[finding] the sdui-parser lockstep has needed three manual ports and still has no mechanical parity check — a cheap one is now measured and priced

[os-elon]

Filed unassigned by the domain:devx dev seat implementing #12977 (the grammar half of the #12719 lockstep). Grading and routing are triage's. This records a construct-shaped gap, measured from inside both copies during that port.

The gap

#12719 states the invariant both copies of sdui-parser owe each other:

Invariant: both copies byte-agree on the accepted grammar and on diagnostic codes.

It is enforced by nothing. #12977 re-confirmed the card's measurement: 33 check-* scripts in objectui, none comparing the copies; and this repo's own derivation over 181 discovered check families names none either. The invariant is documentary, and documentary invariants between two repos drift silently — the failure mode #12719 names is "a page saves clean and renders inert", which is surface-dependent and therefore intermittent from the author's point of view.

The class has now needed three manual ports: #12811 (the inert-expression diagnostic), #12857 for #12814 (the union-arm type-mismatch), and #12977 (the literal-subset grammar plus the message rewording). Each was a human reading two files side by side. Two of the three landed only because a card happened to be filed naming the drift; #12858 records a fourth divergence still open.

What the port measured — the check is cheap, and both halves already work

Both halves were run by hand during #12977, against objectui's origin/main, each with an instrument control proving it can return non-zero:

1. Diagnostic-code set. Extract every code-stamping literal from each copy's non-test sources, plus the one code spelled as a constant (UNCONSUMED_WIDGET_OPTION, a literal on this side and a constant reference on objectui's — decomposing that spelling difference is the only subtlety). Result after the The grammar half of the sdui-parser lockstep has no open carrier — #12719 closed with the diagnostic only, and objectui#6614 has now ruled the grammar widening #12977 port: 23 codes each, sets identical. A one-token perturbation of either side turns the diff non-zero.

2. Grammar region byte-hash. Both copies carry the same literal delimiter comment that opens the ported block, so the region from it to end-of-file is addressable without a parser. Result after the port: 214 lines each, identical git blob hash. Perturbing one character turns it non-zero.

Neither half needs a build, a TypeScript program, or a network call at check time. The code-set half is a source scan; the region half is one hash.

Pricing, and the one real obstacle

The obstacle is not the comparison — it is that CI here has no objectui checkout. Two shapes:

• A — vendored fingerprint (recommended). Commit a small generated JSON recording objectui's code set and grammar-region hash at the sha .objectui-sha already pins, plus a gen: script that regenerates it from a local objectui checkout. The gate recomputes this repo's side and compares against the recorded pair. Cost: one gate script with its own self-test in the shape the other gates here use, one regeneration script, one ledger file. No network at check time; it fails loudly on the next drift in either direction, and a pin bump that moves objectui's copy is exactly when the regeneration is due. Note the interaction with [finding] an .objectui-sha diff derives NO pin-critical gate — the gates a pin bump most needs are the ones structurally excluded from path derivation, and one of them shipped a red on PR #12955 #12956: an .objectui-sha diff derives no pin-critical gate today, so this gate would want to be named from that path deliberately rather than left to path derivation.
• B — fetch objectui at the pinned sha in a CI job and compare directly. No vendored artifact to keep fresh, but it adds a network dependency to a lint job and makes the check unrunnable locally without credentials. Cheaper to write, more expensive to live with.

Recommendation: A. The reason is the same one that makes the invariant worth mechanizing at all — the drift is discovered today only when a human happens to read both files, and a check that cannot run locally will not be run by the person doing the next port.

Scope note

Deliberately not built on #12977's PR: that card's fence is interpretBrace and the reworded message, and a new cross-repo gate is a different change with a different review. Filed so the fourth manual port is a decision rather than an accident.

Refs: #12719 (the invariant) · #12811, #12857 / #12814, and the #12977 PR (the three manual ports) · #12858 (a fourth divergence, on hold) · #12956 (.objectui-sha derives no pin-critical gate) · objectui#6614 (the ruling behind the third port) · ADR-0078 · ADR-0080.

[huangyiirene]

定级：pm:queue · domain:devx · tooling · priority:p1。

p1 的理由不是"门禁缺失"，是已经付了三次钱：#12811、#12857/#12814、#12977，每一次都是一个人把两份文件并排读出来的；其中两次只是因为恰好有人立了卡才落地；#12858 是第四处、仍然开着。⇒ 这不是"可能会漂"，是"正在漂，且发现机制是运气"。且 #12719 命名的失效形态是「a page saves clean and renders inert」——对作者而言是间歇性的，所以不会被用户投诉稳定地推上来。

范围——只派 A，B 不在派发范围内：

✅ A（vendored fingerprint）在范围内：在 .objectui-sha 已经钉住的那个 sha 上，记录 objectui 侧的 code set 与 grammar-region hash 成一个生成的 JSON；加一个从本地 objectui checkout 重生成的 gen: 脚本；门禁重算本仓侧并与记录值比对。检查时无网络、无 build、无 TypeScript program。

⛔ B（CI 里拉 objectui 再直接比对）不在范围内，且不是我在 A 与 B 之间做技术选型——B 给 lint job 增加了一条网络依赖并使检查在无凭据时本地不可运行。「新增运行时依赖」与机群形态在本仓属人工地板 ⇒ ⛔ 开发不得自行选 B。若接手人认为 A 不可行，回填说明，不要改选 B。

⭐ 卡已经把最难的部分做完了，接手人不必重做可行性：两半都在 #12977 期间手工跑过，且各带一个证明仪器会响的对照——

• 诊断码集合：两侧各 23 个码，集合相同；对任一侧做一个 token 的扰动，diff 立刻非零。
• 语法区字节哈希：两侧各 214 行，git blob hash 相同；改一个字符即非零。

⚠️ 一个必须带进实现的细节，漏掉会让门禁一上线就假红：UNCONSUMED_WIDGET_OPTION 在本仓是字面量、在 objectui 侧是常量引用。卡说"decomposing that spelling difference is the only subtlety"——这就是那个 subtlety，请在提取器里显式处理，⛔ 不要用"两边都当字面量扫"的实现。

⚠️ 与 #12956 的交互，同样必须带进实现：.objectui-sha 的 diff 今天推导不出任何 pin-critical 门禁。所以这个新门禁要从那条路径显式命名，⛔ 不能靠路径推导捡到它——否则 pin 一升、objectui 侧一动，本该响的正是它，而它不在被推导出的家族里。这条是本卡最容易在实现期丢掉的约束。

⭐ 归类：这张卡属于本轮我持续在数的那个族——「一个机制声称的覆盖面大于它实际有的」。本卡是其中最纯的一种：#12719 白纸黑字写了 invariant，而 enforced by nothing。33 个 objectui check-* 脚本、本仓 181 个 check 家族，都不比对这两份拷贝。

---

Generated by Claude Code