docs(skills): the four small internal skills — seven dead references, four copies replaced by pointers, and the live-elsewhere never-remove verdict

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Fixes#14317
Fixes#14058

Member card of the skills catalog optimization program (maintainer mandate 2026-09-02, verbatim and untranslated: 「审核所有的 skills，进行全面的优化。」). Sibling PR in objectui carries VFY-C-01 (Part of #14317).

Clause-②: no — nothing here states a platform contract; no needs:contract-review.

Per-finding record

id	落点	before	after
SPR-C-05 (= #14058)	spec-property-retirement §0	four checkboxes; zero mentions of the ledger's fifth verdict	a fifth checkbox: a live-elsewhere row is dead here by measurement, genuinely enforced in a sibling repo, and never a removal candidate; what to check instead is its four attestation criteria and the gate that executes them
SPR-C-01	:329-332	"check-generated runs behind a paths filter in ci.yml; add the path in the same PR"	deleted — that job and that filter no longer exist, and ci.yml has no paths: at all
SPR-D-01	§6 traps	4 of 7 traps restated a binding anchor (stale dist, pipe-eaten exit code, sequential gate masking, --check vs bare)	one pointer naming all four and their two anchors; the two traps unique to retirement stay
SPR-C-04	:300-311	a hand-rolled ten-name check:* loop — the "match by hand" the root instruction file forbids	the one wrapper command, which reports every stale artifact at once
SPR-C-03	:313-314	3 of the 5 pure source audits named	all five named
SPR-C-02	:81 (the route table)	a hand-written $ZodErrorMap passed to z.object(shape,{error}).strict()	the shared strictObject()guidance: slot, its set-keyed form, its worked example and the audit test that now covers it
SPR-C-06	:145	"~28 tombstones follow all five conventions" — a hand count, measured 181 sites today	the census is delegated to the pin test that actually judges every site
SPR-F-01	§4 changeset row	no ADR-0087 disposition marker	the marker and the gate that enforces it, on the row where retirement is exactly the named case
DOG-C-01	dogfood:66-83	18 lines re-stating the dev-agent definition's ablation rule — same criterion, same ledger, same two incidents	a two-line pointer at that anchor
DOG-F-01	dogfood §0	zero mentions of OS_PORT, though the runner doc names this section by name as the trap's origin	the two lines: the self-pinging connectors read the environment, so -p alone leaves them dialing 3000
X-01	dogfood §0	environment facts split across two docs, with the routing and the pricing pointing opposite ways	decided by pointer: the runner doc is the environment anchor, this section says so and carries only what that anchor lacks
CLA-C-01	checklist-author:33-51	declares itself "not a second copy of the process", then restates the sweep doc's six steps one for one	a pointer plus the three clauses measurably absent from that doc
CLA-C-02	:46-49	"four of six waivers expired" — the oldest of three disagreeing numbers	gone with its containing duplicate; the surviving anchor states the measured truth, and the checklist gate reports 0 waived today
CLT-B-01	checklist-test:78-87	Tier-1 batching reads automated.ref "from the selector" — which projects only id / priority / surface / since / revision	the step now reads the field where it lives, in the area file
CLT-C-01	:96-97	a blanket opus tier, sourced from memory	the tier is derived per dispatch from the current tooling output, never recalled
CLT-C-02	:229-235	7 lines re-copying the disclosure rule, opening with a precedence claim the runner doc explicitly denies	the guardrail stays in 4 lines (see the note below), the elaboration and the precedence claim go
CLT-C-03	:102-105	re-copies runner rule 1	a pointer
X-02	both files	the no-subagent degradation clause stated twice, near-verbatim	one copy, in checklist-author; the other points at it. Provenance noted, not re-litigated: both copies came from the same maintainer round

Dead references — each with its positive control

stated	corrected to	control
SPR:103scripts/liveness/orphans.mts	packages/spec/scripts/liveness/orphans.mts	stated path absent, corrected path present; every other citation in this file is repo-rooted
SPR:261packages/cli/src/utils/lint-liveness-properties.ts	packages/lint/src/lint-liveness-properties.ts	stated path absent, corrected path present
SPR:282.changeset/tool-inert-keys-removed.md	any live .changeset/*-retired.md	stated file absent (consumed at release); 6 live files carry the "The retirement kit:" section
SPR:329-332 the ci.ymlpaths filter	deleted with SPR-C-01	ci.yml has no paths:
CLA:42 the README's "deep-test contract"	the README's ## Item anatomy	the term has zero hits in that README; the section exists at the cited line
CLT:78-87automated.ref from --json	the area file	the selector's projection is {id, priority, surface, since, revision}

Line delta per file

file	before	after	ceiling	widest table row
.claude/skills/spec-property-retirement/SKILL.md	337	333	337	328 → 326 (pin 328)
.claude/skills/dogfood-verification/SKILL.md	157	147	157	n/a
.claude/skills/checklist-author/SKILL.md	62	53	62	n/a
.claude/skills/checklist-test/SKILL.md	238	234	238	221 (unchanged)
total	794	767

Every file shrinks; no ceiling is raised and the ratchet script is untouched. The SPR-C-05 addition is paid in the same file by SPR-C-01 and SPR-C-04, so the ceiling raise 337 → 340 authorized on #14058 (2026-09-01, director batch #26) is not taken — the ruled TEXT lands, including the recognizability half that ruling insisted on ("dead here by measurement / genuinely enforced in a sibling repo"); only its funding route changes, from a raise to the sanctioned pay-by-deletion. The gate now reports 2 bytes of table-row headroom on that file; per this flight's constraint that slack is left for the corpus audit's Phase 3 to re-lock, not taken here.

premise_false: none — every finding's premise held when checked against this base.

Three notes where the work differed from the brief, each measured

• CLT-C-02 is trimmed, not deleted. The runner doc states in its own text that this guardrail is "one rule written in both places, not a precedence claim by either". Deleting this copy would falsify that anchor, which this flight may not edit. So the copy shrinks to the operative rule plus that framing, and the false precedence claim — the half the anchor actually denies — is what goes.
• The all-five-verdicts check ordered on [finding] spec-property-retirement skill predates the live-elsewhere verdict — its ledger-verdict discipline never names the never-remove cross-repo case #14058 was carried out: the ledger's status vocabulary is live / experimental / planned / dead / live-elsewhere. inconclusive, noted as a zero-hit term during triage, is not a ledger status at all (zero hits in the vocabulary table and in the liveness scripts), so no vocabulary is missing on its account.
• check:pm-skill-id-lint does not bind anything in these four files — its scan set is the dispatch-protocol tree plus the dev-agent definition and the root instruction file. No section id or finding id in these four is load-bearing for it, which is why sections could be restructured. Verified before the first deletion.

Gates

All at head 7df16bc7, which equals the pushed remote head and the tree every command below read. Exit codes captured by redirect before any pipe. Heavy runs serialized through the shared verify lock (slot issue-14317).

command	exit	its own verdict
pnpm check:pm-skill-ratchet	0	all four files green, e.g. spec-property-retirement/SKILL.md is 333 lines (ceiling 337; headroom 4) and widest table row is 326 bytes (pin 328; headroom 2)
pnpm check:pm-skill-id-lint	0	green
pnpm check:skill-frame-sync	0	green
pnpm check:skill-frame-freshness	0	green
pnpm check:pm-governed-merges	0	green
pnpm check:doc-authoring	0	green
pnpm check:agent-test-spelling	0	green
pnpm check:cross-package-test-inputs · node scripts/check-cross-package-test-inputs.mjs	0	green
node scripts/check-ci-filter-parity.mjs	0	green
node scripts/check-required-contexts.mjs	0	green
node scripts/check-shard-attestation.mjs	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	58 cases passed self-test, then 22 record-scoped formula example(s) across 426 files / 1385 TS blocks judged clean, 9 @example(s) judged clean, 14 predicate(s) judged clean
pnpm --filter create-objectstack exec vitest run --maxWorkers=2 src/template-consistency.test.ts	0	Test Files 1 passed (1) · Tests 35 passed (35) — the metadata.internal: true enforcement; all four files still carry the marker
node scripts/check-test-completeness.mjs	3	NOT MEASURED, not a red: PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named. It does not run tests and cannot produce one; CI tees the log and passes the path

The gate family was re-derived from the real changeset after the last edit (node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands, 4 paths vs merge base aaa4e6578, three-dot semantics) and the whole derived list is above. The first union run returned six exit 3 PREREQUISITE-NOT-MET results because the fresh worktree had no node_modules; they were re-run green after pnpm install, and the formula/lint dependency closures were built before the doc-formula gate could measure anything.

Repo-level pnpm lint was not run, and this is declared rather than implied: it is the CI-owned run. No narrowed substitute is claimed for it.

skip-changeset: this diff is .claude/** only and releases nothing, which is the route the empty-changeset gate's own enumeration names for it (.github/, .claude/, skills/, docs/, content/, examples/ → label, not an empty changeset).

Governed surface (.claude/**) ⇒ this PR stays a draft for human merge; review requests are the dispatching seat's step.

Out-of-scope cards filed after a dedupe search (unassigned, unlabelled for triage): #14595 (the sweep doc sends agents to land a run record in a gitignored directory) and #14596 (a checked-in launch config pointing at a worktree that does not exist).

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code