skills(pm): encode the axis-weighting and decision-batch presentation rulings

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Fixes#14030
Fixes#14032

Two maintainer rulings from the 2026-09-01 PM chat, both of which explicitly order
their own encoding into the skills. Landed together because they are two halves of one
decision-presentation discipline: one governs how to weigh, the other how to
present. Landing either alone leaves the other half still wrong.

HEAD at verification: 58337bf97.

The two rulings, verbatim and untranslated

#14030 — axis weighting

四维分析中，长期合理应该权重最高，至少50%

#14032 — decision-batch presentation

每批5张详细解释，不是这样一堆列给我，这个也要写入skills

What landed, per card

#14030 — long-term soundness carries the highest weight, ≥50%

Encoded in all fourcheck:skill-frame-sync copies plus the card-face
serialization. The rule: when the axes conflict, lead the recommendation with the
long-term reading, and the other three axes together cannot outvote it.

Two guards land with it, because the ruling implies both and neither is safe to leave
to the reader:

1. Read the weight through that axis's own definition — 缩小而非扩大特例/契约增生,
   shrink special-cases and contract accretion rather than grow them. Without this, a
   50% weight on "long-term soundness" is exactly the sentence someone reaches for to
   justify building the grand speculative surface now, which is the opposite of what
   the axis says.
2. It ranks RECOMMENDATIONS, never authority. A 50% weight cannot make a
   human-floor matter dispatchable; 安全/权限边界 and 破坏性难逆动作 stay with the
   maintainer whatever the weight says. This is the boundary triage flagged as
   最要紧, and it is written in as a ⛔ clause rather than implied.

Composition with the 2026-08-27 tie-break, verified rather than asserted. The
existing 分歧推荐序 (「tong y 4」) is untouched and still ranks. The new rule is the
stronger constant floor: the 08-27 order says which way ① points given pull
evidence; this ruling sets the minimum weight ① carries. I tried to construct a
conflicting case and could not — see "Stop conditions" below.

⛔ No axis was reordered. The ruling is about weight, not list position; reordering
would move check:skill-frame-sync's axis NAME SEQUENCE across all four copies and is
a larger change than was ruled. The gate confirms the sequence is unchanged.

#14032 — exactly 5 per batch, full prose per card

The teeth the existing text lacked:

1. 恰好 5 张、呈完即停等回批 — the file said 「每批 ≤5 张」, which permits the
   incident (~20 items in one message). ⛔ 永不一条消息投多批.
2. 「详细解释」= 逐卡散文 — business problem + each option's real cost + the
   four-axis analysis written out + one recommendation. ⛔ A table row, a one-line
   verdict, or a 「变/未变」 grid does not discharge the duty: 压缩批次是违规不是摘要.
3. Re-analysis passes follow the same discipline as first presentations.
4. An applicability boundary, which triage explicitly asked for so the rule is not
   read as "no tables anywhere in skills": the duty attaches to 任何席位向维护者呈递
   需其裁决的决策项 — presenting items for a ruling — and ⛔ 不及结果陈述
   (轮报表格照旧), stating results already decided. Written as 任何席位 rather than
   naming the director seat, because triage measured the same concern arriving on a
   second channel (the 分诊席 batch-processing question the same day).

Landing map — one home per duty, measured not assumed

The stop condition was "the presentation duty has TWO existing homes". It does
not. Measured:

Face	What it owns	Action
references/decision-analysis.md 「常设决裁批流程」	the duty's mechanics (批序与写法)	✅ teeth land here
references/lanes/director.md 「基本原则」	the maintainer's 2026-08-27 charter quote, and an explicit pointer: 「呈报批序与写法见 ../decision-analysis.md」	untouched — the pointer already reaches the new text

director.md delegates rather than duplicates, so it is a pointer, not a second home;
its 「三要素:每批 5 张」 is already consistent with "exactly 5" and needed no edit.
decision-analysis.md is also the correct home on the merits: it is the cross-seat
file, while director.md is director-only, and triage established this rule's readers
extend past the director seat. No third home created; no copy made.

For #14030 the corpus already has an established two-place pattern for this exact rule
class (the 08-27 tie-break sits in both): the authoritative statement in the frame's
single source, and a card-face restatement. I followed it exactly rather than inventing
a placement.

Frame-sync

✓ check-skill-frame-sync: 4 copies of the decision frame are structurally isomorphic across 3 files
— 4 axes, sequence business-need → long-term-soundness → ai-authoring-safety → startup-scope-discipline, binding sentence present in all 4, 4 count mentions agree.

All four copies carry the weighting. That is deliberate and follows the precedent
recorded in the token-ratchet ledger for the 2026-08-27 rule: this published file
carries two enforced frame copies, so a rule that changes which recommendation the
frame yields must reach them, or third-party installers weigh the axes co-equally
while this repo weighs long-term ≥50% — the #5130 drift shape exactly, and a lying
green light (the sync gate compares structure, so it would stay green).

Line and token budgets — before/after, with funding source

File	Metric	Before	After	Ceiling	Funding
references/decision-analysis.md	lines	46	54	46 → 54	⚠️ ruling-quoted raise
.claude/agents/os-dev.md	lines	466	469	466 → 469	⚠️ ruling-quoted raise
.claude/skills/pm-dispatch/SKILL.md	lines	965	971	1005 (unchanged)	existing headroom
skills/objectstack-pm-dispatch/SKILL.md	tokens	14390	14549	14391 → 14549	⚠️ ruling-quoted raise

Verdict lines on the final tree:

✓ .claude/skills/pm-dispatch/SKILL.md is 971 lines (ceiling 1005; headroom 34).
✓ .claude/skills/pm-dispatch/references/decision-analysis.md is 54 lines (ceiling 54; headroom 0).
✓ .claude/agents/os-dev.md is 469 lines (ceiling 469; headroom 0).
✓ skills/objectstack-pm-dispatch/SKILL.md is 14549 tokens (ceiling 14549; headroom 0).

Three ceilings move and all three are flagged for the human merge. Each is
recorded in its map with the standing convention (the raise, the quoted ruling, the
arithmetic, and what funding was attempted first). Both rulings explicitly order these
encodings, which is the ratchet's sanctioned quoted-ruling escape.

Funding, ladder-first

• decision-analysis.md (+8) — no reflow slack exists: all 46 lines are already
  packed to their own minimum at the 120-byte cap (the file was densified at phrase
  boundaries by Densify the pm-dispatch governed corpus at phrase boundaries (whitespace-only) + the ruled funding-ban distinction #13389; the measured saving available is zero). No genuine deletion
  either — the rulings make no existing clause redundant. The one in-place payment left
  was deleting a ruled clause, refused on the state-machine precedent.
• os-dev.md (+3) — ⚠️re-wrap funding was available here and was deliberately
  refused. Three nearby paragraphs carry wrap artifacts (two orphan lines of 6 and 8
  bytes) worth exactly the 3 lines needed; merging them lands this file at 466/466 with
  no raise at all. I measured it both ways, then reverted it: that is 筹行, banned
  by the 2026-08-17 ruling — 「⛔ re-wrap(折行合并)不得用作筹行 —— 棘轮治理的是内容体量,
  行数只是机读代理,新增以删减付账」. The density fix is legitimate on its own and belongs
  in a net-reducing PR (2026-08-29), not this one. The cheaper-looking green was the
  wrong answer and this PR does not take it.
• SKILL.md internal (+6) — paid entirely from existing headroom, no raise. The
  full statement of the weighting deliberately lands here (the frame's single source)
  rather than in the ratcheted-to-zero reference file.

skills/** — the two required readings

Per the 2026-08-21 ruling (「对外发布的 skills 是整个平台的最大价值,尤其要整体考虑和评估。」
/「……不能为了一个小功能扩写很多。」):

• Changed file, whole-file:skills/objectstack-pm-dispatch/SKILL.md14390 →
  14549 tokens (+159); lines +11/−1.
• Whole published bundle: ratcheted (authored) subtotal 171410, bundle total
  (whole shipped tree) 187891. This is the only skills/** file in the diff.

Proportionality: +159 against the +152 the comparable 2026-08-27 frame rule
cost across the same two copies — essentially the same size for the same rule class.
It got there by compression, not by first draft: the initial cut measured +242, and
the provenance was then removed on the #5451 route-B convention — the published copy
carries the rule without this repo's dates or quoted rulings. Verified empirically:
the published file contains zero date-shaped strings and zero 「」 quotations,
and the only ones present before this correction were the ones I had just added. The
verbatim rulings live in the internal copies and in this PR body, which is where the
corpus keeps provenance.

Stop conditions — both checked, neither tripped

"The weighting conflicts with the 08-27 tie-break in some constructible case." I
tried and could not construct one. The candidate is the zero-pull branch: 08-27 says
recommend ④不扩散, and one might expect ①'s ≥50% to override toward building the
general capability. It does not, because axis ①'s own definition is 「缩小还是扩大
特例/契约增生」 — an unconsumed declared surface is contract accretion, so ① and ④
point the same way there. The pull-bearing branch (荐①长远终态) is already the
weighting's own direction. The two compose: 08-27 supplies the direction, this ruling
supplies the floor. This is also precisely why guard (1) above is written in — the
only way the conflict becomes real is if someone reads ① against its own definition,
and the encoded text now closes that reading.

"The presentation duty has two existing homes." It has one, plus a pointer — see
the landing map. No third home created.

Premise verification, and one triage input already satisfied

Both premises re-checked against today's origin/main (the file was reflowed tonight;
I read the current text, not the line numbers on the cards):

• skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 — valid. No weight floor anywhere; only the 08-27 tie-break.
• skills: harden the decision-batch presentation rule — exactly 5 per batch, each card explained in full prose; bulk lists/tables are a violation, not a summary #14032 — valid. 「每批 ≤5 张」 present, all three teeth absent.
• ⚠️The card's landing hint is wrong on one point, worth correcting for the record:
  skills: encode the maintainer's axis-weighting ruling into the four-axis decision framework — long-term soundness carries the highest weight, at least 50% #14030 suggests landing in references/decision-analysis.md because it is "outside
  the ratchet". It is not — the line ratchet covers references/ too, at 46/46 with
  zero headroom. That is why this PR pays a raise there and puts the bulk in SKILL.md
  instead.
• ⚠️Triage's ③-axis implementation input is already in the tree and needs no work.
  Triage recommended listing the four canonical axes verbatim inside the 四棱 card-block
  template. They have been there since 2026-08-20 (feat(pm-dispatch): decision analyses argue from the business standpoint (2026-08-20 ruling) #10186) — decision-analysis.md
  lines 38–41 spell all four in the block itself. So the seat's drift to self-invented
  axis names was not caused by the canon being absent from the point of use, and the
  proposed fix would have been a no-op. What the block genuinely lacked at the point of
  use is the weight, which this PR adds immediately adjacent to those four lines.

Verification

Full derived gate union re-derived on the final tree (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, no hand-built path list; clean derivation,
no stale-tree warning): 29 families, 27 exit 0.

The two non-zero are both self-declared NOT MEASURED, not findings:

• node scripts/check-test-completeness.mjs → exit 3, needs a saved turbo run test
  log; its own text says "the local reading for this gate is NOT MEASURED … ⛔ It is not
  a red".
• node scripts/pm/check-governed-queue-guard.mjs → exit 1, requires
  GITHUB_EVENT_PATH; CI-context only, and it correctly refuses to exit 0 when it cannot
  look.

Exit codes were captured before any pipe throughout (cmd > log 2>&1; rc=$?).

Also run beyond the derived union:

• pnpm lint (repo-wide eslint . --no-inline-config) — exit 0, no narrowing
  needed, run under the shared verify lock (VERDICT command-exit 0 · held the lock 69s).
• Both edited gate scripts' own --self-test legs: line ratchet 111 cases pass,
  token ratchet 64 cases pass. The token self-test initially failed one case — the
  SKILL.md subtotal pin — which is by its own documented convention shifted by exactly
  the authorized raise (118910 → 119068), the same operation the two prior raises
  recorded.
• Verbatim-quote integrity: both rulings verified byte-identical to the card bodies
  fetched from the API, and each occurrence confirmed to sit on a single line
  ([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 — a governed quotation that soft-breaks stops being findable by the things
  that grep it; one had wrapped in a comment and was fixed in its own commit).
• Control-byte self-scan over every changed file: clean.

Merge notes

⚠️ Governed faces (.claude/** + published skills/**) ⇒ draft only, human merge.
Three ceilings move; each raise quotes its authorizing ruling in the map beside the
number, and the os-dev.md one additionally records the re-wrap funding that was
available and refused.

No changeset: the diff publishes nothing from any package, and the five most recent
commits touching this published file all carry zero changesets. skip-changeset applied
and read back.

Generated by Claude Code

---

Generated by Claude Code