[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)

[os-zhuang]

Filed unassigned, recording only. Measured while landing the max-line-length rule ruled on #11106 (option B). The PR there declares these same numbers in its own body rather than hiding them.

What the new rule does and does not reach

The rule caps every line in the CEILINGS corpus at 120 bytes, and the one-off re-wrap brought 555 legacy lines under it. 133 lines remain over budget, every one of them exempt by shape — the exemptions are structural and correct, but they are not free, and nothing measures what flows through them.

exempt class	lines still over 120B	why it is exempt
quotation (multi-line 「…」 ruling)	68	line structure is the author's; re-flowing moved a phrase and turned check-skill-frame-sync red
table (markdown table row)	46	a wrapped | row is a different table
blockquote	10	a continuation must repeat >, i.e. insert a non-whitespace byte into a quotation
unbreakable (URL / path / one long code span)	6	no legal break point exists
anchored (Blocked-by:-family)	1	行锚定 — the contract is that a grep finds it at a line start
fence	2	a wrapped fence is a different program

The measured shape of the hole

Table rows are the sharpest of these, on two counts:

• They are the longest. The five longest surviving lines in the corpus are all table rows: AGENTS.md:702 at 1,081 bytes, SKILL.md:276 at 765, SKILL.md:279 at 631, AGENTS.md:701 at 608, SKILL.md:280 at 484.
• They are unbounded by construction. A table row grows by widening a cell, which costs zero lines and passes both controls the corpus has — the line ratchet counts it as one line, and the new length rule exempts it by shape. That is the same mechanism [finding] The pm-skill ratchet counts LINES while its header prices a per-session token read — a green 134/134 run grew platform-readings.md ~33% in bytes #11106 reported for long prose lines, surviving in the one place the fix could not reach.

quotation at 68 lines is the larger count but a much weaker concern: those are verbatim maintainer rulings, a bounded and human-audited form, and their line structure is deliberately preserved.

Not claimed

• ⛔ Not a claim that the exemptions are wrong. Each of the six is load-bearing and the alternative — a rule that demands an illegal wrap, or one with a per-line allowlist — is worse. The finding is that the exempt set has an unmeasured interior, not that it should be smaller.
• ⛔ Not measured: whether table cells are actually drifting. The byte-drift rider on [finding] The pm-skill ratchet counts LINES while its header prices a per-session token read — a green 134/134 run grew platform-readings.md ~33% in bytes #11106 measured whole files, not per-row. Whether the 46 rows have been widening the way the prose lines were is the cheap next reading, and it is the reading that would tell a grader whether this is worth spending anything on.

Possible directions (triage, not a recommendation)

• A per-row byte budget on table rows, wrapping cells being illegal, so the remedy is to move the row's detail into prose or a reference file. Reaches the largest lines; needs a number the corpus does not obviously supply.
• A whole-file byte ceiling alongside the line ceiling for the handful of table-heavy files (AGENTS.md, pm-dispatch/SKILL.md). This is option A from [finding] The pm-skill ratchet counts LINES while its header prices a per-session token read — a green 134/134 run grew platform-readings.md ~33% in bytes #11106, which the maintainer declined for the whole corpus — narrowed to the files where the line rule provably cannot reach, it is a different and much smaller proposal.
• Nothing. 133 lines out of 4,931 in a corpus that now has a working control, and the exempt classes are the ones a reviewer's eye is best at judging anyway.

Which of these is worth it is a maintainer call, not this seat's.

---

Generated by Claude Code

[os-zhuang]

First-touch grading (skills lane self-triage, session session_01RMTpSRF5CjMmQBFfPtPCwJ) — the card named its own deciding reading ("whether the 46 rows have been widening the way the prose lines were"), so I took that reading before grading. Verdict: finding → needs-user-decision, with the direction menu now narrowed by measurement.

The measurement

Per-row byte drift on markdown table rows, keyed by first cell, between two refs of origin/main:
a749273b (2026-07-24) → ef786a51 (today), plus b3fe0f99 (2026-08-09) → today for files younger than July.

AGENTS.md (Jul 24 → today): 45 matched rows, zero widened. Over-120B rows went 3 → 11 entirely via new rows born long, not via widening.

AGENTS.md (Aug 9 → today): one row widened — and it is the corpus's #1 longest line, the **/translations/*.generated.ts row at AGENTS.md:702: 639 → 1,081 bytes (+442B in ~2 weeks). The other 58 matched rows: unchanged.

pm-dispatch/SKILL.md (Aug 9 → today): 3 rows widened (+750B total), led by the domain:engine row at 101 → 765 bytes (7×). Meanwhile 7 rows shrank and the file's over-120B row count fell 33 → 23 through consolidation.

What it decides

• Direction "Nothing" is refuted. The widening mechanism is not hypothetical: the two single longest lines in the corpus got there by exactly the unmetered path this card describes, both within the last two weeks.
• But this is a heavy-tail problem, not broad decay. ~100 matched rows show zero drift; the harm concentrates in 2–3 rows per file that absorb hundreds of bytes each. A corpus-wide new rule would be sized for a defect that lives in a handful of rows.
• Direction A's missing number has a repo-idiomatic answer. The card said a per-row budget "needs a number the corpus does not obviously supply." It doesn't have to: pin the current max-table-row-bytes per file as a shrink-only ratchet (same idiom as the line-count ceilings and MAX_LINE_BYTES). No invented constant; today's worst row becomes the ceiling, widening any row past its file's pin fails, and consolidation pays it down exactly like the existing ratchets.

Decision requested (maintainer)

1. A-ratchet (recommended): per-file max-table-row-bytes pin in check-skill-line-ratchet.mjs, shrink-only, no new constants — catches the measured defect precisely, at the cost of one more registry.
2. B-narrowed: whole-file byte ceiling on the 2 table-heavy files only (declined corpus-wide in [finding] The pm-skill ratchet counts LINES while its header prices a per-session token read — a green 134/134 run grew platform-readings.md ~33% in bytes #11106; this is the narrowed variant).
3. Nothing — now against the evidence above.

Pick one and I'll charter it as a skills-lane card; the measurement script and refs are reproducible from this comment.

---

Generated by Claude Code

[os-support-ai]

Maintainer ruling — option 1: a per-file max-table-row-bytes shrink-only ratchet

Source: maintainer, 2026-08-25, live PM chat (decision-inbox batch 8 review, session session_01KWRU3s15AJz7PGW7a7wdCh), verbatim: 「同意」 — accepting the batch's presented recommendation for this card: the skills seat's option 1 (A-ratchet).

• check-skill-line-ratchet.mjs gains a per-file max-table-row-bytes pin over the CEILINGS corpus: shrink-only, seeded at today's widest table row per file — no invented constants, the same idiom as the line-count ceilings and MAX_LINE_BYTES. Widening any row past its file's pin fails; consolidation pays it down.
• Grounds as measured and accepted: the widening channel is real and heavy-tail — the corpus's two longest lines both got there through unmetered table cells within the last two weeks (AGENTS.md:702 639 → 1,081 bytes; the domain:engine row 101 → 765) — while ~100 matched rows show zero drift, so a corpus-wide new rule would be oversized; the ratchet catches exactly the measured defect. Options 2 (whole-file byte ceiling on two files) and 3 (nothing — refuted by the measurement) are not taken.
• The other five structural exemption classes (quotation/blockquote/unbreakable/anchored/fence) are untouched — they remain load-bearing and human-audited.

This card is the charter — the skills seat implements directly on it (no separate charter card; the measurement script and refs in the grading comment are the seed data). The implementation PR touches scripts/pm/check-skill-line-ratchet.mjs (not governed surface; the self-test rides per the script's own rules).

State: needs-user-decision → pm:queue in the same write; domain:skills stands.

---

Generated by Claude Code

[os-zhuang]

Live instance for this decision, recorded while it waits: PR #12086 (SKILL.md discipline pack, ACCEPTed with a caveat) landed six graded rules in the zero-headroom SKILL.md by lowering the line count while raising the byte count (~+2.1KB, 1005→999 lines / 79,260→81,320 bytes) — line-budget re-wrap funding new content, honestly declared by the dev and mechanically confined to touched paragraphs. This is the unmetered channel this card measures, occurring in the corpus's most load-bearing file, under full disclosure. It strengthens option A-ratchet (per-file byte-width pins, shrink-only): with that in place, such a payment would be visible as a ratchet transaction instead of a PR-body confession.

---

Generated by Claude Code

[os-zhuang]

Claim: skills seat, session session_01RMTpSRF5CjMmQBFfPtPCwJ. Folded dispatch — ratchet-script pack (this card + #12098 + #12081; the three share scripts/pm/check-skill-line-ratchet.mjs as the serial unit). Branch: claude/issue-11947-table-row-byte-pins. Implementing the maintainer-ruled option 1 (A-ratchet, ruling comment 5406811814): per-file max-table-row-bytes shrink-only pins seeded at today's widest row per file.

---

Generated by Claude Code