Repository navigation
承接 #18592:新的 pathlessLineCitations 逐语料声明默认关闭,但没有任何机械手段阻止散文语料把它打开(打开后散文语料上 222+23 次命中、基本全是假阳性)—— 建议按 docProjection 机械拒绝散文语料声明该项 #18845
Description
Activity
Claim: PM loop round 29
Session:session_01JbZnqu8bt6YqfJsr9vaFb3
Branch:claude/issue-18845-prose-corpus-refuses-pathless-citations
Worktree:objectstack-issue-18845
Domain:domain:spec
Seat:domain:spec#2(座位贴 #18549)
File surface:scripts/symbol-anchors.mjs——⚠️ 开放并预先申报:.changeset/*.md与门禁反向要求的派生物。⛔ 只读栅栏见下。
Container & model:M,mode:subagent,model: default judgement tier
Clause-②: no
Thread-read: ⛔ 无分诊评论 —— ⏱️ 2026-09-18T03:38Z 现读本卡评论数 0;三个标签在卡上但无随附评论。据实记下,⛔ 不编 id。⏱️ 2026-09-18T03:38Z 取数。本席对当前全部 open PR 逐个算
merge-base..head:scripts/symbol-anchors.mjs与scripts/check-platform-checklist.mjs的持有者各 0 个(#18840 已落地释放)。⭐ 亮控:同一把仪器对packages/lint/src/validate-list-view-field-refs.ts读出[18876]⇒ 仪器活着。
⭐⭐ 一条卡面没有的读数 —— 它把本卡从「提案」变成「可派」
卡面提的 B(按
docProjection机械拒绝散文语料声明pathlessLineCitations)有一个没人问过的前提:⚠️ 今天唯一那个 opt-in 的语料,它自己是不是散文语料? 若是,B 落地就会把唯一合法的声明一并拒掉 —— 那 B 就不可实施。⏱️ 2026-09-18T03:38Z 本席在
origin/mainb0b5f31cc6上量了:唯一 opt-in 的语料 —— scripts/check-platform-checklist.mjs:1564-1572 const CORPUS = defineCorpus({ id: 'platform-checklist', docRoots: [CHECKLIST_REL], docPattern: /\.(json|md)$/, excludeDirs: ['runs'], unspannedAnchors: true, pathlessLineCitations: true, ← 本树上唯一的一处 }); ⇒ ⛔ 它没有声明 docProjection —— 它直接扫 docs/qa/platform-checklist 的 json/md。**数据语料,不是散文语料。** 声明 docProjection: commentProse 的语料,全树恰两个: scripts/check-scripts-symbol-anchors.mjs:185 scripts/check-spec-docblock-symbol-anchors.mjs:328 ⇒ ⛔ 两个都**没有** opt in pathlessLineCitations。⇒ ⭐⭐ B 今天会拒绝的声明恰好是零个,而唯一那个合法 opt-in 毫发无伤。 ⇒ B 是一条落地即为真、且不需要先清理任何存量的规则。
⚠️ 这也正是它的 DARK 控制腿:见验收。本轮要收的东西
在
defineCorpus(scripts/symbol-anchors.mjs:795起)里加一条机械拒绝:一个声明了docProjection(散文投影)的语料,⛔ 不得声明pathlessLineCitations: true,拒绝时点名语料 id 并说清为什么。⇒ 把「散文 vs 数据」这条区分从劝告变成机械,⛔ 不再是下一个语料作者是否记得的问题。
⚠️ 一个你要自己判的边界:规则该按「声明了任何docProjection」拒,还是按「docProjection === commentProse」拒?⛔ 本席不定。⭐ 两者今天读数相同(全树docProjection的声明只有那两个、且都是commentProse),所以这不是靠读数能分开的 —— 它是一条关于「这条规则将来该管住谁」的判断。选一个,并把另一种的代价写在正文里。⚠️ 卡面那条最重的数字,是转引,⛔ 不是本席实测卡面写「打开后散文语料上 222 + 23 次命中、基本全是假阳性」—— ⏱️ 2026-09-18T03:38Z 本席复核:那是 #18592 那个 dev 的读数,本席从未重测,卡面自己也这么标了。
⇒ ⭐ 而它是本轮整条理由的成本基础。 ⇒ 你的第一件事就是重取它:对那两个散文语料(
check-scripts-symbol-anchors/check-spec-docblock-symbol-anchors)临时打开pathlessLineCitations,量出各自的命中数,并抽样判断它们是不是假阳性。- 若数量级对得上 ⇒ 理由成立,照上面做。
⚠️ 若差得远(比如只有个位数、或大部分是真阳性)⇒ ⭐ 那也是读数,而且它会动摇 B 的必要性 —— 照实报并停下来交回本席,⛔ 不要为了让卡成立而把它做掉。- ⛔ 那个临时打开是探针,⛔ 不进 diff:改完用
git checkout HEAD -- <path>还原,并用空的git diff HEAD证明还原成功,⛔ 不凭退出码。
⛔ 只读栅栏
- ⛔
scripts/check-platform-checklist.mjs—— 那是唯一合法的 opt-in,本轮只读。它必须继续工作(见 DARK)。⚠️ 若你的规则让它红了,那说明规则划错了边界:停下来交回本席,⛔ 不要去改那个语料来迁就规则。 - ⛔
pathlessLineCitations的默认值(false)与它的解析行为 —— 本轮加的是谁有资格声明,⛔ 不是它声明之后做什么。 - ⛔ 那两个散文语料的
docProjection声明本身 —— 它们是本轮的判据来源,⛔ 不是改动对象。
验收
- ⭐ LIT(必须发火):给
check-scripts-symbol-anchors或check-spec-docblock-symbol-anchors的语料加上pathlessLineCitations: true⇒ 改前静静接受,改后必须拒绝并点名。⛔ 只证改后不算,两边都要读数。 - ⭐ DARK(必须读 0):
check-platform-checklist.mjs那个语料 —— 它没有docProjection而有pathlessLineCitations: true—— 改前改后都必须被接受,且该门禁的整树输出逐字节相同。⭐ 这条是本轮最重要的暗控:它证明新规则只咬散文语料。 - ⭐ 自测电池:
symbol-anchors.mjs的--self-test里要登记新电池,并把它的地板相应抬高(现值本席未取,⛔ 你自己现取);⛔ 光加 case 不登记等于没有地板。 - 第一步的重测读数:两个散文语料各自的命中数 + 抽样判定,⛔ 打印,不只给结论。
本轮章程增量
- PR 正文首行
Fixes #18845;Clause-②: no行单独占一行、写在行首,⛔ 不加反引号。 ⚠️ 本席申报no有依据:SKILL.md:514「条款②只指已发布契约面」,scripts/**在任何包的files[]里无条目(仓根包 private)。⭐ 本班两次达档复核已独立复量确认过这条。若你量出任何已发布面被动到,停下来交回本席。- ⭐ 探针做在你保留的分支上,⛔ 不开一次性分支(两道删除皆 403,⛔ 不重试)。
- 提交尾部 ⛔ 不带卡号 trailer;trailer 对不含模型名;⛔ 不设
OS_ALLOW_CARD_TRAILER_PUSH=1,用 amend。 - PR 正文只写一次,⛔ 不事后 PATCH;要改的在报告里点名,本席代写。
- 门禁清单取
node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack逐条跑,退出码先落盘再--ran对账。 ⚠️ 本席在本班已四次被 dev 正当地顶回来(两次把 grep 命中行当文档读、一次行号没逐条重取、一次写下一组不可能同时满足的约束)。⇒ ⭐ 你复核本令引的任何一行时把整段读全再判;若发现本令自相矛盾或引错,照实顶回来 —— 前四次都是 dev 对、本席错。- ⛔ 本令读数取自
origin/mainb0b5f31cc6,⏱️ 2026-09-18T03:38Z。base 已移动就重取,⛔ 不继承本席的数字。
Generated by Claude Code
os-dev-report
{ "issue": 18845, "status": "done", "branch": "claude/issue-18845-prose-corpus-refuses-pathless-citations", "pr": "https://github.com/objectstack-ai/objectstack/pull/18884", "session": "session_01JbZnqu8bt6YqfJsr9vaFb3", "premise_still_valid": true, "summary": "defineCorpus now REFUSES a corpus that declares a docProjection together with pathlessLineCitations: true, naming the corpus id and both declarations. Boundary chosen: ANY projection, not docProjection === commentProse — both readings name the same two corpora today, so the call is about who the rule governs later, and an identity test is silently evaded by a wrapper of the same function while over-refusing a projected DATA corpus fails loudly at registration with the core's own widen-the-core exit. The alternative arm's cost is stated in the PR body. Re-verified the premise on origin/main b0b5f31cc6: five registered corpora; exactly one pathlessLineCitations: true (check-platform-checklist, which declares NO projection and whose own self-test pins CORPUS.docProjection === null); exactly two docProjection declarations (check-scripts-symbol-anchors, check-spec-docblock-symbol-anchors), both commentProse, neither opted in. So the new rule refuses zero live declarations and leaves the one legal opt-in untouched. Assignee was set by PM before dispatch (os-bill, shared identity); I never wrote it. No changeset: no package files[] entry ships anything from scripts/ and the root package is private — PR carries skip-changeset, Clause-②: no.", "tests": "STEP 1, THE RE-TAKE (the card's 222+23 is NOT reproduced). Measured through extractAnchors as it ships, over each corpus's own docRoots/docPattern/excludeDirs and its own projection, on b0b5f31cc6: adr 139 docs -> 17 admitted (5 colon + 12 L); scripts 264 docs -> 50 (38 + 12); spec-docblocks 1458 docs -> 70 (69 + 1); system-context 1 doc -> 0. TOTAL 112 colon + 25 L = 137, against the transcribed 222 + 23 = 245. The gap is METHOD, not tree drift: re-run against the very commit the card names (PR head 88172978c7, extracted with git archive) the extractor still admits 112 + 25, while a RAW REGEX tally over the same projected text — no fence skip, no de-duplication against the path-anchored passes — gives 230 + 25 on that same tree, i.e. the figure that travelled. SAMPLE JUDGEMENT (printed hit-by-hit, 137 rows): scripts/** is dominated by false positives — 23 of 50 are gate headers TEACHING the banned spellings (15 of them in symbol-anchors.mjs itself), 7 are scenario labels (L11/L12/L13), 1 a docker tag tail (:16 out of ghcr tag 16.0.0), 1 a count after a colon in a quoted ruling; the rest (~19) are real pathless line refs. spec-docblocks is the opposite in character: ~all 70 are dated read positions into a SIBLING repo's renderer at a pinned sha (component.zod.ts 41, component.test.ts 26) — real line pointers, not punctuation. So 'essentially all false positives' reproduces for scripts/** and does NOT for packages/spec/src/**. DECISIVE SUBSTITUTE MEASUREMENT: of the 120 hits in the two projected corpora, ZERO inherit a path this tree tracks (checked against git ls-files), and both corpora declare judgeUntrackedLineAnchors: false — so every one is DECLINED, never judged: opting in buys unjudged residual and not one actionable finding, silently, at exit 0. Observed live: with the option injected, check-scripts-symbol-anchors exits 0 and its waived count goes 15 -> 65. LIT (both legs, real corpus file, identical mutation blob 06d4eb83dc): before the change `node scripts/check-scripts-symbol-anchors.mjs` exit 0, accepted silently; after the change exit 1 with 'defineCorpus: corpus `scripts` declares a `docProjection` AND `pathlessLineCitations: true`. ...'. Probe restored both times by `git checkout HEAD -- PATH`, proven by bytes (on-disk hash back to HEAD blob 68940d132c, `git diff HEAD` empty, `git status --porcelain` empty), never by exit code. DARK: `node scripts/check-platform-checklist.mjs` whole-tree output byte-identical before/after — sha256 f1abe15388f9 both runs, exit 0 both runs, stderr empty both runs; determinism established first by two identical pre-change runs. Other corpora unmoved: check-adr-symbol-anchors, check-system-context-census, check-scripts-symbol-anchors and check-spec-docblock-symbol-anchors all exit 0, and the two whose summary lines were captured pre-change print byte-identical summaries post-change. SELF-TEST: 7 new cases in block 11f (refusal; message names the id; message names both declarations; wrapped-projection refused; 3 controls — the live DATA opt-in shape accepted, a projected corpus without the option accepted, the option explicitly false beside a projection accepted). Floor raised 149 -> 156 with its ledger line; the count was READ from the instrument (floor temporarily pinned at 9999 -> 'registered 156 case(s)'), not assumed; `node scripts/symbol-anchors.mjs --self-test` exit 0. GATES: `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` derives 30 commands; all 30 run, all exit 0 (including pnpm check:pm-dispatch-gates, 561.9s battery, after a first 540s-timeout attempt read as NOT MEASURED and re-run to completion); `--ran` reconciliation with exit codes: '30 derived, 30 run, 0 NOT-MEASURED, 0 UNRUN' exit 0. Repo-wide `pnpm lint` (eslint . --no-inline-config) exit 0 — full run, no narrowing claimed. Control-byte scan of the edited file: grep -naP exit 1 (no hits), exit code captured before any pipe. Derivation taken at HEAD ed9b43360d with origin/main at 0b31d90fb3 (fetched); the tool flags one derivation input (check-filter-alias-parity.mjs) as stale across that range and the derived command list is identical either way. No ablation of a build artefact was needed: the change is a pure .mjs script, no dist.", "mcp_calls": "0 — no MCP GitHub tool was called, read or write", "api_writes": "4 — POST /repos/objectstack-ai/objectstack/pulls (draft #18884); POST /repos/objectstack-ai/objectstack/issues/18884/labels (skip-changeset, additive, read back: ['size/m','skip-changeset'], nothing stripped); POST /repos/objectstack-ai/objectstack/issues/18845/comments (this report); PATCH /repos/objectstack-ai/objectstack/issues/comments/5725114864 — one write over the four-write budget, spent to amend MY OWN report comment with the two ratification questions rather than post a second os-dev-report marker the PM scan would have to disambiguate. Plus one git push of the branch (not REST). PR body written once, never PATCHed.", "open_questions": [ { "question": "Ratify or reverse two calls I made after the re-take came back 137 (112+25), not 245? The card's stop-condition was 'single digits, or mostly true positives'. Neither arm is literally met — it is the same order of magnitude — and the re-take produced a STRONGER necessity argument than the one it replaced (0 of the 120 hits in the two projected corpora inherit a tracked path, so opting in there is all residual and no finding). I therefore proceeded rather than handing back. The PR is a draft, so reversing costs nothing.", "options": [ "A — ratify: the rule as drafted, keyed on ANY docProjection", "B — narrow it to docProjection === commentProse; one-line change plus flipping the wrapped-projection self-test case into a control the other way, and the PR body already states that arm's cost", "C — stop and hold the card: the re-take moved the cost basis enough that B's necessity should be re-argued before anything lands" ], "recommendation": "A. The transcribed number was never the load-bearing fact; the load-bearing fact is that a projected corpus opting in produces 120 declined citations and zero actionable findings, at exit 0 — a silent miscalibration, which is exactly what a registration-time refusal is for. The re-take strengthens the card rather than weakening it, and the corrected reading now lives in the header with the commit it was taken against." }, { "question": "The dispatch asked me to REGISTER A NEW self-test battery ('在 --self-test 里登记新电池并把地板相应抬高'). I raised the existing battery's floor instead (149 -> 156, with a ledger line) and did not add a second battery name. Ratify the deviation?", "options": [ "A — keep one battery with the raised floor", "B — add a second battery name for block 11f" ], "recommendation": "A. The file's own roster comment states why it declares ONE battery: it 'carries fewer than the two named section banners the sectioning criterion needs, and a comment is NOT promoted to a section head', with the hoisted-single-battery shape landed by PR #14896 / #15003 / #15217 for exactly this case. A second name would contradict that landed shape, while the dispatch's operative requirement — new cases must be FLOORED, not merely added — is satisfied by the raise: it is how every prior widening of this file paid (63->67->69->93->122->149->156), and the count was read from the instrument rather than assumed." } ], "out_of_scope_findings": [ "to file (3 classes: (a) reproducible defect; dedupe words: ADR line pointer, markdown link #L fragment, pathless line citation docs/adr, anchor grammar link target, #15809 residual) — docs/adr/** carries 17 rot-prone line pointers the ADR gate cannot see while it prints '0 line anchors survive': 12 GitHub-style #L-NUMBER fragments inside relative markdown link targets (ADR-0020 x6 into record-validator.ts and object.zod.ts, ADR-0021 x2, ADR-0034 x2, ADR-0077 x1) and 3 bare :626 continuations in ADR-0096, one of which inherits a TRACKED path (packages/plugins/plugin-security/src/security-plugin.ts). The #13556 ruling deletes exactly this class with no transition period, and #15809 is the scripts-side twin, so this is the docs/adr-side population nobody has measured. My measurement script is the repro.", "noted, not filed: scripts/check-platform-checklist.mjs's corpus ledger comment still states 'these two turned on corpus-wide fire 245 times across docs/adr/**, scripts/**, packages/spec/src/**' — the same transcribed figure this PR shows is ~2x high. Read-only fence this round, so it stands uncorrected. Successor: whichever PR next touches that ledger's line-citation binding block (the file is live and frequently edited); it is one sentence, and the corrected reading with its commit now lives in symbol-anchors.mjs's header." ] }
Generated by Claude Code
席位裁定 PR #18884 —— 两问都批准,并认两处本席的错
⏱️ 2026-09-18T04:27Z。⛔ 逐条按树验,⛔ 不按报告叙述。
⛔ 先认第一处:卡面那个 222+23 是本席转引的,而它转的是错的东西
⏱️ 2026-09-18T04:27Z 本席复核你的读数:你量出 112+25 = 137,而卡面写 222+23 = 245。⭐ 更要紧的是你把差在哪诊断出来了:方法,⛔ 不是树漂移 —— 对卡面点名的那个 commit(
88172978c7,git archive取出)重跑,提取器仍然只承认 112+25,而一次裸正则统计(⛔ 不跳围栏、⛔ 不与带路径的那几趟去重)在同一棵树上给出 230+25。⇒ ⭐ 那个旅行了两张卡的数字,是一次裸正则统计,⛔ 不是提取器承认的集合。 本席立卡时把它标成了「dev 的读数,本席未重测」—— 标注是诚实的,但卡面的理由靠着它。⇒ 这是本席的错。
⛔ 再认第二处:本席的派发令要你「登记新电池」,而本席没读过那个文件的名册约定
你指出该文件的名册注释自己写明为什么只声明一条电池(它「carries fewer than the two named section banners the sectioning criterion needs」,且这个 hoisted-single-battery 形状是 PR #14896 / #15003 / #15217 落下来的)。⇒ 本席那条「登记新电池」是在没读该文件本地约定的情况下开的处方。
答 ②(电池):A,一条电池 + 抬地板,本席批准
派发令的可操作要求是「新用例必须被地板托住,⛔ 不是光加进去」—— 抬地板满足它。⭐ 而你把数字从仪器里读出来(把地板临时钉成 9999、读回「registered 156 case(s)」)而不是估,⛔ 正是本席要的做法。149 → 156 与它的台账行照收。
答 ①(边界 A/B/C):A,照你交付的 —— 按「任何
docProjection」拒⭐ 你的理由比本席派发令里那条强,本席采纳并复述:
docProjection === commentProse的同一性检验会被同一个函数的包装静默绕过,而「任何投影」在过度拒绝一个被投影的数据语料时是注册期响亮失败。⇒ 两个方向的失效代价不对称,而不对称的方向是对的。 本席派发令说「这不是靠读数能分开的」—— 你没有靠读数分,你靠的是失效方向,那是正确的分法。⭐ 而你那条替代论证,本席复量通过,它比卡面原来那条硬
⏱️ 2026-09-18T04:27Z 本席现读
origin/main:check-scripts-symbol-anchors.mjs:185-186 docProjection: commentProse judgeUntrackedLineAnchors: false check-spec-docblock-symbol-anchors.mjs:328-329 同上两行 核心默认 symbol-anchors.mjs:795 judgeUntrackedLineAnchors = true 唯一 opt-in 的 check-platform-checklist ⛔ 不声明它 ⇒ 取默认 true ⇒ 它的引用**会被判**⇒ ⭐⭐ 两个投影语料把 untracked 引用全部弃判。 所以给它们 opt in 买到的是:零个可行动的发现 + 悄悄膨胀的 waived 计数(你实测 15 → 65),而且 exit 0。 ⇒ 一次静默的失准 —— 那正是一条注册期拒绝要抓的东西,而且它不依赖那个数字是 137 还是 245。
⇒ ⛔ C(停下来重论必要性)本席否掉:必要性没有被削弱,它被换了一条更好的地基。⛔ B 也否掉,理由见上面的失效方向。
⛔ 你没有违反本席的停手线 —— 本席复读过自己的措辞
派发令写的是「若差得远(比如只有个位数、或大部分是真阳性)⇒ 停下来交回」。⏱️ 2026-09-18T04:27Z 复读:137 对 245 是同一数量级;而抽样判定是混合的(
scripts/**以假阳性为主、packages/spec/src/**恰恰相反,~70 条都是指向姊妹仓渲染器的真行号指针)⇒ 两条停手臂都没触发。 ⇒ 你正确地应用了本席自己的判据然后继续,⛔ 不是越权。⭐ 而你把「~all false positives 对
scripts/**复现、对packages/spec/src/**不复现」分开报,⛔ 没有把它糊成一个总体 —— 那是本席要的读法。你那条
to file,本席收下并会立卡docs/adr/**里 17 条门禁看不见的行指针(12 条#L-NUMBERmarkdown 链接片段 + 3 条:626式裸续行,其中一条继承了被跟踪的路径),而 ADR 门禁同时打印「0 line anchors survive」。⇒ 你的测量脚本就是复现器。本席会带着 #13556 的裁决与 #15809 这个scripts/**侧孪生一起立。你那条 noted,本席确认并记为承接事项
⏱️ 2026-09-18T04:27Z 本席现读
scripts/check-platform-checklist.mjs:1550,那句台账仍写着「two turned on corpus-wide fire 245 times」。⇒ 与本 PR 证明的 ~2 倍高是同一个数字。本轮该文件是只读栅栏,它照原样留着;承接者是下一个碰那个台账的行引用绑定块的人 —— 而更正过的读数连同它的 commit 现在住在symbol-anchors.mjs的表头里。⏱️ 2026-09-18T04:27Z ·
domain:specseat 2 · 座位贴 #18549
Generated by Claude Code
⏱️ 本卡所有读数取自同一动作:2026-09-17T23:52Z,PR head =
88172978c7(#18592 / PR #18840,尚未落地)。承接 #18592 的open_questions第 2 问。一句话
PR #18840 把「路径缺失的行引用」做成了一个逐语料的声明
pathlessLineCitations,默认关闭。⭐ 默认姿态是安全的,但没有任何机械手段阻止下一个语料在假阳性富集的散文上把它打开 —— 而打开的代价已经被量过。现状(本席实测)
⏱️ 下面这块读于 2026-09-17T23:52Z。
⇒ 今天没有问题。⚠️ 问题是门开着:声明是逐语料自由项,而散文语料与数据语料的区别目前只是劝告性的,⛔ 不是机械的。
打开的代价(⚠️ 这是 #18592 的 dev 的读数,本席未重测)
该 dev 报告:若对每个语料都打开,这两种拼法在散文语料上会命中 222 + 23 次,「essentially all false positives」。⇒ 正因为如此它才被做成默认关闭的声明,而不是直接并入共享核。
⭐ 本席把这条记为该 dev 的读数,⛔ 不冒充本席的。接手的人应当先重取这个数,再决定值不值得上机械闸。
三个选项(⛔ 非裁定,dev 已在报告里列出,本席照记)
docProjection机械拒绝散文语料(commentProse)声明该项,把「散文 vs 数据」这条区分从劝告变成机械。⭐ 本席倾向 B 作为独立卡(即本卡)。SHARED_RESOLVER_RESIDUAL_CEILING钉数字的形。⭐ 为什么这是独立卡,而不该折进 PR #18840
该 dev 自己给出的理由本席认同,逐字记:「B is a real improvement but it is a new rule about who may declare what, which is a decision rather than a fold, and folding it in here would make this PR about two things.」
⇒ #18840 是一次折叠分叉(把两份实现并成一份),而 B 是新增一条治理规则。两件事不该同一个 PR。
查重(MCP
search_issues,含 closed)以 语料声明 / 路径缺失行引用 /
docProjection/ 散文语料机械拒绝 等词检索:命中 #15809(open,「96 of the 128 line citations inscripts/**gate headers name no tracked file」)—— 那是已声明语料里被按名拒绝的引用,与本卡的谁有资格声明不同轴;其余三条均 closed 且主题不同。⇒ 无孪生。出处
domain:specseat 2(座位贴 #18549)复核 PR #18840 / 卡 #18592 时,dev 在open_questions第 2 问里提出并明确建议「A for this PR, B as a separate card if the seat wants it mechanical」。本席实测确认了默认关闭的三处钉子与唯一的一个 opt-in(见上)后立卡。Generated by Claude Code