docs(pm-dispatch): release is an explicit act — clear the assignee and post a Release: line whenever a card leaves a session's hands (#15846)

[claude]

Fixes #15846

What

Ownership had an explicit entry (assign + Claim:) and no explicit exit. Under the maintainer's 2026-09-05 design ruling (R3, recorded on the ownership-contradiction card by the skills seat: the assignee field stays a presence bit under one shared identity — 「暂时不区分」 — and a presence bit is only truthful if its exit is written down), release is an explicit act: whoever moves a card out of a session's ownership — re-route, hand-back on premise_still_valid: false, an abandoned flight with no takeover, a cross-lane hand-off — clears the assignee and posts a line beginning Release: (session ID, which of the four reasons, destination = the new label state or lane) in the same write; the next owner claims afresh. Two definitional readings follow: a pm:queue card never carries an assignee; a pm:dispatched card always does (what H1 and H24 read).

Line-neutral on both files, both at headroom 0. Every new fact is folded into an existing line or takes the place of a line whose content survives in its neighbour; no ceiling moves and the ratchet script is untouched.

.claude/skills/pm-dispatch/SKILL.md — 811 → 811 lines, widest table row 342 → 342 bytes

State-model table (状态模型):

• open + 队列标签 + 无 assignee row, 85 → 128 bytes: adds 「pm:queue 卡恒无 assignee,有即半态」.
• assignee 已设 row, 63 → 216 bytes: adds 「离手恒走释放 —— 四因 = 改路由、前提证伪、弃飞无接管、跨车道移交,去向 = 新标签态或车道;⛔ 不静默摘 assignee」.
• pm:dispatched row, 106 → 122 bytes: adds 「恒带 assignee」.

认领 family (beside 「assignee 字段归 PM」, which is unchanged):

• New line, 118 bytes: 「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。」
• Paid for by folding 「选中的单派发前按序执行一个原子对。」 into the ① line, now 「派发前按序执行原子对:① Assign @me,并把 pm:dispatched 与摘 pm:queue 放进同一次标签写入。」 (119 bytes). Nothing dropped.
• Dead-claim reclamation line, 113 bytes: 「摘 assignee(注明原因)回队」 → 「释放回队(Release: 行载因)」. Reclamation is the abandoned-flight reason of this same rule; left as it was, the file would spell the same exit two ways.

.claude/skills/pm-dispatch/references/state-machine.md — 42 → 42 lines

• New line after 摘标者, 119 bytes: 「释放者 = 让卡离手的席位:行首 Release:(会话/因/去向)与清 assignee 同笔;下一任重新认领。」
• Paid for by folding 「awaiting 卡在在飞视图里计停放库存。」 into the retriage-authority line, now 「retriage 原定级在改判前仍是权威;awaiting 卡在在飞视图里计停放库存。」 (93 bytes). The only phrase dropped is 「挂/摘分属两方」, which the 挂标者 / 摘标者 lines directly beneath already state.

Byte widths (LC_ALL=C awk, bytes)

• SKILL.md widest non-table prose line: line 98 at 120 (pre-existing, unchanged). Widest table row: line 250 at 342 (the domain:engine row, pre-existing, unchanged). Line 810 at 172 bytes sits inside the claim template's code fence (pre-existing, fence-exempt). No line this PR writes exceeds 120 bytes; the widest new line is 119.
• state-machine.md widest line: line 18 at 120 (pre-existing, unchanged); the new line is 119.

Where the four reasons and the destination live — the dev's call inside the ruling

The 120-byte cap on a 认领 line cannot hold the act, the four reasons and the three fields at once; the assignee 已设 row has ~240 bytes of room under the 342-byte pin. So the act and the fresh claim sit in the 认领 line and in state-machine.md (the 行契约: 行首 Release: with its three fields, same write as the clear), while the four reasons, the destination definition and the ⛔ on silent clearing sit in the assignee row of the state model, and the two definitional readings sit in the pm:queue / pm:dispatched rows they define.

Four-axis note. 实际业务需求: measured, not speculative — H1/H24 report residue as live half-states and a seat's list view shows taken cards that are free; all four exits already exist in the protocol (改路由 in the pm:retriage row, premise_still_valid: false in 收集 / 报告契约, 死认领回收, 跨车道 接手). 项目长远合理性: the exit is written where the entry is written — the state model and the 认领 section — not as a patrol heuristic. 防 AI 犯错: the definitional readings turn the field into a checkable invariant, and 「⛔ 不静默摘 assignee」 forbids the tolerant path. 创业阶段不扩散: no new label, field, comment marker or gate row; the patrol row for a Release: newer than the last Claim: stays the declared follow-up.

Not changed

The claim's first-line spelling Claim: (ruled 2026-08-11); CLAIM_COMMENT_MARKER; the H1/H24 code in scripts/pm/check-half-states.mjs (its prose is a follow-up once this lands); the two declared frame-block copies; the ratchet script's CEILINGS and pins.

Verification — all at ec0a7f8d8c, the PR head

Family derived by node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack on the two files (16 commands: 9 by path, 7 whole-tree), cross-checked by the no-path derivation from the merge-base (same 16), every command run with the exit code captured before any pipe, then reconciled: ✓ dispatch-gates --ran: 16 derived famil(ies) accounted for — 16 run, 0 NOT-MEASURED.

• ✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md: widest table row is 342 bytes (pin 342; headroom 0).
• ✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/SKILL.md is 811 lines (ceiling 811; headroom 0).
• ✓ check-skill-line-ratchet: .claude/skills/pm-dispatch/references/state-machine.md is 42 lines (ceiling 42; headroom 0). (widest table row 0, pin 0)
• ✓ check-skill-line-ratchet self-test: 155 cases pass.
• ✓ check-skill-id-lint: 24 file(s) clean (pattern /#[0-9]{3,}/g). (+ self-test 14 cases)
• ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
• ✓ check-half-states self-test: 2267 cases pass. — no H-row prose pin fired on the edited sentences.
• ✓ check-skills-token-ratchet: 36 authored bundle file(s) within their ceilings; 10 generator-owned file(s) measured, not ratcheted.
• ✓ check-governed-prose: 2 instruction surface(s) name all 5 registered governed surfaces … and claim no others.; check-governed-merges --self-test: 274 assertions; check-governed-queue-guard self-test: 144 cases pass; check-closing-keyword-parity --self-test: 24 assertions; comment-mask corpus sweep: 5998 files, 0 disagree; check-agent-test-spelling: 0 violations; doc authoring guard: 46 published skill files clean; check-nul-bytes --self-test: 75 assertions; check-refd-timer-probe self-test: 11 cases pass; check-watch-hint-literal: 61 declaration(s) across 4 rostered name(s); check:driver-memory-census exit 0.
• check:doc-formula-expressions: first run PREREQUISITE NOT MET (exit 3: @objectstack/formula, then @objectstack/lint, not built — NOT MEASURED, not a finding); after turbo run build --filter=@objectstack/lint through the lock: ✓ check:doc-formula-expressions self-test: 58 cases passed and ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 428 files / 1370 TS blocks judged clean by @objectstack/formula.
• Whole-repo pnpm lint (eslint . --no-inline-config) through scripts/pm/os-verify-lock.sh (slot issue-15846): os-verify-lock: VERDICT command-exit 0 · held the lock 67s (1m07s) · waited 65s (1m05s).
• Control-character self-scan of both files (grep -naP over the C0/DEL range): no hits.

Negative control on the ratchet, run from the committed state: one line appended to each file → ✗ … SKILL.md is 812 lines; the ratchet ceiling is 811 and ✗ … state-machine.md is 43 lines; the ratchet ceiling is 42 (exit 1); restored with git checkout HEAD -- …, proven by git hash-object equal to the HEAD blob on both paths (a32af80219…, b3b537e59a…) and git diff HEAD empty.

check-governed-merges --test on the two paths: exit 3 (.claude/**, governed) ⇒ draft PR, in-seat contract-tier review, human merge, skip-changeset (no package publishes anything from this diff).

Generated by Claude Code

🤖 Generated with Claude Code

https://claude.ai/code/session_019RfFHiRCSs3JXLK4cwcfox

---

Generated by Claude Code

[os-steve]

ACCEPT — contract-tier in-seat review of PR #15848, head ec0a7f8d8 (skills seat, 2026-09-05T11:3xZ).

Implemented-by: claude/issue-15846-release-is-an-explicit-act (os-dev subagent)
Reviewed-by: session_019RfFHiRCSs3JXLK4cwcfox (skills seat)

Verified in a detached compare worktree at the PR head, three-dot against origin/main (d701e657d), not from the report: two files, +8 / −8, line-neutral — SKILL.md 811 / 811 with the widest table row still line 250 at 342 bytes, state-machine.md 42 / 42. The rule lands where ruled: the 认领 family gains 「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。」, paid for by folding the 原子对 header into its ① line with nothing dropped; the state-model rows carry the two definitional readings (pm:queue 卡恒无 assignee,有即半态 / pm:dispatched 恒带 assignee) and the assignee 已设 row carries the four exits, the destination and ⛔ 不静默摘 assignee — the dev's placement of the four reasons in the row rather than the 120-byte prose line is the right call inside the ruling; the dead-claim reclamation line now routes through 释放回队 with the Release: line carrying the reason; state-machine.md gains the releaser line beside the applier and remover, paid for by folding the awaiting-inventory line into the retriage-authority line (only the redundant 挂/摘分属两方 dropped). No issue number in any changed line; no model names in the diff or the PR body; the Claim: spelling, CLAIM_COMMENT_MARKER, the H1/H24 code and the frame copies untouched. Governed on both paths (check-governed-merges --test exit 3); tier read MANDATORY from the SKILL.md path, met.

Seat gate runs on this head in the compare worktree: skill-line-ratchet exit 0 (both files at headroom 0, as ruled), skill-id-lint 24 files clean, half-states self-test 2267 (no H-row prose pin fired), governed exit 3. The dev's evidence beyond that: 16 derived families all exit 0 and reconciled 16 of 16, frame-sync green (two declared copies), the formula-expressions gate green after building its prerequisites (its one-build-short refusal filed as #15850 for devx), whole-repo lint through the lock, the negative control (+1 line on each file reds the ratchet, restore proven by blob hash), widest line this PR writes 119 bytes; CI on ec0a7f8d8: 29 check runs completed, 16 success, 13 skipped by the path filter, none failing (seat's read at 11:3xZ).

Landing regime: governed surface ⇒ this PR stays a draft; reviewers os-zhuang and hotlong requested by the seat; the human merge is the review record; no seat flips it ready, enqueues it, or arms auto-merge. Fixes #15846 closes the card on merge. A check-half-states H1/H24 prose follow-up is the seat's to file once this lands.

---

Generated by Claude Code

[os-steve]

Tally correction to the ACCEPT above (skills seat, 2026-09-05T11:2xZ): the seat's REST read of the check runs on ec0a7f8d8 at 11:27Z gives 29 completed — 17 success, 12 skipped, none failing (the verdict's 16 / 13 was read while the last run was still in progress). The verdict stands.

---

Generated by Claude Code