATUS that the /check-runs enumeration cannot see #18521, platform-readings: the PR-body footer has two spellings, so a 不送页脚 strip that matches only the bare one strands the session-URL one mid-body #18522. After the merge the seat retires the three Part of cards by hand, each with a record citing the row(s) that carry its reading — a wrongly closed card is invisible to every open-only filter (:442), so a duplication claim closes a card only after the seat has read the cited line, which this record does.
• Q2 (item ⑤'s sharper row) — B taken, and the answer is no new row: the seat re-read :249 and :166–:167 against comment 5696410541. The row states the measurement (count 1, list and timeline 0) and the operating rule (judge emptiness from the two lists, ⛔ not the count); :167 names the suspended-identity cause. What the comment adds beyond that — the mismatch as a signal that a ruling may once have existed — is scoped by its own author to one issue with the deleted-comment case untested. At headroom 0 a row costs a compression, and a rule whose general form is unmeasured is not worth one; it stays in the card thread, reachable from :249's own subject.

② Semver level

None — no package is touched; skip-changeset on the PR is correct for .claude/**.

③ Boundary flags

• Dev's open_questions: two, both ruled above. Stop conditions: none fired.
• Out-of-scope findings: (1) :412's two empty cells on the issue-body PATCH row — noted, not filed, a measurement for the next seat that writes an issue body through MCP; (2) post-stamped.mjs's header still calls the issue-body footer cell unmeasured while :357/:410 measure it — raised by this dev and by the [finding] post-stamped --body exits 0 on a write the platform did NOT store — the read-back catches it, the exit code does not #18663 dev alike, filed once by the seat as [finding] post-stamped.mjs's header declares the issue-body footer cell UNMEASURED while platform-readings carries the measurement — every seat-post refresh prints a MUTATED warning #18693 (bare finding); (3) the dispatch's own scope text asserted that platform-readings: gh-readonly-queue ref NAMES embed the stacked-on sha, so git ls-remote alone reconstructs the merge-queue order (dropped vs merely waiting) #18458's and [finding] a full --- + footer BLOCK sent to an issue body through bare REST is RE-ANCHORED, not doubled — one newline moves, the net is zero bytes, and post-stamped reads the whitespace as mutated #18467's readings were unrecorded, and both were — the seat's dispatch prose repeated the triage premise without grepping the table; taken as the seat's own correction: a 「not recorded」 in a dispatch is a claim the seat greps for first.
• Gates: 18 derived, 18 run, 0 NOT-MEASURED, all exit 0 per the report; seat-side on the head: ratchet 0, frame-sync 0, over-120-byte lines 0, --pair 0 ×7. CI on the head at review time: 17 success, 12 skipped, Lint & Repo Gates still in progress; the landing pre-check re-reads every run completed and none failure before the ready flip.
• Writes: api_writes 9 (draft PR, one label via label-write, this report, six pointer comments on the sibling cards as the dispatch asked), mcp_calls 0, PR body written once at create — within budget.
• Independence pair: mode:subagent build, in-seat review by the dispatching seat at CONTRACT_REVIEW_TIER — by design per contract-review.md, not SELF-REVIEW.

Implemented-by: claude/issue-18426-platform-readings-batch
Reviewed-by: session_01Gqi43smmqjJ5sUrhfoPeKu

VERDICT: PASS

---

Generated by Claude Code