docs(pm-skill): the fifth readings increment — 26 measured platform facts

[claude]

Fixes #15977
Fixes #16006

Refs #16104 (section 1 only — its sections 2 and 3 stay open on that card for their own split)
Refs #15803 (section C only — that card's section A is in the decision box and its labels are untouched here)
Refs #16386

The fifth increment into .claude/skills/pm-dispatch/references/platform-readings.md, one flight over four carriers. Every candidate fact was deduplicated against the ledger as PR #15955's head leaves it (04b61b5f6a, 362 lines) rather than against origin/main (359 lines), because that PR's three fourth-increment lines are not on main yet. Facts stated by two carriers land as one line with both carriers named. Nothing was deleted and nothing was re-wrapped.

Ruling executed

The ceiling decision on #16386 was ruled A — director seat, decision batch #62, 2026-09-07, maintainer verbatim and untranslated: 「同意」 — in #16386 (comment). Its words: the ledger's line ceiling "rises to the measured count", 385 on that day's main, "or 388 if PR #15955's ruled raise (362) lands first — the same +26 either way". PR #15955 landed first, so this round executes the 388 leg.

This patch round does exactly two things, and nothing else.

1. origin/main merged in as a merge commit (never a rebase, never a force-push), which brings PR docs(pm-dispatch): three measured platform readings into the readings ledger #15955's ruled ceiling of 362 and its own ruledRaises record. The ledger file itself is untouched by this round; it is now 388 lines on the branch — 362 plus the same 26 measured lines the candidate table below accounts for.
2. scripts/pm/check-skill-line-ratchet.mjs, +15 −1: the CEILINGS row for .claude/skills/pm-dispatch/references/platform-readings.md moves 362 → 388 (before → after), and one ruledRaises record is appended. No other line in that script, no other ceiling, and nothing in the ledger file.

The appended record, as it stands in the map:

{
  ruling:
    'the fifth-increment raise on'
    + ' `.claude/skills/pm-dispatch/references/platform-readings.md` — maintainer,'
    + ' decision batch #62, 2026-09-07, verbatim and untranslated: 「同意」 on option'
    + ' A, and the sentences that ruling adopts, in the director record\'s own words'
    + ' (comment 5563909225): "`references/platform-readings.md`\'s line ceiling rises'
    + ' to the measured count: 385 on today\'s `main` (359 → 385), or 388 if PR'
    + ' #15955\'s ruled raise (362) lands first — the same +26 either way."',
  date: '2026-09-07',
  delta: 26,
}

The declared cross-file move is untouched by the raise: the destination now reads +11 (314 → 388, less 63 lines of ordinary ruled raise — 34 + 3 + 26) against the sources' unchanged net decrease of 20, which is exactly the +11 the move landed with.

Reverse verification — the raise is paid by the RECORD, not by the number

Every exit captured before any pipe.

leg	exit	reading
live run, pre-patch head 65a6c814d4 (after the merge)	1	✗ … platform-readings.md is 388 lines; the ratchet ceiling is 362. — 1 of 1 failing assertions
live run, patched head 34f19d1b89	0	✓ … platform-readings.md is 388 lines (ceiling 388; headroom 0). — 0 failing assertions
--self-test, patched head	0	✓ check-skill-line-ratchet self-test: 155 cases pass., the last of them ✓ every live ruled-raise record quotes its ruling, dates it, and names a positive line count
the new record deleted in a scratch copy, ceiling left at 388	1	✗ … raises its ceiling 314→388, less 37 lines of ordinary ruled raise (+37) against a net source decrease of 20 and ✗ … the declared cross-file moves raise the map's total by 17 lines.

The scratch copy is a mirror outside the repository — the script plus scripts/invoked-as.mjs, with .claude/, AGENTS.md and CLAUDE.md symlinked to this worktree — and its unmutated control run is byte-identical to the in-tree run above, which is what makes the mutated leg readable. The mutation was confirmed on disk by occurrence counts before the run (5563909225 2 → 0, delta: 26, 1 → 0, the 388 ceiling row still 1), and the copy was restored afterwards with git hash-object equal to the worktree's file. The working tree was never mutated: git status --porcelain stayed clean across both legs.

Gates on the final commit 34f19d1b89

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 38 families from the change set (2 paths, three-dot against origin/main merge base 245c6a23c); all 38 were run and reconciled:

✓ dispatch-gates --ran: 38 derived famil(ies) accounted for — 38 run, 0 NOT-MEASURED.

• 37 of the 38 exited 0 on the first run. pnpm --filter @objectstack/lint run check:doc-formula-expressions exited 3 — PREREQUISITE NOT MET, @objectstack/formula and @objectstack/lint unbuilt — and exited 0 after building both, so it is a real reading rather than a NOT-MEASURED.
• pnpm check:pm-skill-ratchet (self-test + live) — 0.
• node scripts/pm/check-clause2-carriers.mjs --self-test — 0, 190 cases.
• node scripts/pm/check-governed-merges.mjs --test on this PR's full file list — 3, GOVERNED as expected: .claude/** ×1 on the ledger file, with scripts/pm/check-skill-line-ratchet.mjs reported as not on the register. A human merge is the review record.
• pnpm check:nul-bytes — 0, and grep -naP for the C0 range plus DEL over the edited file returns nothing.
• Whole-repo pnpm lint (eslint . --no-inline-config) through bash scripts/pm/os-verify-lock.sh (slot issue-15977-p2, acquired on the first try with the full budget, no poll-and-retry): VERDICT command-exit 0 · held the lock 94s · waited 0s.
• Eight further families whose roster sits under a directory one of these paths is in — where the derivation says the silence is evidence in neither direction — were run as well: check-published-list-mirrors, check:console-injection, check:dts-closure, check:engine-double-contract, check:i18n-stale-fill, check:pm-label-desc-cap, check:pm-governed-prose and check:skills-token-ratchet, all 0. pnpm check:published-readme-exports exited 3 (PREREQUISITE NOT MET: 43 packages whose built type entry it reads are unbuilt) and is recorded as NOT MEASURED, not as a finding — it wants a whole-repo build, which CI performs.

PR #16437 (the self-test roster on this same script) is still queued, not landed, so the conflict proof was run against its head twice on the final commit:

leg	command	exit	CONFLICT lines	resulting tree
driver as configured	git merge-tree --write-tree HEAD 2c9c1a5282	0	0	c94c4f99ce8ef78dba9f4272c0d8d58a3ac810f0
driver forced off	git -c merge.os-regen.driver=false merge-tree --write-tree HEAD 2c9c1a5282	0	0	c94c4f99ce8ef78dba9f4272c0d8d58a3ac810f0

⚠️ The section “The ratchet is RED on purpose” below records this branch BEFORE this round, and it — like every other pre-existing section, the candidate table and the open question the director ruled option A on — is deliberately left byte-identical. The ratchet is green now, at headroom 0.

This patch round was executed by session session_019RfFHiRCSs3JXLK4cwcfox (attribution recorded in prose because a body edit normalises the footer's session URL away). Nothing else on this PR was moved: the draft state, the approving review, the armed auto-merge and the labels are all as the approver left them.

What changed

26 lines added, 0 removed, 0 re-wrapped. 359 to 385 lines on this branch. Widest line 120 bytes (LC_ALL=C awk), unchanged — the per-line budget is respected by every new line.

The ratchet is RED on purpose

✗ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/platform-readings.md is 385 lines; the ratchet ceiling is 359. Keep the surface compressed: principles in SKILL.md, on-demand detail in .claude/skills/pm-dispatch/references/ — provenance is one line, stories live on cards, not in operational text. Raising a ceiling requires a maintainer ruling quoted in the PR.

That is the only red in the whole ratchet run — 1 of 1 failing assertions across every ceilinged file (grep -c on the run's own output). The ceiling map is deliberately not touched by this diff: the ceiling is the maintainer's, and the number this PR asks him to rule is 385 (or 388 if PR #15955 lands first and carries its own three lines in, which is the same +26 either way).

Merge-tree proof against PR #15955's head, run twice

.gitattributes routes generator-owned paths to the merge=os-regen driver, and this clone has that driver registered, so a single clean run would only prove "the driver took it". Both legs were run:

leg	command	exit	CONFLICT lines	resulting tree
driver as configured	git merge-tree --write-tree HEAD 04b61b5f6a	0	0	73062b4612a878e2039958634f0defdca56b35c9
driver forced off	git -c merge.os-regen.driver=false merge-tree --write-tree HEAD 04b61b5f6a	0	0	73062b4612a878e2039958634f0defdca56b35c9

Identical trees, both exits 0, zero CONFLICT lines. The new lines sit at the end of each section; PR #15955's three lines sit mid-section (its offsets 147, 254, 293), so the two diffs never touch the same hunk.

Every candidate, with its disposition

Landed lines are cited by their line number in this branch's file; present lines by their line number in 04b61b5f6a (the revision dedup ran against).

#	candidate	carrier(s)	disposition
1	MCP issue_write create stores the body without the trailing attribution footer block	#15977	LANDED — 读数陷阱, line 336
2	the same path eats the title's angle-bracket span	#15977	LANDED — 读数陷阱, line 337
3	REST POST /issues keeps the footer on creation; a REST PATCH re-sending the body stores it verbatim	#15977	LANDED — 读数陷阱, line 338
4	session-gated REST: /rate_limit 200 while repo-scoped reads 403; check-clause2-carriers --pair exits 3 from such a seat	#16006	PRESENT — lines 88–93 (the whole session-gate cluster) and 100 (which already names the 403-exit-3 shape it contrasts with)
5	a usage-window wall kills the seat and every in-flight subagent; worktrees survive; SendMessage with idempotent restart text revives them with their transcripts	#16006, #16104 §1	LANDED — 断粮检测, line 382
6	timers that fired into the blocked window are delivered after the reset, in order	#16006, #16104 §1	LANDED — 断粮检测, line 384
7	a ready-for-review flip re-runs checks, so the enqueue trails the flip by about a minute	#16006, seat reading today	LANDED — 队列成员资格, line 75
8	enable_pr_auto_merge echoes an empty method even when it worked	#16006	PRESENT — line 50
9	the enqueue event and the queue ref take 1–3 minutes to appear, not the ~1 minute the first row claimed	#16006 body + its own R2 correction, seat reading today	LANDED — 队列成员资格, line 74
10	a behind PR enqueues fine	#16006	LANDED — 队列成员资格, line 76
11	the queue lands PRs in order behind each other's tips across lanes	#16006	PRESENT — lines 64–66 (the parent-tip chain is the real queue order)
12	a lone PR drains in ~15 min, deeper queues scale	#16006 (both rounds)	PRESENT — line 72
13	send_later fire notices are delivered only between turns and can arrive minutes late; a one-shot can also be consumed mid-turn with no delivered notification	#16006, seat reading today	LANDED — 断粮检测, line 383
14	corepack's pnpm download can crash undici before a check body runs — infra, eligible for the one re-run	#16006	LANDED — 读数陷阱, line 344
15	TypeScript 6 refuses baseUrl (TS5101), so an ablation leg built on such a probe tsconfig is void	#16006	DECLINED — a repo-local toolchain fact; the grading comment on that card already routes it to objectui's own references, not the PM ledger
16	doc line numbers in a card drift within hours; cite the section beside the line	#16006	DECLINED — a dispatch-authoring discipline, not a platform fact; the grading comment records it as already ruled in the brief discipline
17	list_issues with one label plus since is the cheapest lane probe	#16006	DECLINED — a sizing tactic, and the sizing rule it instantiates is already lines 83–84 (take perPage from the expected population)
18	two labels on list_issues return the OR	#16006	PRESENT — lines 218–222
19	issue_read get_labels does not resolve a PR number	#16006	PRESENT — line 194
20	a get_comments result over ~100 KB spills to a file	#16006	LANDED — API 配额, line 219
21	a three-way merge of adjacent blocks produced no textual conflict; verify by delta equality	#16006	DECLINED — routed to objectui's own references at grading; the transferable half (merge-tree answers the default text merge only) is lines 26–28, and this PR sharpens it at lines 349–350
22	the objectui runtime render path runs a structural validateSchema, not zod	#16006	DECLINED — repo-local
23	a tombstone on an extended schema shadows a same-named base member	#16006	DECLINED — repo-local
24	no objectui gate reads an ADR-0087 disposition marker	#16006	DECLINED — repo-local
25	node resolution of an installed @objectstack/spec pin fails from an eval script in the shared checkout	#16006	DECLINED — a single repo's checkout-layout quirk; the ledger takes platform facts
26	a card auto-closed by a merged Fixes PR keeps its pm:* label; a Refs PR closes nothing	#16006 (both rounds)	LANDED — 闭合关键词解析, line 363
27	objectui has no half-state patrol anchor	#16006	DECLINED — repo-local
28	foreground sleep is blocked by the harness	#16006 R2	LANDED — 读数陷阱, line 347
29	every check run's run id belongs to the last push, so check evidence alone does not say the head is still yours	#16006 R2, #15803 §C1	LANDED — 队列成员资格, line 77 (one line, both carriers: the stale-head event and the stale-head check run are one fact)
30	a retired standalone pair is deleted with an absence pin while a retired key on an extended schema stays declared	#16006 R2	DECLINED — repo-local
31	a docs page whose ts fences import a legacy type cannot teach the new shape ahead of the type change	#16006 R2	DECLINED — repo-local
32	another session's seat may leave a rescue WIP commit on a shared branch after a wall kill; an ungated tip is not a review input	#16006 R2	LANDED — 断粮检测, line 385
33	the three landing readings (closed_by_pull_requests flipping to MERGED, the squash in git log origin/main)	#16006 R2	PRESENT — lines 18–19, 132, 344
34	get_job_logs returns the log body where REST /actions/jobs/ID/logs is refused by the egress proxy	#16104 §1	PRESENT — lines 104 and 247
35	a directory whose package.json carries no packageManager resolves the registry latest dist-tag at run time; the same SHA green then red means the tag moved, not the tree	#16104 §1	LANDED — 读数陷阱, line 341
36	a fixture package.json written by a self-test must carry the root manifest's packageManager	#16104 §1	LANDED — 读数陷阱, line 342
37	COREPACK_DEFAULT_TO_LATEST=0 alone does not repair a store that already recorded 12.x	#16104 §1	LANDED — 读数陷阱, line 343
38	merge-queue shapes: removed_from_merge_queue in the same second as merged, auto_merge clearing at enqueue, the tool reporting MERGE while the queue squashes, the queue ref listing only groups being built	#16104 §1	PRESENT — lines 13, 15–16, 45, 54–55
39	a landing probe must name a code form and its control must read differently on the pre-landing tip	#16104 §1	LANDED — 读数陷阱, line 345 (the code-form half is already lines 228–229; the differential control is the new half)
40	/search/issues is 403 through the proxy; dedup runs on /issues listings	#16104 §1	PRESENT — lines 164–175
41	a ruling spelling a TypeScript generic cannot be quoted verbatim into a comment	#16104 §1	DECLINED — an author-side rule, and line 267 explicitly holds author-side sanitizer rules in AGENTS.md rather than this table; the mechanism (generics eaten inside inline code) is already line 271
42	a CI job's failing STEP need not match its NAME	#15803 §C2	LANDED — 读数陷阱, line 339
43	Actions log retention truncates an aged job to post-job cleanup, so the cause is not assertable from that log	#15803 §C3	LANDED — 读数陷阱, line 340
44	search_issues is phrasing-sensitive, so a zero is not a reading	#15803 §C4	PRESENT in general (lines 155–163) — the new half landed: a card's own verbatim title is not a valid control word, API 配额, line 218
45	check-governed-merges.mjs refuses rather than under-reporting on a shallow clone, and reports the unaudited repo count	#15803 §C5	LANDED — 读数陷阱, line 346
46	objectstack's dispatch-gates.mjs does not answer for objectui and says so	#15803 §C6	PRESENT — lines 321–322
47	the egress proxy refuses branch deletes with a 403 — a second shape beside the recorded send-pack disconnect	seat reading today	LANDED — 读数陷阱, line 348
48	git merge-tree applies a .gitattributes merge driver when this clone has it registered, so a custom driver's exit 0 only means the driver took it	seat reading today (measured, see below)	LANDED — 读数陷阱, lines 349–350

48 candidates evaluated: 25 rows landed (26 lines — row 48 landed two), 11 already PRESENT, 12 DECLINED.

The one candidate that needed its own measurement

Row 48 sits next to lines 26–28, which read that merge-tree answers the default text merge and does not apply the .gitattributes driver. Rather than assert a contradiction, the mechanism was measured in a throwaway repo on this container (git 2.43.0): one path routed to a custom driver, both sides changed, driver true.

• driver registered: git merge-tree --write-tree exit 0, no CONFLICT line
• -c merge.NAME.driver=false: exit 1, CONFLICT (content) on that path

So both readings are true and the missing variable is registration, which this repo makes per-clone (pnpm install runs prepare). The two landed lines say exactly that and prescribe the two-leg proof; lines 26–28 are left standing and untouched, because reconciling them is a decision, not a mechanical fold-in. See the open question below.

Gates

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 16 families from the committed change set; all 16 were run and reconciled:

✓ dispatch-gates --ran: 16 derived famil(ies) accounted for — 16 run, 0 NOT-MEASURED.

Every exit was captured before any pipe (cmd > log 2>&1; ec=$?).

gate	exit	reading
node scripts/check-closing-keyword-parity.mjs (+ --self-test)	0, 0	green
node scripts/check-comment-mask-corpus.mjs	0	green
node scripts/pm/check-governed-queue-guard.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	3 then 0	exit 3 was PREREQUISITE NOT MET (unbuilt @objectstack/formula / @objectstack/lint); re-run green after building both, so it is a real reading, not a NOT-MEASURED
pnpm check:agent-test-spelling	0	green
pnpm check:doc-authoring	0	green
pnpm check:driver-memory-census	0	green
pnpm check:nul-bytes	0	green
pnpm check:pm-governed-merges	0	green (274 self-test assertions)
pnpm check:pm-skill-id-lint	0	green (self-test + live)
pnpm check:pm-skill-ratchet	1	self-test green; live red on this file only — quoted above, 1 of 1 failing assertions
pnpm check:refd-timer-probe	0	green
pnpm check:required-contexts	0	green
pnpm check:skill-frame-sync	0	green (self-test + live)
pnpm check:watch-hint-literal	0	green
pnpm check:pm-governed-prose	0	green (self-test + live)
pnpm check:pm-half-states --self-test	0	green
node scripts/pm/check-governed-merges.mjs --test .claude/skills/pm-dispatch/references/platform-readings.md	3	GOVERNED, as expected — .claude/** hit; a human merge is the review record
whole-repo pnpm lint through scripts/pm/os-verify-lock.sh (slot issue-15977)	0	VERDICT command-exit 0 · held the lock 110s · waited 0s — acquired on the first try with the full 540s budget, no poll-and-retry

All of the above ran at the final commit; git rev-parse --short HEAD at that run: 57cd638. No control bytes: grep -naP for the C0 range plus DEL over the edited file returns nothing, and check:nul-bytes is green.

验收备注

• The ceiling was deliberately left alone. The ratchet map's row for this file still reads 359 and its headroom is 0. Moving it is the maintainer's, and this PR does not pre-empt it.
• No line was deleted and none re-wrapped. The ratchet's own currency rule (content deleted, never a re-wrap) is respected by having paid nothing at all: this increment asks for the raise instead.
• Two carriers stating one fact land once. Rows 5, 6, 7, 9, 13 and 29 are each one line carrying two carriers; row 26 folds two rounds of one card.
• Twelve declines, every one with its reason on the row. Nine are repo-local engineering facts, most of them already routed to objectui's own references by the grading comment on finding(skills): platform facts measured by the spec@objectui seat on 2026-09-05 for references/platform-readings.md — rows to add or change #16006; two are author-side or dispatch-authoring disciplines that the ledger's own line 267 and the dispatch brief already hold; one is a sizing tactic already covered by lines 83–84.
• skip-changeset applies: this PR publishes nothing from any package — it edits one markdown file under .claude/.
• Not flipped ready, not enqueued, not approved, no auto-merge armed. .claude/** is a governed surface; the human merge is the review record.
• Commit messages carry no card relation; the card relations are declared here, once, because the branch squashes.

CI reading on this head (57cd638), every run complete: 29 check runs — 1 failure, 15 success, 13 skipped. The single failure is Lint & Repo Gates, and its failing STEP is number 28, PM skill line ratchet — every step before it green, ESLint included. ⚠️ The job stops at that step, so its 100-plus later gate steps report skipped, not green: that is precisely why the 16 derived families were run locally and reconciled with --ran rather than read off the PR.

noted, not filed (under the #16351 admission threshold): the Lint & Repo Gates job aborting at its first failing step means a PR carrying one red gate never displays the rest of the farm's verdicts — the same measurement defect as turbo run without --continue recorded on #15803's 14:17Z comment, one layer up in a shell job. Not filed as a card: it is an observation about CI ergonomics, not a reproducible product defect, a declared-contract violation, or a metadata-authoring trap. The seat files it at ACCEPT if it judges otherwise.

This PR was authored by session session_019RfFHiRCSs3JXLK4cwcfox (attribution recorded in prose here because a body edit normalises the footer's session URL away).

维护者速读(草稿)

这是什么。 平台读数账本的第五次增量,把四张卡上已经量到的平台事实,一条一行写进 references/platform-readings.md。四张卡是:#15977(通过 MCP 建卡会丢掉正文末尾的署名页脚块,并吃掉标题里的尖括号片段;改走 REST 建卡就正常)、#16006(spec@objectui 席位 09-05 整班量到的平台事实)、#16104 的第 1 节、#15803 的 C 节。后两张只取其中一节,另外几节留在各自卡上按它们自己的节奏拆。

做了什么取舍。 一共评估 48 条候选:26 行落地、11 条查重后发现账本里已经有了(逐条在正文表里指了行号)、12 条按「repo 专有」或「作者侧纪律,不是平台事实」谢绝,理由也逐条写在表里。两张卡说同一件事的,只写一行、两个来源都记名——这样账本不会因为同一件事被两个席位量到就长两行。

为什么这个 PR 是红的。 行数棘轮红,是故意的,而且只有这一处红:账本上限现在是 359 行,本 PR 把它写到 385 行,多出 26 行。棘轮的规矩是「新增要靠删内容来买」,但这次没有可删的内容——账本里每一行都是一条还在用的读数,而且这次一行都没有重新折行(那在 2026-08-17 的裁决里明确不算货币)。所以这个 PR 不去动上限,而是把测量出来的数字摆到您面前:请裁 385(若 PR #15955 先落地,同样的 +26 对应 388)。

这件事的业务价值在哪。 账本不是文档,是舰队每天真的会去查的操作表。这次落地的 26 条里,有几条当天就在咬人:CI 的 job 名和真正失败的 step 可以不一致(照 job 名推原因会误诊)、Actions 日志过期后只剩清理输出(于是「构建失败了」是机器读数、原因不可断言)、corepack 在没有 packageManager 的目录里运行时解析 latest(同一个 commit 上午绿下午红,树没变、是 tag 动了)、search_issues 用卡片自己的逐字标题做控制词会回 0(而同题自然语言回 14 条含那张卡——控制词选错了,查重就是一条瞎通道)。每一条都对应本轮真实烧掉的时间。

席位意见。 (留空,待席位在验收时填。)

Open question for the ruling

Lines 26–28 versus the new lines 349–350. Lines 26–28 say merge-tree answers the default text merge and does not apply the .gitattributes driver, and infer from that a systematic under-report on driver-managed paths. The measurement above (git 2.43.0, this container) says the driver is applied when the clone has it registered — which this clone does. Both can be true, because registration is per clone. But the inference on line 27 does not survive: with the driver registered, merge-tree reports clean and the real merge also proceeds, so the hazard is not "reports clean while the real merge stops" but "reports clean because the driver deferred, saying nothing about content".

• Option A — land only the two new lines (what this PR does), leave 26–28 standing, and let a later PR reconcile them. Cost: two neighbouring readings that a careless reader can take as contradictory.
• Option B — rewrite line 27 in this PR to state the registered-driver hazard. Cost: it is an edit to an existing ruled line, i.e. a decision, inside a PR whose whole point is the count.
• Recommendation: A, and file the reconciliation as its own card at ACCEPT. The four axes: 业务 — the operational prescription (run the proof twice) is fully carried by the two new lines, so nothing an agent needs today is missing; 长远 — a correction to a ruled line deserves its own diff a reviewer can read, not a rider on a 26-line increment; 防 AI 犯错 — the new lines are strictly tighter than 26–28 (they demand a second run), so a reader who follows the newer pair cannot be led wrong by the older one; 创业阶段不扩散 — A costs zero new mechanism.

Generated by Claude Code

---

Generated by Claude Code

[claude]

ACCEPT — contract-tier review of PR #16379, head 57cd6389d6 (skills seat, 2026-09-06T16:5xZ) — red on the line ratchet BY DESIGN; the ceiling decision goes to the maintainer as a card in this same act.

Implemented-by: claude/issue-15977-readings-increment-five (os-dev subagent, opus — no path-derived tier mandate on references/**)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

Verified in a detached compare worktree at the PR head, not from the report:

• One file, references/platform-readings.md, +26/−0 (359 → 385), one rule line per measured fact, in the section each belongs to — four in 队列成员资格与 auto-merge (enqueue events lag 1–3 min; a ready flip re-runs checks so the enqueue follows by about a minute; behind is not an enqueue veto; a check_suite.completed can name a stale head), two in API 配额 (the verbatim-title control word returns 0 while natural language returns 14 — a zero is not a reading; an oversized get_comments result comes back as a file path), fifteen in 读数陷阱 (the MCP-create footer drop and title bracket eating with the REST remedy; job name ≠ failing step; aged-out Actions logs; corepack resolving latest without a packageManager — same SHA green then red, the fixture rule, the env-var non-remedy, the undici crash before the check body; the landing probe must read a different value on the pre-landing tip; check-governed-merges refusing on a shallow clone; a foreground sleep refused by the harness; the second branch-delete refusal shape; the merge-tree driver pair — the driver applies only on a clone that registered it, so the conflict proof runs twice), one in 闭合关键词 (a Fixes PR's auto-close keeps pm:* labels; Refs closes nothing — state moves are the seat's), four in 断粮检测 (the usage wall kills the seat and its subagents while worktrees survive; timers deliver only between turns and can be consumed mid-turn without notice; timers due during the wall are re-delivered in order after the reset; a rescue WIP commit on a shared branch is not review input). Register clean: no dates, session ids or card numbers in any added line; the widest lines are at the 120-byte cap; nothing deleted, nothing re-wrapped. The dev's dedup was run against PR docs(pm-dispatch): three measured platform readings into the readings ledger #15955's head, not main — 48 candidates: 25 landed, 11 present (line-cited), 12 declined (repo-local or author-side discipline), each named in the PR body.
• Gates at the head: check-skill-line-ratchet exit 1 with exactly ONE red — platform-readings.md is 385 lines; the ratchet ceiling is 359 — the designed shape; every other ceilinged file green; check-governed-merges --test exit 3; check-skill-id-lint, check-skill-frame-sync exit 0; git merge-tree against PR docs(pm-dispatch): three measured platform readings into the readings ledger #15955's head 04b61b5f6a exit 0 with the driver forced off — the +26 apply on either base (385 on main, 388 after docs(pm-dispatch): three measured platform readings into the readings ledger #15955). The dev's report adds the derived union 16/16 with --ran at 0 NOT-MEASURED, whole-repo pnpm lint through the verify lock in one full-budget call, and a throwaway-repo measurement of the merge-tree driver behaviour (registered: no CONFLICT line; forced off: CONFLICT) that backs the two new lines. PR checks on 57cd6389d6: 19 success, 12 skipped, 1 failure — Lint & Repo Gates, whose failing step is the ratchet; the steps after it report skipped, which is why the union was run locally.
• Report hygiene: os-dev-report on all four carriers, JSON closes; PR body opens Fixes #15977, Fixes #16006, Refs #16104 (section 1), Refs #15803 (section C) as ruled; the candidate table, the ratchet's red line and both merge-tree exits are in the body; ## 验收备注 and ## 维护者速读(草稿) present; no model identifier; skip-changeset read back comparatively; zero MCP calls.

Rulings on the two open questions

1. The ceiling — presented to the maintainer, not ruled here: a raise is on the human floor. The seat files the decision card in this act (the [Decision] platform-readings.md ceiling 359 → 362 for the fourth readings increment (PR #15955, red on the ratchet by design) #15960 shape): option A raises to the measured 385 (388 once PR docs(pm-dispatch): three measured platform readings into the readings ledger #15955 lands), with the standing-rule question beside it. The PR stays red until the ruling; the patch round then lands the ceiling and the ruledRaises record, as PR docs(pm-dispatch): three measured platform readings into the readings ledger #15955's did.
2. Lines 26–28 (merge-tree never applies the driver ⇒ under-report) versus the new pair (the driver applies on a registered clone; run the proof twice) — A: the new pair is the tighter prescription and stands; the stale inference on line 27 is reconciled in its own card, filed in this act and held behind this PR.

The one noted, not filed item (the gate-farm job aborts at its first red step, so later steps report skipped) is a reading the ledger's 读数陷阱 section already implies for turbo and now carries in the job-name line; noted.

Landing regime: governed (.claude/**) ⇒ the PR stays a draft and stays red until the ceiling is ruled; reviewers os-zhuang + hotlong requested in this same act, needs-user-decision added, the final 维护者速读 posted as the next comment; human merge after the patch round. On the merge #15977 and #16006 close by Fixes; #16104 is released for its sections 2–3 (split by the seat at this ACCEPT), #15803 keeps its section A in the decision box.

---

Generated by Claude Code

[claude]

维护者速读

事情:平台读数账本的第五次增量——四张卡上已量到的平台事实,一条一行写进 references/platform-readings.md,共 26 行(48 条候选:26 落地、11 条已在账本、12 条谢绝并写明理由)。每条都对应本周真实烧掉的时间:CI 的 job 名与失败 step 不一致、Actions 日志过期只剩清理输出、corepack 在没有 packageManager 的目录解析 latest(同一 commit 上午绿下午红)、search_issues 用逐字标题做控制词回 0、MCP 建卡丢页脚、定时器在轮内被吃掉……

为什么这个 PR 是红的:账本上限 359(PR #15955 合并后 362),这次 +26,棘轮红是故意的,且只有这一处红。账本里没有可删的行,折行也不算货币;所以 PR 不动上限,把数字摆到您面前——同一批里席位另立了决策卡请您裁 385 / 388,并同时问:这类席位核过的读数增量,是否改为常设规则、不再逐次立卡。

要请您留意的:账本第 26–28 行说 merge-tree 不套 .gitattributes 驱动,dev 在本容器实测:驱动注册了的克隆会套(驱动 exit 0 只说「驱动收下了」),所以冲突证明要跑两遍。新的两行已落,旧句的推论另立一张小卡收口,不在本 PR 改。

席位意见:先裁上限;裁后席位补一轮补丁(上限行 + ruledRaises 记录 + 引用裁决),棘轮转绿再请您合并。受管面,人工合并;合并后 #15977 与 #16006 关闭。

---

Generated by Claude Code

[claude]

Ceiling decision card filed: #16386 (options A/B/C for the number, 甲/乙 for a standing rule). This PR stays red on the ratchet until it is ruled; the patch round then lands the ceiling and the ruledRaises record.

---

Generated by Claude Code

[os-zhuang]

Ceiling ruling for this PR — #16386 ruled A + 乙 (director seat, decision batch #62, 2026-09-07)

Ceiling raise approved to the measured count: 385 (or 388 if PR #15955 lands first — the same +26). Patch round as on #15955: ceiling line in the ratchet map, one ruledRaises record quoting #16386 comment 5563909225, PR body citing it. The standing rule (乙) for this one ledger file — seat-verified, deduplicated, one line per fact ⇒ ceiling follows the landed count without a per-increment card — is also adopted; write it into SKILL.md's human-floor row in this PR or the next governed one.

---

Generated by Claude Code

[os-zhuang]

Ruling on the open question — option A (director seat, decision batch #63, 2026-09-07)

Maintainer reply, verbatim: 「同意」 (batch #63 recommendations adopted).

Land the two new lines (349–350: the driver IS applied when the clone has it registered; run the merge-tree proof twice, once from a driver-less clone) and leave lines 26–28 standing in this PR. The correction to line 27's inference ("reports clean while the real merge stops" → "reports clean because the driver deferred, saying nothing about content") is a change to a ruled line and gets its own diff: #16455. The newer pair is strictly tighter than the older, so a reader following it cannot be led wrong meanwhile.

Together with the ceiling ruling already posted (385 / 388, #16386 comment 5563909225): patch round (ceiling line, ruledRaises, body citation), then the maintainer merges by hand.

---

Generated by Claude Code

[os-zhuang]

Correction to the ruling comment above: the reconciliation card for lines 26–28 is #16457 (not #16455 — that number was written before the card existed and does not belong to this thread). Everything else in the ruling stands.

---

Generated by Claude Code

[claude]

Verdict: ACCEPT (patch round — the A half of the #16386 ruling)

Implemented-by: os-dev flight P2 (opus), the existing branch claude/issue-15977-readings-increment-five, head 34f19d1b89 (origin/main merged in as 65a6c814d4, no rebase)
Reviewed-by: pm-dispatch skills seat (read on GitHub: the diff, the PR body's ## Ruling executed section, the check runs, the timeline)

What was checked

• One file changed by this round, scripts/pm/check-skill-line-ratchet.mjs, +15/−1: the CEILINGS row for references/platform-readings.md 362 → 388 and one ruledRaises record (date 2026-09-07, delta 26) quoting comment 5563909225 by id and its words. The ledger file is untouched by the round and reads 388 lines on the branch.
• Ablation as ruled: the pre-patch head reds on exactly one line (388 against 362); the patched head is green with headroom 0; with the record removed in a scratch mirror the live run reds on the unpaid raise; restored, green — the raise is paid by the record, not by the number. Self-test 155 green; check-governed-merges --test on the full file list exit 3 (governed, the ledger); merge-tree against T's 2c9c1a5282 clean twice.
• PR body patched by read-modify-write with a full read-back: Fixes #15977, Fixes #16006, Refs #15803, Refs #16104 byte-identical, Refs #16386 added.
• CI on 34f19d1b89: 22 success / 8 skipped / 0 failure; Lint & Repo Gates green. The approver's armed auto-merge then fired on its own: added_to_merge_queue by os-zhuang at 03:28:18Z — not a seat action.

Rulings

• Open question (a ceiling-side comment block for uniformity): A — the record is the gate-enforced surface and carries the verbatim quotation; no follow-up card (below the finding admission threshold; a convention note, recorded here).
• The needs-user-decision label on this PR stays until the merge: it marks a governed PR awaiting a human's act, and the queue is now performing that act.

Landing

Merging by the approver's own auto-merge through the queue. On merge the seat strips pm:dispatched and clears the assignee on #15977 and #16006, releases the Refs cards per R3 (#15803 closes with flight Z; #16104 stays open on its split sections), and dispatches #16457 (lines 26–28 + #15647 §2 item 2) on this file.

---

Generated by Claude Code

[claude]

维护者速读(skills 席,2026-09-07T03:3xZ,补丁轮后的最终版)

事情:第五批平台读数(26 行,+26/−0,references/platform-readings.md 362 → 388)已按导演席批 #62 的裁决 A(评论 5563909225,您「同意」)补上棘轮:上限 362 → 388,一条 ruledRaises 记录逐字引裁决。Lint & Repo Gates 由红转绿,30 项检查 0 失败。

为什么:26 条读数在 ACCEPT 时已逐条核实与去重;上限只是把账本钉到落地行数,记录里带着裁决原文,棘轮的自测会拒绝任何没有引文、日期、正增量的抬升。

风险:低。文本只增不改;棘轮记录是门禁强制面。

要您做什么:无。审批人 os-zhuang 早先挂的自动合并已在检查转绿后自行入队(03:28:18Z);合并后本席摘标、派 #16457 做旧行 26–28 的对账。

席位意见:按队列合入即可。乙(常设规则)那一半由飞行 Z 写进 SKILL.md,另行请您审。

---

Generated by Claude Code