tooling(pm): H8, H19 and H37 quote the LANDED release act, both halves

[claude]

Fixes #15899

scripts/pm/check-half-states.mjs presented a shorthand inside quotation brackets at four
places. In this file's register those brackets signal a verbatim
.claude/skills/pm-dispatch/SKILL.md quotation — and that phrase is in no SKILL.md. Two
measurements, both cheap to repeat:

• git grep for it outside this one script: zero hits. SKILL.md and
  references/state-machine.md carry 同笔摘标, a label removal on pm:retriage — a different
  act on a different field, deliberately left alone here.
• git log -S for it over SKILL.md's whole history, non-shallow: empty. It was the script's
  own coinage, wearing a promise the source could not answer.

Why this is not a spelling nit

The release act is ONE act with TWO halves:

释放是显式动作:让卡离手者同笔清 assignee + Release: 行(会话/因/去向);下一任重新认领。

Clearing the assignee is half of it; the Release: line is the other half. The coinage named
only the first, so a reader following those remedy sentences exactly produced the very shape the
H47 row exists to report — the field cleared, no Release: line, an exit that left no record.
The sweeper's own remedy text prescribed a half-state.

What changed — one file, no predicate changes

A new module constant RELEASE_ACT_RULE carries the landed line verbatim, kept unbroken on one
line so it stays greppable against its source. Its text was extracted from the transcriptions
already in the tree (the RELEASE_COMMENT_MARKER docblock and selfTest's RELEASE_RULE_LINE)
and asserted equal to SKILL.md's line during the edit, never retyped — a hand-retyped quotation is
how a verbatim quote stops being one. Three emitted sentences now quote that constant and name
both halves:

Site	Row	Was
full-delivery remedy	H8 (h8MergedPrStillDispatched)	assignee drop only
release text, closed and unresolved branch	H19 (h19BlockOutlivedBlocker)	assignee drop only
dispatched-residue remedy	H37 (h37FamilyMemberDrift)	assignee drop only

The fourth site is the H24 header-inventory paragraph, which narrated where the coinage came
from; it now reads the rule instead of the history.

No predicate changes, no new rows, and CLAIM_COMMENT_MARKER / RELEASE_COMMENT_MARKER /
ACCEPT_VERDICT_MARKER / MAINTAINER_BRIEF_MARKER are untouched. H24's own remedy and H47's two
legs already quoted the landed line and were not edited.

Verification — all at head 67bec961f4

• Self-test 2414 → 2427 cases, all passing. Case-name set diffed: exactly two names left
  (the two that spelled the coinage), fifteen arrived; the other two of the four pinned cases keep
  their names with re-pointed assertions. Each re-pointed sentence gained a case that the rule
  line is quoted UNBROKEN, one that the second half is named, and one grep-negative that fails if
  the coinage returns.
• The grep-negative needle is assembled, not spelled. The measure of this repair is
  git grep -c for the coinage over this file reading zero; a literal needle in a negative case
  would be the single hit that defeats the check it protects. It now reads zero, and the
  neighbouring 同笔摘标 occurrences are untouched.
• Ablation. The coinage was restored at H8's site on a scratch copy laid out beside a copy of
  scripts/invoked-as.mjs, with an unmutated control copy at the same path. Mutation confirmed on
  disk by occurrence counts either side (coinage 1 / landed phrase 0 on the ablated copy; 0 / 1 on
  the control). Control: 1 failure — H43 register, environmental, a sibling-repo copy cannot see
  the governed register. Ablated: 4 failures, i.e. the same environmental one plus exactly the
  three H8 pins (names the release act in the spelling that LANDED, quoting the rule line unbroken, and the grep-negative flipping to got true, want false). No rebuild leg applies —
  node loads this .mjs source directly, there is no dist/ in the resolution path. The tree
  file's blob hash was identical before and after (3161d0b325…, equal to its HEAD blob).
• Gates. dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 32
  commands from the committed change set; all 32 ran, all exit 0, and --ran reconciles:
  32 derived, 32 run, 0 NOT-MEASURED, 0 UNRUN. The 7 roster gates whose baseline sits under
  scripts/ were also run green. check-governed-merges.mjs --test on this file list:
  NOT governed.
• Whole-repo pnpm lint through the shared verify lock: VERDICT command-exit 0. Not a
  narrowed run — the full farm.
• NOT MEASURED, named rather than skipped: check:dts-closure and
  check:published-readme-exports both exit 3 PREREQUISITE NOT MET here (no package dist/;
  they need a whole-repo build). Neither can be moved by a one-file scripts/pm/ diff, and CI
  builds before running them.

Known-red step, inherited and untouched

main has been red on Lint & Repo Gates at pnpm check:merge-driver since 17:05Z —
Corepack resolving pnpm 12.3.4, which rejects the fixture's pnpm -s. It is tracked as #15992 in
the devx lane and is not this branch's; nothing here works around it. For the record, that gate
passes locally on this branch under pnpm 10.31.0. Every other check is expected green.

skip-changeset: scripts/pm/** publishes nothing from any package.

---

Generated by Claude Code

[os-steve]

ACCEPT — lands #15899 as ruled. Pure code (one file under scripts/pm/, governed-surface predicate exit 0), so ready + SQUASH auto-merge follow in this same act; the merge waits on main's own red (below), not on anything in this diff.

What the seat verified, at head 67bec961f4, in its own compare worktree (merge base f7db8f4fd2)

• One file, +89/−15. The retired coinage 「同笔摘 assignee」 reads 0 in the file; the neighbouring 「同笔摘标」 (a label removal on pm:retriage, a different act) is untouched at 2.
• RELEASE_ACT_RULE's literal is a verbatim substring of SKILL.md line 466 (the seat compared the two strings, not the report), and the self-test's independent transcription RELEASE_RULE_LINE equals it — so the pins compare two sources rather than a constant with itself. The three emitted sentences (H8's remedy, H19's shared release text on both branches, H37's dispatched-residue verdict) quote the constant and name both halves of the act; H24's header paragraph reads the rule, not the history.
• No predicate changed; zero diff lines touch any of the four comment markers; self-test 2414 → 2427 (+15 / −2 renamed, none dropped), exit 0 in the seat's run; check-governed-merges --test exit 0.
• CI, seat's read 19:0xZ: 35 check runs — 23 success, 11 skipped, 1 failure: Lint & Repo Gates at its Merge-driver wiring gate step, red on main itself at f7db8f4fd2 (main red: Lint & Repo Gates › Merge-driver wiring gate — check-regen-pending.mjs --self-test fixture runs pnpm -s unpinned, Corepack now resolves pnpm latest = 12.3.4 which rejects -s; merge queue blocked since 17:05Z #15992, p0, devx lane). Not this PR's; every other step and check is green. Auto-merge is armed now and will enqueue when that gate is green on this head — the seat re-runs the failed job or updates the branch once main red: Lint & Repo Gates › Merge-driver wiring gate — check-regen-pending.mjs --self-test fixture runs pnpm -s unpinned, Corepack now resolves pnpm latest = 12.3.4 which rejects -s; merge queue blocked since 17:05Z #15992's fix lands.
• Body line 1 Fixes #15899; skip-changeset present.

Open question answered: A for this PR; C filed as #16016 (a gate for bracketed pseudo-quotations that no SKILL.md line carries — the class, not the instance). The two tidiness items (inline release-rule spellings in H24/H47; drifted line numbers in two docblocks) are folded into #16003's flight on this file.

Implemented-by: os-dev executor, flight #15899, branch claude/issue-15899-release-quote-both-halves
Reviewed-by: pm-dispatch skills seat, https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

---

Generated by Claude Code