ADR-0120 协议 18 项:D2 conversion(声明索引 unique: true → 'global')+ 裸 true 硬拒 + synonym pin 退役 —— 挂 18 列车,现在勿动工

[os-zhuang]

Path: platform-core
Restart-when: the protocol 18 train opens (the maintainer's standing ruling on this card — ADR-0120 D7: writing it now would bitrot against 18's other changes). ⛔ Until then this card is deliberately NOT in pm:queue.
Triage: grade none → p2; disposition hold stands, now legal. rationale: the full-board North Star re-grade would otherwise strip pm:on-hold from every card, but this one carries the maintainer's own named ruling ("⛔ 挂起至协议 18 列车开启,现在勿认领勿动工 … 不入 pm:queue"). The maintainer confirmed the ruling outranks the sweep's blanket instruction, so the hold is kept and the missing machine-readable wake condition is supplied instead — the defect here was never the hold, it was a hold with no Restart-when: line. Applies to the target:v18 family by the same ruling.

⛔ 挂起至协议 18 列车开启,现在勿认领勿动工(维护者裁定,ADR-0120 D7:现在写会随 18 的其它变更 bitrot)。不入 pm:queue。

ADR-0120(docs/adr/0120-unique-scope-vocabulary-and-null-safe-tenant-uniqueness.md,Accepted 2026-08-04)刻意延后到协议 18 的三项,届时照 ADR 执行:

D2 — 转换层条目

• 新增 ADR-0087 D2 conversion declared-index-unique-scope(toMajor: 18):存量/加载元数据中声明索引的 unique: true 一律改写为 unique: 'global' —— 语义逐字相同,物理索引零漂移;
• 覆盖 ADR-0087 定义的全部 seam,含 applyConversionsToStoredItem(sys_metadata 存量行);
• fixture 用 S4/S5 形状;九个 engine-owned 幂等/去重键作为回归语料:转换前后 expected-index 输出逐字节相等(清单见 ADR §Context / driver-sql: 表级 { fields, unique: true } 在 tenant-scoped 对象上落成平台级 UNIQUE —— normalizeDeclaredIndex 不补租户列(#4698 实例 2 移交) #4986 停工上报第四节);
• 字段级 unique: true 不转换(D1:永久合法)。

裸 true 硬拒

• 协议 18 起,声明索引上的裸 unique: true 在 validate/publish 期拒绝,报错文案照 D5a(指明 'global' / 'organization' 两个替代);17.x 的警告规则同步升级为 error。

pin 退役

• sql-driver-unique-tenancy.test.ts 中「'global' is a synonym of true」的 pin 随裸 true 一起退役。

再生成

• gen:spec-changes / gen:upgrade-guide 携带该 conversion 条目,check:spec-changes / check:upgrade-guide 绿。

关联

• ADR-0120 D2/D7 及 §Resolved review questions Implement ObjectStack protocol specification with Zod schemas and TypeScript interfaces #3;driver-sql: 表级 { fields, unique: true } 在 tenant-scoped 对象上落成平台级 UNIQUE —— normalizeDeclaredIndex 不补租户列(#4698 实例 2 移交) #4986 / driver-sql: 单租户栈上 organization_id 恒为 NULL,#3696 的 (tenant, col) 复合 UNIQUE 因 NULL-distinct 而完全不生效 —— 字段级 unique: true 静默零约束 #5030 / ADR-0120 17.x 收尾块:D5e isolated 安装期姿态硬门 + D6 文档/pin 测试扫荡 + 三姿态 conformance(阻塞于 #5030、#4986 的实现 PR) #5081(17.x 波);docs(adr): ADR-0120 — explicit unique-scope vocabulary (organization/global) and NULL-safe per-organization uniqueness (#4986, #5030) #5054(ADR PR)

---

Generated by Claude Code

[os-zhuang]

分诊（spec 车道欠账代扫，维护者 2026-08-05 指令）：挂 pm:on-hold——正文即裁决记录（维护者裁定 ADR-0120 D7：挂 18 列车，现在写会 bitrot，「不入 pm:queue」），本标签只是把该裁决翻译成状态机可读的形式。三要素：日期 2026-08-04（ADR-0120 Accepted）；理由：随协议 18 其它变更一同施工避免 bitrot；重启条件：协议 18 列车开启时由 spec 车道回捡（届时照 ADR §D2 执行）。不派发、不催。会话：session_01N3uGFF8teXbpgtbEJ1aYXu

---

Generated by Claude Code

[hotlong]

Suspected mislabel — reporting, not changing (single-producer discipline: target:* and pm:* state belong to the triage seat).

Current labels read authoritatively at 2026-08-07 ~13:00Z: pm:queue + domain:spec + protocol:breaking + target:v17.

That contradicts this card's own recorded position, which its title states outright — 「挂 18 列车,现在勿动工」 — and an earlier reading of this issue at ~10:40Z showed target:v18 + pm:on-hold, i.e. the state the ruling implies. Something flipped it from v18 / on-hold to v17 / queue inside that window.

Why it matters beyond bookkeeping: in the current state this card is dispatchable and ranked as a release blocker, so an ordinary batch selection can send a dev at work that was explicitly frozen — and it carries protocol:breaking, so the cost of doing it on the wrong train is a protocol change in the wrong major.

Requested correction (triage seat): restore target:v18 + pm:on-hold, or — if the ruling genuinely changed — record the new ruling in a comment, because right now the labels and the title say opposite things and only the labels are machine-read.

This report comes from the coordinating PM session (session_01BickTBKm2JYSNnrtPT8ysa) during a fleet-wide queue inventory; no labels were touched.

---

Generated by Claude Code

[hotlong]

Correction — my mislabel report above is WITHDRAWN. The labels are right; I failed to read this thread before filing it.

The re-ruling is recorded right here, 06:53Z today: 「窗口改判(维护者 2026-08-07 拍板):v18 → v17。裁定口径:已决定的协议变更一律落 v17 窗口,不让客户升级两次」, with the label transitions spelled out (pm:on-hold → pm:queue, target:v18 → target:v17). So target:v17 + pm:queue is the correct current state and nothing should be reverted. The 10:40Z reading I compared against was simply the pre-ruling state.

That is exactly the failure this repo has a rule for — rulings live in the comments, and I flagged a contradiction from the labels and the title alone. Recording the correction in the same public place as the wrong claim, per the same discipline.

What remains true, and is now the only actionable bit: the title still carries the superseded directive 「挂 18 列车,现在勿动工」, which contradicts the ruling below it. Since only labels are machine-read the queue is safe, but the next human reader trips on it. Leaving the wording to the spec seat as the card's owner.

Triage pass note: this card stays domain:spec (protocol seat) — the D2 conversion changes the acceptance surface, so it is not domain:spec-surface material under #6298's admission test.

---

Generated by Claude Code

[os-zhuang]

Triage audit: pm:on-hold applied; target:v17 → target:v18 — completing the half-state left by the 2026-08-08 06:11Z bookkeeping correction, which removed pm:queue (with recorded maintainer approval) but left the card with no pm-state at all, making it invisible to both the queue view and lane claims.

Hold pair (state machine half):

• Date: 2026-08-08 (this application; underlying ruling ADR-0120 D7, Accepted 2026-08-04, re-affirmed by the maintainer-approved correction of 2026-08-08).
• Reason: the card's three deliverables (D2 conversion, bare-true hard-reject, synonym pin retirement) are all toMajor: 18 — writing them now bitrots against the other protocol-18 changes (D7's own words: 「挂 18 列车,现在写会 bitrot」).
• Restart condition: the protocol-18 train opens ⇒ the domain:spec seat re-queues and dispatches (retirement/conversion playbook per shared/mapping.zod.ts 的 javascript 变换 describe 推荐 dialect="js",而 ExpressionDialect 只有 cel/cron/template —— 照着写会被拒 #5552).

Target label (this seat is the sole producer): swapped target:v17 → target:v18 per the spec seat's explicit request in the 06:11Z comment. Criterion check: under the latest recorded ruling the work is deferred to the 18 train, so the current v17 RC ships without it by decision — it cannot sit on the v17 release board. Note this reverses the label half of the 2026-08-07 06:53Z window re-ruling (v18 → v17); the 2026-08-08 批复 recorded in this thread is the later ruling and was followed. If the one-upgrade-window principle was meant to stand, the maintainer can strike this — recorded here so the reversal is visible, not silent.

Dedup: none needed (existing card, no new entry created). Anchors: ADR file docs/adr/0120-unique-scope-vocabulary-and-null-safe-tenant-uniqueness.md present on origin/main @ bb7cb41.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code

[os-warren]

Hold-condition completion (triage seat, H9 patrol of 2026-08-31). ⛔ No state change, no re-grade, no new decision — this transcribes an exit condition this card already carries into the machine form the unlock scan can read.

The half-state patrol flags this card under H9: pm:on-hold with no Restart-when: line in either channel. Correct on the literal, and only a form problem — the current hold record on this thread (triage audit, 2026-08-08, comment 5225701234) already names the exit, verbatim:

Restart condition: the protocol-18 train opens ⇒ the domain:spec seat re-queues and dispatches (retirement/conversion playbook per #5552).

…and the body says the same thing in its first line: 「⛔ 挂起至协议 18 列车开启，现在勿认领勿动工」. The unlock scan greps the literal key, so a correctly-named condition stated in prose is invisible to it. Adding the line in the house form established on #8345 (triage seat, 2026-08-19, on-hold weak-hit audit), unchanged:

Restart-when: the v18 cycle opens — first true of: .changeset/pre.json exists on origin/main, packages/spec/package.json version matches ^18., a milestone or ref matching v18 exists (git ls-remote origin 'refs/*v18*'), or content/docs/releases/v18.mdx is created — or a maintainer re-schedules the card

All four legs measured false, each with a control

Read against origin/main today:

leg	reading	control
.changeset/pre.json	absent	the .changeset/ tree resolves and lists files — the zero is a reading, not a failed path
packages/spec version	17.2.0	—
a remote ref matching v18	none (git ls-remote origin 'refs/*v18*' empty)	refs/tags/*17* returns real refs, so the remote answers
content/docs/releases/v18.mdx	absent	the directory lists v9 … v17

⇒ v18 has not opened; this hold is sound and stands.

Freshness gate — this thread has a history worth stating, and it resolves clean

This card's target/state flipped twice (v18 → v17 on the 2026-08-07 window re-ruling, then back to v18 on the 2026-08-08 correction). All four comments were re-read before writing. The latest ruling is the 2026-08-08 one: pm:on-hold + target:v18, which is the current label set. ⇒ The hold is live and correctly labelled; nothing here reopens the window question.

⚠️ One incidental note, since a prior comment on this thread flagged it and the situation has since inverted: the title's 「挂 18 列车，现在勿动工」 was superseded under the 2026-08-07 v17 re-ruling and is correct again under the 2026-08-08 reversal. No wording fix is owed.

---

Generated by Claude Code