pm-ratchet: cap ceilinged lines at 120 bytes, re-wrap the legacy corpus, re-pin the map

[claude]

Fixes #11106

The line ratchet counts lines while its own docblock states it is pricing a per-session token read — and tokens track bytes. The card measured the gap: one PR grew references/platform-readings.md 16,953 → 22,559 bytes (+33%) at a flat 134/134 line count, green on every run, with no ceiling raised and nothing in the gate able to see it.

Maintainer ruling 2026-08-23 (option B), recorded verbatim and untranslated in comment 5386683155:

「10950 不考虑存量,其他接受你的建议」

Ruled: a max-line-length rule (~≤120 bytes/line) on the ceilinged files, a one-off re-wrap of the existing long lines, and the ratchet-table sync — not a parallel byte/token ratchet (A), not a disclaimer-only docblock edit (C). The rider from the analysis — measure the other ceilinged files' byte drift in the same pass — is at the bottom of this body.

Governed surface: .claude/** + AGENTS.md. Draft, human merge. Not ready-flipped, no auto-merge.

---

1. The rule — 120 bytes, and why that number

scripts/pm/check-skill-line-ratchet.mjs now checks every line of every file in its CEILINGS table against a 120-byte budget, alongside the existing per-file line ceiling.

120 is the corpus's own upper bound, not a new house style. Measured over the 4,028 lines of the pre-change corpus: 34 lines land at exactly 120 bytes, and the two wrap conventions already in use both sit under it — ASCII prose wraps at ≤91 bytes (~88 columns) and CJK prose at ≤120 bytes (50–60 characters, ~100–120 display columns, a CJK character being 3 bytes and 2 columns wide). 120 is the one number both conventions already satisfy, so the rule codifies what the careful authors were doing instead of reflowing the corpus to an invented width. Bytes rather than characters or columns, because bytes is what the header prices.

The exemptions live in the RULE, and there is no allowlist

A length rule that can demand an illegal wrap is worse than no rule — it teaches authors to break a table, split a fence, or wrap a 行锚定 directive whose whole contract is that a grep finds it at a line start. Every exemption is therefore derived from the line's own syntax. There is deliberately no exemption registry and no per-file allowlist: a list of blessed long lines is exactly the ratchet-expanding remedy #8435 keeps out of an author's reach, and it would rot.

class	what it exempts	why
fence	fenced code blocks, fence lines included	a wrapped fence is a different program
table	markdown table rows	a wrapped | row is a different table
heading	ATX headings	a heading has no continuation line
frontmatter	YAML front matter	a scalar is not markdown prose
anchored	line-anchored Blocked-by: / Restart-when: / Restart-touch: / Unlock-when:	行锚定 — the contract IS that a grep finds it at a line start
blockquote	blockquote lines	its continuation must repeat >, so wrapping inserts a non-whitespace byte into a quotation
quotation	lines inside a MULTI-line 「…」/『…』 verbatim ruling	the quote already spans lines as its author broke it; moving the break moves phrases across it
unbreakable	lines with no legal break point (bare URL, long path, one long code span)	wrapLine returns them unchanged — the gate never demands a wrap it cannot itself produce

Two of those were found by running the re-wrap and reading what broke, not designed up front, and both are declared as judgment calls the reviewer can reverse in one predicate each:

• blockquote — the re-wrap's content-identity proof went red on three files because a blockquote continuation needs a >. Wrapping would have edited verbatim maintainer rulings. Cost: 10 lines stay long.
• quotation — an earlier cut split 「我们是一个创业项目,…」 across two lines and turned check-skill-frame-sync's self-test red: a phrase inside a governed quotation stopped being findable on one line. A soft break does not change what a reader sees, but it does change what a matcher sees, and quoted rulings are the text most likely to be matched. Cost: 68 lines stay long.

A quote that fits on one line is not exempt — it is an atom instead, so surrounding prose still wraps around it intact.

What a wrap may move

Only whitespace: a space becomes a newline, or a newline is inserted between two East Asian characters (a segment break there is removed by the CSS segment-break transformation rules, which is why the corpus already wraps CJK prose mid-run). No break is offered at a CJK↔Latin junction without a space, because that one would render as a space. Inline code spans, links, autolinks, URLs and single-line 「…」 quotes are atoms — never broken into. A continuation line is never allowed to begin with a sequence markdown reads as a new block.

Self-test

--self-test goes 26 → 71 cases, with a red/green pair for the rule and for each exemption class — every exemption case carries a RED twin in a non-exempt shape, so a case can only pass by the rule actually discriminating:

✓ check-skill-line-ratchet self-test: 71 cases pass.

2. The legacy re-wrap — 555 lines, zero content changes, proved per file

The re-wrap tool imports classifyLine / wrapLine / advanceState from the gate itself, so the re-wrap and the rule it satisfies cannot disagree. Two equality proofs run per file, and the tool refuses to write a file that fails either:

1. whitespace-stripped byte identity — every non-whitespace character unchanged, in the same order, so the only thing that changed is whitespace;
2. inline code-span sequence identity — the whole-file sequence of `…` spans byte-identical, which is what protects every path, every Blocked-by: #N, and check-governed-prose's code-span scan.

3. Per-file arithmetic, ceiling sync, and the equality proof

Counts against origin/main. Every raise is the mechanical consequence of wrapping, and headroom is 0 on every row. landing-operations.md moved the other way (82 → 80, its standing headroom locked in); six files were already within budget on every line and did not move at all.

file	lines before → after	Δ	ceiling	bytes	longest line B (before → after)	equality proof	exempt long lines
.claude/skills/pm-dispatch/SKILL.md	666 → 1008	+342	pinned 1008	78376 → 78919	824 → 765	ws-identical + spans-identical	quotation×42 table×23 blockquote×4 unbreakable×1
.claude/skills/pm-dispatch/references/dispatch-runbook.md	242 → 274	+32	pinned 274	20480 → 20554	172 → 171	ws-identical + spans-identical	quotation×18
.claude/skills/pm-dispatch/references/state-machine.md	43 → 43	+0	pinned 43	3408	119 → 119	untouched	—
.claude/skills/pm-dispatch/references/contract-review.md	44 → 48	+4	pinned 48	3320 → 3331	132 → 120	ws-identical + spans-identical	—
.claude/skills/pm-dispatch/references/decision-analysis.md	38 → 48	+10	pinned 48	3072 → 3087	140 → 135	ws-identical + spans-identical	quotation×5
.claude/skills/pm-dispatch/references/platform-readings.md	130 → 315	+185	pinned 315	29660 → 30153	1437 → 120	ws-identical + spans-identical	—
.claude/skills/pm-dispatch/references/rest-channel.md	46 → 87	+41	pinned 87	6718 → 6817	714 → 129	ws-identical + spans-identical	unbreakable×1
.claude/skills/pm-dispatch/references/review-checklist.md	82 → 84	+2	pinned 84	7944 → 7949	270 → 119	ws-identical + spans-identical	—
.claude/skills/pm-dispatch/references/landing-operations.md	80 → 80	+0	lowered to 80	6140	117 → 117	untouched	—
.claude/skills/pm-dispatch/references/release-aftercare.md	58 → 58	+0	pinned 58	4109	119 → 119	untouched	—
.claude/skills/pm-dispatch/references/seat-post-protocol.md	101 → 105	+4	pinned 105	9074 → 9086	181 → 119	ws-identical + spans-identical	—
.claude/skills/pm-dispatch/references/lanes/engine.md	40 → 40	+0	pinned 40	2612	118 → 118	untouched	—
.claude/skills/pm-dispatch/references/lanes/services.md	30 → 30	+0	pinned 30	1762	115 → 115	untouched	—
.claude/skills/pm-dispatch/references/lanes/cli.md	35 → 35	+0	pinned 35	2117	120 → 120	untouched	—
.claude/skills/pm-dispatch/references/lanes/devx.md	37 → 39	+2	pinned 39	2633 → 2639	162 → 119	ws-identical + spans-identical	—
.claude/skills/pm-dispatch/references/lanes/skills.md	35 → 35	+0	pinned 35	1998	116 → 116	untouched	—
.claude/skills/pm-dispatch/references/lanes/spec.md	43 → 46	+3	pinned 46	2663 → 2671	184 → 120	ws-identical + spans-identical	—
.claude/skills/pm-dispatch/references/lanes/hotcrm.md	44 → 57	+13	pinned 57	3800 → 3831	224 → 224	ws-identical + spans-identical	blockquote×2 anchored×1 quotation×2
.claude/agents/os-dev.md	405 → 470	+65	pinned 470	35845 → 36017	182 → 170	ws-identical + spans-identical	quotation×1 unbreakable×1
.claude/skills/checklist-test/SKILL.md	232 → 238	+6	pinned 238	14965 → 14973	221 → 221	ws-identical + spans-identical	table×6
.claude/skills/checklist-author/SKILL.md	61 → 62	+1	pinned 62	3856 → 3860	121 → 119	ws-identical + spans-identical	—
.claude/skills/dogfood-verification/SKILL.md	155 → 157	+2	pinned 157	10951 → 10965	126 → 125	ws-identical + spans-identical	unbreakable×2
.claude/skills/spec-property-retirement/SKILL.md	334 → 337	+3	pinned 337	24452 → 24460	328 → 328	ws-identical + spans-identical	table×6 blockquote×1
AGENTS.md	961 → 1149	+188	pinned 1149	82233 → 82864	2092 → 1081	ws-identical + spans-identical	blockquote×3 table×11 unbreakable×1
CLAUDE.md	86 → 86	+0	pinned 86	5264	91 → 91	untouched	—

Totals: 25 files · 555 lines re-wrapped · line delta +903 · bytes 367,452 → 369,586 (+2,134, +0.58% — the inserted newlines and list-continuation indents, which is precisely what "whitespace-only" costs) · all equality proofs pass.

How to read the jumps. The pre-change numbers are not comparable to these. A line that used to hold 400–830 bytes is now three to seven lines holding the same bytes; 666 → 1,008 is not slack, it is the same text measured in a unit that finally means something. Headroom is 0 on every row, and each line is now capped at 120 bytes — which is what makes the count track the token read the header prices.

133 lines remain over budget, all structurally exempt: quotation×68, table×46, blockquote×10, unbreakable×6, fence×2, anchored×1. Table rows are the largest remaining unmetered channel (the five longest surviving lines are all table rows, topping out at 1,081 bytes in AGENTS.md) and are filed as a separate finding — see below.

4. Rider from the ruling: byte drift across the whole ceilinged corpus

"Measure the other ceilinged files' drift in the same pass (it falls out for free)." Bytes at the commit where each ceiling was last set (derived by replaying the gate script's own history and reading each CEILINGS value at each commit) against bytes on origin/main:

file	ceiling set	ceiling	lines then → on main	bytes then → on main	byte drift the ratchet could not see
…pm-dispatch/references/lanes/hotcrm.md	2026-08-24	44	44 → 44	2982 → 3800	+818 (+27.4%)
…pm-dispatch/references/state-machine.md	2026-08-23	43	43 → 43	3134 → 3408	+274 (+8.7%)
…pm-dispatch/references/dispatch-runbook.md	2026-08-23	242	242 → 242	19442 → 20480	+1038 (+5.3%)
…pm-dispatch/references/review-checklist.md	2026-08-21	82	82 → 82	7670 → 7944	+274 (+3.6%)
AGENTS.md	2026-08-21	961	961 → 961	80901 → 82233	+1332 (+1.6%)
.claude/agents/os-dev.md	2026-08-23	405	405 → 405	35375 → 35845	+470 (+1.3%)
…pm-dispatch/SKILL.md	2026-08-23	666	666 → 666	77695 → 78376	+681 (+0.9%)
…pm-dispatch/references/seat-post-protocol.md	2026-08-21	101	101 → 101	8991 → 9074	+83 (+0.9%)
…pm-dispatch/references/landing-operations.md	2026-08-21	82	82 → 80	6253 → 6140	−113 (−1.8%)
the other 16 files	2026-08-21 → 08-24	—	flat	flat	+0 (0.0%)

The answer to the card's open question is yes. Eight files grew in bytes at a perfectly flat line count, one of them by 27.4%, every one of them green on every run — the same silent mechanism the card reported for platform-readings.md, confirmed across the corpus. And the windows are days, not weeks: every ceiling in the map was set between 2026-08-21 and 2026-08-24, which makes +27.4% in that span sharper evidence, not softer. platform-readings.md reads +0% only because its ceiling was re-pinned on 08-24, after the incident this card reports.

Verification

Reverse verification — plant, prove on disk, read the gate's own verdict, restore. The on-disk probe is anchored on the text being changed in both directions, and it earned its keep: the first attempt used a hand-written anchor that was not in the file, the probe reported joined=0 and the run was discarded rather than re-tried until something landed.

leg	on-disk probe	gate	verdict line
0 baseline	joined=0 continuation-at-line-start=1 max=115B	exit 0	✓ … services.md is 30 lines (ceiling 30; headroom 0).
1 mutate (join a wrapped pair into one 209B prose line)	joined=1 continuation-at-line-start=0 max=209B	exit 1	✗ … services.md has 1 line(s) over the 120-byte budget: L12 (209B).
2 control (the SAME line in an exempt shape — blockquote)	blockquoted=1 bare-joined=0	exit 0	✓ … services.md is 29 lines (ceiling 30; headroom 1).
3 restore	joined=0 continuation-at-line-start=1 max=115B	exit 0	✓ … services.md is 30 lines (ceiling 30; headroom 0).

Leg 1 is worth reading twice: on the mutated tree the line ratchet went green with headroom 1 while the new rule went red. Fewer lines, more bytes, old control blind — the drift this card is about, reproduced on demand.

Derived gate union — node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no path args; 160 families discovered, 19 matched). All 19 run green, plus check:ratchet-remedy-authority run explicitly. Exit codes captured before any pipe; each line is the gate's own verdict.

gate	exit	verdict
check:pm-skill-ratchet	0	✓ check-skill-line-ratchet self-test: 71 cases pass. + 25 file rows, all headroom 0
check:skill-frame-sync	0	✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
check:pm-governed-prose	0	✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces
check:pm-skill-id-lint	0	✓ check-skill-id-lint: 22 file(s) clean (pattern /#[0-9]{3,}/g).
check:pm-governed-merges	0	✓ check-governed-merges --self-test: 129 assertions
check:nul-bytes	0	check-nul-bytes: OK (scanned 6650 text file(s) … no raw ASCII control bytes).
check:agent-model-declared	0	✓ check-agent-model-declared: 1 agent definition(s) … all declare a model
check:agent-test-spelling	0	✓ check-agent-test-spelling: 0 violations — 364 file(s)
check:doc-authoring	0	✓ doc authoring guard: 389 files clean
check:docs-audit-scope	0	✓ release-owned pages are in scope and read-only
check:entry-guard	0	✓ check:entry-guard: 154 scripts/ file(s)
check:parse-guard	0	✓ check:parse-guard: 153 scripts/ file(s)
check:pnpm-filter-targets	0	✓ check:pnpm-filter-targets: 135/168 --filter occurrence(s)
check:required-contexts (×2 spellings)	0	✓ check-required-contexts: 6 required context name(s) pinned
check:cross-package-test-inputs (×2)	0	OK: 16 package(s) read outside themselves, all declared
check-ci-filter-parity	0	OK: all 96 declared cross-package glob(s) … covered
check:doc-formula-expressions	0	✓ … 9 @example(s) judged clean across 1001 packages/spec/src files
check:ratchet-remedy-authority	0	OK … 138 scripts swept (scripts/*.{mjs,mts}); 8 mark the expanding remedy ⛔ MAINTAINER-ONLY, 5 turn it down outright, 125 hand out no ratchet-expanding remedy.

Green union run at 117fc0bd, the final commit.

On check:ratchet-remedy-authority. It did not appear in the derivation (consistent with #11935) and it does not reach this script either way: its corpus is scripts/*.{mjs,mts}, non-recursive, and this gate lives in scripts/pm/. Run explicitly regardless, and it neither flags this file nor reports it UNCLASSIFIED — 138 scripts swept, none of them this one. Independently of that: the new failure text offers no registry-expanding remedy to mark. Its remedy is "wrap the line"; the exemptions are structural predicates, not a list, and the message says so in terms (there is no allowlist to add a line to), which is a self-test case. The pre-existing ceiling-raise remedy in the ratchet's other message is untouched by this PR.

ESLint — the repo-wide sweep is CI's run; the local run is a declared narrowing, with all three readings:

1. Receptacle read from ESLint's own config, not guessed: running ESLint on the changed .md files returns File ignored because no matching configuration was supplied — markdown is outside the population. The only file in this diff ESLint judges is scripts/pm/check-skill-line-ratchet.mjs.
2. Count read from --format json: 1 file linted, 0 errors, 0 warnings.
3. Invariance for untouched files: eslint.config.mjs never enables type-aware linting (its own docblock, verbatim: "this repo runs one eslint.config.mjs, which never enables type-aware linting (no parserOptions.project, no typed @typescript-eslint rules) for ANY file, test or not"), so nothing in this diff can move an untouched file's verdict.

Not touched: the published skills/ catalog. This diff is .claude/** + AGENTS.md + scripts/pm/** only, so the published-catalog line-budget clause does not apply and no two-reading report is owed. The gate's header boundary — the published root is deliberately outside the ceiling, and extending it is a policy change needing its own ruling — is unchanged and still pinned by its self-test case.

Changeset: none. .claude/** + scripts/pm/** publish nothing; skip-changeset applied.

Out-of-scope finding filed

• [finding] 133 lines stay over the 120-byte budget by structural exemption — markdown table rows are the largest remaining unmetered channel (up to 1,081 bytes on one line) #11947 — 133 lines stay over the 120-byte budget by structural exemption; markdown table rows are the largest remaining unmetered channel (up to 1,081 bytes on one line). Filed unassigned, finding + domain:skills, not fixed here.

---

Generated by Claude Code