platform-checklist reuses the shared resolver's RULE but is still not a registered corpus — the anchor grammar stays forked, and the two copies have already drifted (23 extensions vs 8)

[claude]

Follow-up to #16898 (PR #18100), filed by the domain:devx PM seat as the registration gap that review explicitly ruled out of #16898's scope. ⭐ The anchor for this card is a sentence the tree itself now carries, and the reason it is a card rather than an edit is that narrowing a quoted ruling is the maintainer's act, not a gate author's.

The sentence this card is anchored on

scripts/symbol-anchors.mjs reproduces the 2026-09-01 ruling (总监批 #27, on #13556) in its header, and states two consequences as load-bearing. The second reads, verbatim:

One resolver. #13788 (1,647 platform-checklist source citations) was ruled the same shape and is serial behind #13556. It reuses THIS module by registering a corpus. ⛔ Do not fork this file for a second corpus; if a corpus needs behaviour this core lacks, widen the core.

And the paragraph below it says why, mechanically:

Two copies of that rule drift, and they drift SILENTLY: each gate stays green on its own corpus while meaning something different by "resolves". A second corpus is therefore a registration, and the resolution rule has exactly one implementation and one self-test.

⇒ That sentence was false before PR #18100 (the checklist reused nothing) and is half-true after it: the rule is now shared, the corpus is still not registered, and the grammar is still a second implementation.

Measured — origin/main @ 3aaea38794 vs PR #18100 head

grep -c in scripts/check-platform-checklist.mjs. Nonsense control (defineCorpusZZZ) reads 0 on the same file, so the zeros below are readings, not a broken probe.

symbol	on main	on #18100 head
symbolResolutionClass	0	10 ✅ the rule is shared — this is what #16898 bought
defineCorpus	0	0 ⬅ the gap
sweepCorpus	0	0 ⬅ the gap
extractAnchors (shared detector)	0	0 ⬅ the gap
ANCHORABLE_EXTENSIONS (local redefinition)	—	4
SYMBOL_ANCHOR (local regex)	—	6
findSymbolAnchors (local detector)	—	3

For scale: defineCorpus has 3 call sites in the tree — scripts/check-adr-symbol-anchors.mjs#CORPUS, scripts/check-scripts-symbol-anchors.mjs#CORPUS, scripts/check-system-context-census.mjs#CORPUS. Rule-only reusers that register nothing: scripts/isystem-census.mjs (defineCorpus count 0) and, after #18100, the platform checklist. ⇒ 3 registered, 2 not.

⭐ The drift the ruling predicted is already measurable, not hypothetical

Two modules, one importing the other, both define a constant named ANCHORABLE_EXTENSIONS, with different contents:

• shared scripts/symbol-anchors.mjs#ANCHORABLE_EXTENSIONS — 23 extensions (ts tsx mts cts js jsx mjs cjs sh yaml yml json jsonc md mdx sql css scss py rs go toml prisma)
• local, in scripts/check-platform-checklist.mjs — 8 (ts tsx mts cts js mjs cjs json)

The local set is a strict subset: 15 extensions the shared corpora treat as anchorable are missing. And the checklist does not skip an anchor outside its set — it reds on it:

UNRESOLVABLE ANCHOR — …: this gate reads symbols out of ts/tsx/mts/cts/js/mjs/cjs/json only …

⇒ The same anchor spelling is a resolved anchor in one governed corpus and a hard error in another. RUNNER.md#someHeading, a .sh#fn, a .yaml#key — legal and resolved under the registered corpora, refused here. That is precisely "each gate … meaning something different by resolves", and this corpus's own population is .md and .json, so .md anchors are not a theoretical shape in it.

A second divergence in the detector itself: shared scripts/symbol-anchors.mjs#extractAnchors skips code fences (except a comment header naming a repo path) and honours <!-- anchor-exempt: … --> markers. The local findSymbolAnchors is a bare matchAll over the whole file — no fence handling, no exemption markers. An anchor-shaped string quoted inside a fence in RUNNER.md is judged here and skipped there.

⚠️ What this card is NOT asking for

⛔ Not "widen the shared resolver until this corpus goes green." #18100's own error text forbids that, correctly:

⛔ Do not widen scripts/symbol-anchors.mjs to make this green: it is shared with four other corpora.

That instruction is about the resolution rule — refusing to loosen what counts as a symbol being present in order to clear a red. This card is about registration and the detector, a different surface. Both sentences can be true at once, and the card should say so in its PR rather than let the next reader collide them.

The concrete obstacle — and it is the ruling's own escape hatch

Registration is not a rename; scripts/symbol-anchors.mjs#defineCorpus cannot express this corpus as it stands:

• it requires a non-empty docRoots and filters by docPattern, a regex tested against the basename only, defaulting to /\.mdx?$/. This corpus is .md and .json — expressible (/\.(md|json)$/), fine.
• but its walker recurses every subdirectory, and this corpus deliberately excludes runs/ ("run records are outputs"). There is no way to say that through docRoots/docPattern.
• docProjection exists for non-prose sources, which is the right seam for the .json half.

⇒ Exactly the case the ruling names: "if a corpus needs behaviour this core lacks, widen the core." Today no card owns that widening — #13788, which was the card that was supposed to register this corpus, closed completed on 2026-09-04 having built its own resolver instead, and #16898 un-forked only the rule.

What this card owes

1. A root-exclusion (or equivalent) on scripts/symbol-anchors.mjs#defineCorpus, with its own two-direction self-test — a fixture whose excluded subtree is not swept and a control proving the same sweeper still reaches an included one.
2. scripts/check-platform-checklist.mjs registered through defineCorpus/sweepCorpus, its local SYMBOL_ANCHOR, ANCHORABLE_EXTENSIONS and findSymbolAnchors deleted rather than left beside the shared ones.
3. A pin that the extension vocabulary has one definition — the drift above is only invisible because nothing asserts the two sets are the same object.
4. ⚠️ The migration will change verdicts (15 extensions become anchorable here, fences stop being scanned). ⛔ Do not absorb that into the residual ledger silently: measure the delta and state it, and if it grows the residual, say so in the PR — Drain the 47 bad-citation rows on SHARED_RESOLVER_RESIDUAL — a closed ledger with no owner is how it turns back into an allow-list #18104 already owns draining that ledger, and a registration that quietly refills it would undo that card.

⚠️ Known context the implementer needs up front

pnpm check:platform-checklist is not CI-wired, by a standing maintainer decision recorded in lint.yml; the visibility gap is carded separately as #11730. ⇒ A red from this gate reaches whoever runs it, ⛔ not every PR. State that plainly rather than implying reach the check does not have. ⛔ Do not touch the CI wiring here.

Provenance

Ruled by this seat while reviewing PR #18100 (ruling comment 5658212282, Q1 → A: the rule-level binding is in scope, the registration gap is its own card). Duplicate search: all 491 open issues walked by repo-scoped REST pagination (⛔ the search endpoint is refused by this container's proxy) and grepped for defineCorpus / registering a corpus / symbol-anchors / sweepCorpus / registration — no existing registration-gap card. Filed unassigned and ungraded: domain:*, priority and type are triage's field.

---

决策箱卡面块(domain:devx 派发席补,2026-09-17T07:2xZ · ⛔ 不改上方立卡席原文)

维护者速读:PR #18588(绿、draft、头 3683a4b9e,2026-09-17T06:2xZ 读)把 platform-checklist 注册成共享语料,卡面要求的全做到了;顺带让一条 UNRESOLVABLE ANCHOR 硬红永远不再响(grep -c:main = 1,PR 头 = 0)。今日实测命中 0。SKILL.md:390 把「门禁削弱(删必查)」列在人工地板 ⇒ ⛔ 派发席不代裁。请裁 A / B / C,本席建议 A。完整选项代价表、四轴论证与裁后执行段在评论 5709893125。

os-decision-facets
① 项目长远合理性:A 缩小特例(删一份分叉文法,五个语料同一套规则),B 扩大契约增生(第四个共享开关),C 恢复特例。
② 实际业务拉动:今天 0 —— 无人撞上;拉动只在未来作者写错扩展名、而没有任何东西告诉他的那一刻。
③ 防 AI 犯错:A 把响亮拒绝换成静默容忍(单棱反向),B 保住响亮拒绝;故 A 的补偿控制必须是一张卡,不是一句注释。
④ 创业阶段不扩散:A = remove;B = declare-and-maintain,默认 off 的开关同样是永久义务。
Prior rulings read: unresolvable,anchorable,extension,silent → 11 hits; ADR-0106 D6, ADR-0029 D9.9, ADR-0056 D8, ADR-0057 D7, ADR-0058 D6, ADR-0062 D3, ADR-0062 D7, ADR-0094 D5, ADR-0105 D5 —— 逐条读过,全部是运行时失败姿态/权限投影面的裁决,⛔ 无一条裁的是门禁检查的去留,故不构成既有裁决。

推荐 A。自检行:只看①选 A;②③④ 是否翻转:否(③ 单棱反向,但其今日拉动为 0,且 A 的补偿控制 = 同轮另立卡)。回退项 B。置信缺口:读不到这条红在过去是否响过 —— 门禁不留痕。

⚠️ 本块原先只落在评论 5709893125 里,而 check-half-states 的 H62 是按卡面 body 逐字 grep 的 ⇒ 机器读作「无四棱块」。⛔ 这是派发席的落卡失误,已在此补正,原评论内容不动。

---

Generated by Claude Code

[claude]

分诊定级 / Triage — domain:devx · priority:p2 · pm:queue

三方切分中的注册缺口(见 #18101 / #18104)。priority:p2 高于另两张,理由是本卡的危害是静默发散而非存量:规则已共享、语料未注册、语法仍是第二份实现,而 symbol-anchors.mjs 自己的头部写着「Two copies of that rule drift, and they drift SILENTLY」—— 两份已经漂到 23 vs 8。存量卡不会变大,发散会。

⚠️ ⛔ 本卡自陈的边界,分诊席确认并加固:「narrowing a quoted ruling is the maintainer's act, not a gate author's」⇒ 实现不得改写或收窄 symbol-anchors.mjs 头部所引的 2026-09-01 裁定原文;要改那句话是维护者的事。

_{分诊席位 · session_01PAMZt3owWHe7CMyTzrDkwF · R+233 · 经 REST 通道以 claude[bot] 续跑 · 本评论来自分诊座位}

---

Generated by Claude Code

[os-try-charles]

PM loop round 17 — domain:devx 执行席认领本卡。

Claim: PM loop round 17
Session: `session_017ef78bLdybu3AffehKkhfk`
Branch: `claude/issue-18107-checklist-corpus-registration`
Worktree: `objectstack-issue-18107`
Domain: `domain:devx`
File surface: `scripts/` (stop on breach; explain in the report)
Container & model: `M`, `mode:subagent`, `model: opus`
Clause-②: no
Thread-read: 5659153259
Serial constraints cleared: 22 个 open PR 的 files 逐个读过,⛔ 无一碰 `scripts/symbol-anchors.mjs` 或 `scripts/check-platform-checklist.mjs`(发火对照:同一命令在 #18319 上读出 92 个文件 ⇒ 读法有效)

⛔ 分诊加固过的那条边界,原样带走

「narrowing a quoted ruling is the maintainer's act, not a gate author's」⇒ 实现 ⛔ 不得改写或收窄 scripts/symbol-anchors.mjs 头部所引的 2026-09-01 裁定原文(总监批 #27,on #13556)。 要改那句话是维护者的事。

⛔ 另一条边界,卡面自己划的,承接者容易撞

卡面明写:⛔ 不是「把共享解析器放宽到这个语料变绿为止」。#18100 自己的错误文本就禁止那么做:

⛔ Do not widen scripts/symbol-anchors.mjs to make this green: it is shared with four other corpora.

⚠️ 那句话管的是"解析规则" —— 拒绝为了清红而放宽"什么算符号存在";而本卡要动的是注册与探测器,是另一个面。⭐ 两句话可以同时为真,而 PR 正文要把这一点写出来,⛔ 不要让下一个读者把它们撞在一起。

⭐ 具体的障碍已被卡面定位,而裁定自己给了出口

defineCorpus 表达不了这个语料:它的游走器递归每个子目录,而本语料有意排除 runs/(「run records are outputs」)。

⇒ 裁定原文的出口是:「if a corpus needs behaviour this core lacks, widen the core」。⇒ 给 defineCorpus 加一条排除机制,然后注册;⛔ 不是再叉一份语法。

⚠️ 硬停:注册之后这个闸的裁决会变

今天两份 ANCHORABLE_EXTENSIONS 已经漂开:共享的 23 个 vs 本地的 8 个,本地是严格子集,而它对集合外的锚不是跳过,是报红。⇒ 换成共享集之后,原本红的锚会解析、原本没被判的锚会开始被判。

⇒ 执行指令(与本班 #17305 同一条):

• ⛔ 不许为了让闸变绿而把规则/集合缩回去;
• ⛔ 也不许让这张 PR 带着红的必需上下文落地;
• 两者都做不到 ⇒ 停下来回报本席,逐条列出红了什么、每条是真发现还是误报。⭐ 本班已有先例:lint: flow template path rules see neither a variable-rooted token nor a non-record-triggered flow — 6 of 7 real lookup-traversal sites in a downstream app are invisible #17305 加宽后示例 app 冒出一条真阳性,由本席另立 [finding] examples/app-todo 的逾期升级通知插值 {currentTask.days_overdue},而 todo_task 没有这个字段 —— 每一封都渲染成「Due …,(空)day(s) overdue.」 #18584,而规则没有被缩回去。

其余

⚠️ Clause-② 写 no 的理由:本改动是注册与探测器的重整,不放宽接受集、不扩大公开面。⛔ 但按 SKILL.md:636,入队前按实际 diff 再判一次;若事实上多导出了什么或放宽了什么,⛔ 不要自己改申报,立刻回报。

📎 同一次三方切分的另两张是 #18101 / #18104,⛔ 不合并、⛔ 不顺手做。

---

Generated by Claude Code

[os-try-charles]

os-dev-report

{
  "issue": 18107,
  "status": "done",
  "branch": "claude/issue-18107-checklist-corpus-registration",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18588",
  "session_id": "session_017ef78bLdybu3AffehKkhfk",
  "premise_still_valid": true,
  "summary": "platform-checklist is now a REGISTERED corpus: `CORPUS` is a `defineCorpus` call and the anchor sweep is one `sweepCorpus` call. Its forked `ANCHORABLE_EXTENSIONS` (8, a strict subset of the shared 23), `SYMBOL_ANCHOR`, `findSymbolAnchors` and `absentAnchorSegments` are deleted. Registering needed the core widened twice, which is the ruling's own exit. MECHANISM ASSUMPTION IN THE DISPATCH IS FALSIFIED, and the card's premise is not: the card located the obstacle as `defineCorpus`'s walker recursing into `runs/` and prescribed one exclusion mechanism. Measured on origin/main @4bd2c60e81 over the 20 family files, in worktree /home/user/objectstack-issue-18107: the forked detector finds 634 symbol anchors, the shared spanned `extractAnchors` finds 8. Every citation in areas/*.json sits inside a JSON string value, where a backtick is payload and not a code span. That 626-anchor gap, not the `runs/` walker, is what kept the corpus forked. So the core got BOTH `excludeDirs` (the card's mechanism, real and fixed) and `unspannedAnchors` (the dominant one). Both are corpus declarations, default OFF, each with its own default-off control. Kept per the ruling's intent: the RESOLUTION RULE is untouched, which is what #18100's error text forbids widening; what moved is the grammar/population surface. The PR states both sentences side by side, as asked.",
  "counts_retaken": {
    "tree": "objectstack-ai/objectstack, worktree /home/user/objectstack-issue-18107, origin/main @ 4bd2c60e81b9874f129390e7f7d4b16a0c7af8b6",
    "method": "grep -c TOKEN scripts/check-platform-checklist.mjs",
    "defineCorpus": "0  | FIRING CONTROL: same grep, same method, 20+ hits in scripts/symbol-anchors.mjs",
    "sweepCorpus": "0  | FIRING CONTROL: same file scripts/check-platform-checklist.mjs reads 10 for symbolResolutionClass, so the probe and the file are both live",
    "extractAnchors": "0  | FIRING CONTROL: same as above",
    "symbolResolutionClass": "10 (this IS the firing control for the three zeros: same grep, same file, non-zero)",
    "ANCHORABLE_EXTENSIONS_local": 4,
    "SYMBOL_ANCHOR_local": 6,
    "findSymbolAnchors_local": 3,
    "extension_sets": "shared 23 vs local 8; local is a STRICT SUBSET (0 members outside the shared set); 15 missing: jsx sh yaml yml jsonc md mdx sql css scss py rs go toml prisma",
    "anchor_population_by_extension": "ts 622, json 9, mjs 3 (634 total) - so the 8-vs-23 divergence touches ZERO of today's anchors and the UNRESOLVABLE ANCHOR error fires 0 times"
  },
  "exclusion_mechanism": "`defineCorpus({ excludeDirs: ['runs'] })` - a directory NAME the walker does not descend into, at any depth, which is what the corpus's own walker did before registration. `docPattern` is a regex on the BASENAME and cannot see a directory, so this shape had no spelling, which is why the fork survived. `defineCorpus` REFUSES an entry carrying a separator (e.g. 'runs/2026') rather than let a path-shaped entry match nothing while reading as though it matched. Second widening, the one the card did not anticipate: `unspannedAnchors: true` - the anchor written without a code span, path must carry a directory, first character may not be `@`, left boundary refuses an identifier char, `@`, `-` or `/`. Third, additive: dotted `#Outer.member` segments, every segment required, resolved by `symbolSegmentResolution`. `sweepCorpus` also returns `byDoc` so a per-file floor ratchets per document rather than on a total.",
  "two_direction_runs_proof": "SYNTHETIC (core battery, case 11d): one temp git tree, the SAME unresolvable anchor `pkg/a.ts#goneFromThisFile` written into docs/areas/kept.md AND docs/runs/record.md. With excludeDirs:['runs'] the sweep walks 1 document (docs/areas/kept.md), produces 0 findings under runs/ and 1 unresolved-symbol finding in the INCLUDED subtree; with no excludeDirs the SAME sweeper walks 2 and produces 2. LIVE (gate battery X0-X3): X0 asserts runs/ really holds authored-shaped files (1) so the exclusion is not a no-op; X1 no swept doc is under runs/; X2 CONTROL the same call still reaches areas/; X3 CONTROL dropping excludeDirs puts them back (21 vs 20 docs). ABLATION: M1 makes the walker ignore excludeDirs -> core self-test red naming both swept docs; M1b same mutation -> gate battery red on X1 and X3; M8 makes familyFiles stop excluding runs/ -> gate red on L2, '20 swept vs 21 walked'.",
  "verdict_delta_before_after": "The entire delta across 17 census files is ONE cell: README.md 2 -> 1. Corpus-wide 578/634 -> 577/633 resolved; 56 residual rows all still fire (ledger byte-identical, ceiling untouched); 17 file floors held. THE ONE ANCHOR: `packages/core/src/security/platform-admin.ts#parsePlatformAdminEmails` at README.md:172, inside a ```text fence where the README DOCUMENTS how to write an anchor. TRUE FINDING for the new behaviour, not a false positive: the forked matchAll was judging the README's own worked example, which is the same fabrication direction symbol-anchors' header states for itself ('a gate header that teaches the grammar must not fail it'). Each anticipated change, measured: extensions 8->23 = 0 anchors affected (population is ts/json/mjs only); fences now skipped = 1 anchor (the above); anchor-exempt now honoured = 0 (no marker exists anywhere in the family); existsSync -> git-tracked = 0 (no target untracked, none missing). NOTHING was narrowed to reach green, no floor lowered, no row joined the residual. HEADROOM NOTE: README.md now sits exactly at its floor of 1, zero headroom.",
  "corpora_not_moved": "Reverse verification of the core widening, from a comparison worktree at 4bd2c60e81 (both trees pnpm-installed): check-adr-symbol-anchors exit 0 -> 0 BYTE-IDENTICAL stdout+stderr; check-scripts-symbol-anchors 0 -> 0 BYTE-IDENTICAL; check-system-context-census 0 -> 0 BYTE-IDENTICAL. Measured before writing the dotted widening: 0 new dotted-span anchors across docs/adr (139 files), scripts/** (261 files through commentProse) and the system-context page. NOTE the first run of check-system-context-census in the cmp tree returned exit 3 'PREREQUISITE NOT MET - typescript is not installed'; that is NOT a reading, the tree was installed and it was re-run.",
  "gates": "32 derived families (node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack, derived at final HEAD 3683a4b9e from the merge base 4bd2c60e8, three-dot). ALL 32 exit 0 at final HEAD. Reconciled: 'dispatch-gates --ran: 32 derived famil(ies) accounted for - 32 run, 0 NOT-MEASURED (a DERIVED zero - all 32 recorded an exit code and none of them is 3)'. check:pm-dispatch-gates ran to completion in the foreground on a recorded PID (tail --pid), exit 0, 'dispatch-gates self-test: 1746 cases pass', 813.0s on this box (slower than the 430-450s note because a duplicate copy of the same gate was briefly running beside it; I killed the duplicate by its recorded PID, not by name). ONE GATE CAUGHT A REAL DEFECT IN THIS PR: check:entry-guard exit 1 on `export const CORPUS` - a scripts/** file that exports a binding AND runs on import can be imported for that binding, running the gate (and its process.exit) inside the importer. Fixed by dropping the export (nothing outside reads this corpus; the sibling corpus gates export theirs because their dispatch sits behind isEntrypoint). Commit 3683a4b9e, re-run exit 0. Also run though not owed: pnpm lint (full repo, eslint . --no-inline-config) exit 0 at 3683a4b9e - so no narrowing claim is being made. Control-byte self-scan of both touched files: clean.",
  "self_tests": "node scripts/symbol-anchors.mjs --self-test exit 0, battery floor raised 93 -> 122 (+29: the unspanned form and every #-shaped neighbour it must refuse, the dotted segments, excludeDirs both directions, each option with its DEFAULT-OFF control). Count read by probing the floor, not estimated. node scripts/check-platform-checklist.mjs --self-test exit 0, 189 assertions, symbol-anchor battery floor raised 40 -> 42. The 25 detector cases MOVED into the core's battery rather than stopping. EVERY re-judged #16898 case survives the transplant unchanged in verdict (R1-R4, P1-P4, P3b, N1-N6), per that card's own in-file rule that deleting one would leave the tree unable to say the narrowing happened - I had deleted them in a first pass and restored them when the battery floor caught it.",
  "ablation": "11 mutations, 11 red. Each: on-disk landing proved by anchor-text count before/after AND injected-text count (a zero-hit editor exits 0, so the editor's exit code was never the evidence); ran; restored with `git checkout HEAD -- PATHS`; restore proved byte-identical by `git hash-object` against the HEAD blob plus an empty `git diff HEAD`, with an empty-hash guard reading as FAILURE. trap RESTORE_FN EXIT INT TERM on absolute paths from `git rev-parse --show-toplevel`. Mutations: M1 walker ignores excludeDirs; M1b same seen by the gate; M2 sweepCorpus stops forwarding unspannedAnchors (gate red, 70 problems); M3 unspanned pass stops de-duplicating against spans (spanned anchor counted twice); M4 only the head dotted segment resolved (R3, P3); M5 residual ledger matches nothing (123 problems); M6 corpus population drifts from familyFiles; M7 a private detector returns (V2); M7c shared vocabulary replaced by a copy of itself (V1 + V2); M8 familyFiles stops excluding runs/ (L2); M10 the V2 source read made to return nothing (V3). ONE ABLATION FOUND A DEFECT IN MY OWN CHECK: M7b injected an ANCHORABLE_EXTENSIONS_SHADOW set and V2 PASSED - the pin was written as an exact-name test and a near-name fork slipped it. Tightened to a name PREFIX with any suffix allowed, re-ablated, red. The residual gap (a fork under a name sharing none of these tokens) is stated in the source rather than implied away.",
  "changeset": "skip-changeset, measured not assumed. Of 75 package manifests, 0 name scripts/symbol-anchors.mjs or scripts/check-platform-checklist.mjs in files[]; the root manifest is private:true with no files[]. FIRING CONTROL on the same reading: 70 manifests DO ship CHANGELOG.md in files[], so the 'docs and scripts are always exempt' reflex is not what this rests on. A second probe (grep my symbols in built dist/) is VOID and reported as such: this worktree has 0 dist directories, so its own positive control did not fire. Label applied via the additive endpoint POST /issues/18588/labels (HTTP 200 with Content-Type: application/json; the first attempt without that header was HTTP 415). Comparison read-back: read-set was {size/l}, target {skip-changeset}, union {size/l, skip-changeset}, read-back {size/l, skip-changeset}, STRIPPED: none. Re-read after the gate run: still present, so the size-labeler PUT did not take it.",
  "api_writes": "5 - POST /repos/.../pulls (draft) x1; POST /repos/.../issues/18588/labels x2 (first HTTP 415, no write; second HTTP 200); POST /repos/.../issues/18107/comments x1 (this report); plus 4 git pushes to the feature branch. Reads not counted.",
  "mcp_calls": "0 - every GitHub write and read went through the REST proxy with curl and $GITHUB_TOKEN; no MCP GitHub tool was called.",
  "open_questions": [
    {
      "question": "Clause-2 was dispatched as `no` and I did NOT change the declaration - but the actual diff does two things the dispatch told me to report rather than re-declare: it ADDS one export, and it WIDENS the shared accept set.",
      "options": [
        "A - leave Clause-2: no. The widened surface is scripts/**, which 0 of 75 package manifests publish, so no released contract moves; and the accept-set widening was measured inert (0 new anchors, all three registered gates byte-identical).",
        "B - flip to Clause-2: yes and attach needs:contract-review. `scripts/symbol-anchors.mjs` is the single shared resolver for five corpora and its own header calls a widening of the accept list something that 'moves every census in the same stroke'; the dotted-SYMBOL half is NOT opt-in and lands on every corpus."
      ],
      "recommendation": "PM's call, not mine - the dispatch is explicit that I must not edit my own declaration. FACTS: (1) one new export from scripts/symbol-anchors.mjs, `symbolSegmentResolution`; none removed; check-platform-checklist's own export count is unchanged at 2 (both of those are fixture text, and the real `export const CORPUS` I briefly added is gone). (2) Accept-set widenings: (a) `SYMBOL` now admits dotted segments - GLOBAL, no opt-out, measured 0 new matches across docs/adr, scripts/** and the system-context page, and all three gates byte-identical; (b) `UNSPANNED_ANCHOR` - a new anchor FORM, but gated behind `unspannedAnchors`, default false, and no existing corpus sets it. (3) Shape additions: defineCorpus's returned object gains 2 fields, sweepCorpus's return gains `byDoc` and `counts.citedSources`, extractAnchors gains an optional 2nd parameter, anchor records gain `unspanned`, unresolved-symbol/unresolved-path findings gain path/symbol/absent. (4) One new REFUSAL: defineCorpus rejects a separator in `excludeDirs` - a narrowing, but only on a field that did not exist before. If B, the PR needs the label and a `--pair` exit code, which I have not applied."
    },
    {
      "question": "The gate's `UNRESOLVABLE ANCHOR` check became unreachable, and I did not reintroduce it.",
      "options": [
        "A - accept as-is and card it. The old detector matched any 1-4 letter extension and ERRORED on one outside its 8; the shared grammar simply does not match a non-anchorable extension, so an author writing `#symbol` on a .txt now gets silence where they got a red. Measured: fires 0 times on today's population (ts 622 / json 9 / mjs 3).",
        "B - add a fourth corpus knob to the core so a corpus can ask for anchor-shaped tokens with a foreign extension to be reported."
      ],
      "recommendation": "A. Reintroducing a local matcher here is precisely the fork the ruling forbids, and B is a fourth widening beyond this card's acceptance list on a check that fires zero times. It is stated in the PR body under its own heading and listed below to be filed."
    }
  ],
  "out_of_scope_findings": [
    "to file (3 classes, dedupe words: UNRESOLVABLE ANCHOR, anchorable extension, foreign extension, silent skip, platform-checklist anchor) - (c)-class authoring trap introduced by this PR and stated openly in its body: an anchor whose extension is outside the shared anchorable vocabulary is now SILENTLY not an anchor, where check-platform-checklist used to red with 'UNRESOLVABLE ANCHOR ... Cite that file BARE'. Repro: put `\"source\": \"a/b.txt#Symbol\"` in any areas/*.json - before 4bd2c60e81 it reds, after this PR the gate is green and the citation is judged by nothing. Measured impact today: 0, the live population is ts 622 / json 9 / mjs 3. The honest fix is a corpus-declarable knob on scripts/symbol-anchors.mjs, not a local matcher here.",
    "noted, not filed: docs/qa/platform-checklist/README.md now sits exactly at its symbol-anchor floor of 1, zero headroom, because the anchor it lost was its own fenced worked EXAMPLE of the grammar. The next citation re-authored out of that file reds the gate with a MAINTAINER-ONLY remedy. Who will meet it: the next PR that edits that README - #18104's residual drainage is the likeliest, since it re-points citations across the family.",
    "noted, not filed: check-platform-checklist still carries a SECOND detector beside the registered one - `SOURCE_LINE_CITATION` / `findSourceLineCitations`, with spellings the shared LINE_ANCHOR does not reach (`~L7246-7331`, a bare `L7477-7493`, a bare `:140` continuation). That is the same fork one surface over. Deliberately NOT merged here: the line-citation half is another card's, and the shared rule finds 0 line anchors in this family today so nothing is currently unjudged. I added a live population-agreement check (CORPUS POPULATION DISAGREES / L2) so the two walkers can no longer drift apart silently. Who will meet it: whoever takes the line-citation half."
  ]
}

---

Generated by Claude Code

[os-try-charles]

决策卡:注册后一条硬红消失 —— 门禁削弱,按 SKILL.md:390 恒交维护者

维护者速读(2026-09-17T06:21Z 读):PR #18588(绿、draft、头 3683a4b9e)把 platform-checklist 注册成共享语料,卡面要求的全做到了;但顺带让一条 UNRESOLVABLE ANCHOR 硬红永远不再响。今日实测命中 0。我不代裁门禁削弱,请裁 A / B / C。我的建议:A。

一句话问题

引用文档里把源引用的扩展名写错(例如 a/b.txt#Symbol)时,门禁以前当场红并告诉作者怎么改;注册成共享语料后它会安静地什么都不说 —— 写的人以为「这条引用已被校验」,而其实没有任何东西在看它。

选项 × 真实代价

选项	做什么	客户/作者可感知的后果
A(推荐)	照此落地,另立一张 finding 卡记这条静默类	今日 0 人撞上(实测:该家族现存锚点 ts 622 / json 9 / mjs 3,该红在 main 上今天也一次不响);且注册后净判更多 —— 原先无条件红的 15 个扩展名现在真的能解析。代价:x.txt#Symbol 这类写法在那张卡做掉前从「红」变「哑」
B	本 PR 内给共享核再加第四个语料开关,把「锚点形状但扩展名不在词表」的 token 报出来	作者仍拿到响亮拒绝。代价:超出本卡验收清单的第四次放宽,落在五个语料共用的解析器上,只为一个今天 0 命中的检查;本 PR 再涨一轮
C	退回本 PR,恢复本地 detector	红回来了,但直接违反 2026-09-01 总监裁决(⛔ 不得为第二语料 fork scripts/symbol-anchors.mjs),该裁决正写在那个文件头部

业务含义直译

A = 「先删掉两份会悄悄漂移的文法,欠下一条今天没人踩的告警」;B = 「为一个今天没人踩的告警,给全平台共用的解析器再加一个永久开关」;C = 「为保住一条告警,留着那份已被判死的重复实现」。

四轴(业务立场)

• 长远合理性:A 缩小特例(少一份分叉文法,五个语料同一套规则);B 扩大契约增生;C 恢复特例。
• 业务拉动:今天 0。拉动发生在未来某位作者写错扩展名、而没有任何东西告诉他的那一刻。
• 防 AI 犯错:这一棱 A 是退步(响亮拒绝 → 静默容忍),B 最优。所以 A 的补偿控制必须是一张卡,不是一句注释。出错时谁看到什么:A 下 —— 作者看到绿,引用指向不存在的符号,无人看到任何东西;B 下 —— 作者当场看到红和改法。
• 创业阶段不扩散:A 是 remove;B 是 declare-and-maintain,每个已声明的开关(哪怕默认 off)都是永久义务。

os-decision-facets
① 项目长远合理性:A 缩小特例(删一份分叉文法),B 扩大契约增生(第四个共享开关),C 恢复特例。
② 实际业务拉动:今天 0 —— 无人撞上;拉动只在未来作者写错扩展名时出现。
③ 防 AI 犯错:A 把响亮拒绝换成静默容忍(单棱反向),B 保住响亮拒绝;故 A 必须配一张卡。
④ 创业阶段不扩散:A = remove;B = declare-and-maintain,默认 off 的开关同样是永久义务。
Prior rulings read: unresolvable,anchorable,extension,silent → 11 hits; ADR-0106 D6, ADR-0029 D9.9, ADR-0056 D8, ADR-0057 D7, ADR-0058 D6, ADR-0062 D3, ADR-0062 D7, ADR-0094 D5, ADR-0105 D5 —— 逐条读过,全部是运行时失败姿态/权限投影面的裁决,无一条裁的是门禁检查的去留,故不构成既有裁决。

推荐 + 回退 + 置信缺口

推荐 A。 自检行:只看①选 A;②③④ 是否翻转:否(③ 单棱反向,但其今日拉动为 0,且 A 的补偿控制 = 同轮另立卡)。回退项 B(若你认为「静默」这一格在门禁面不可让步)。置信缺口:我没有实测未来作者写错扩展名的频率;我只能读到「今天 0 命中」,读不到这条红在过去是否响过 —— 门禁不留痕。

裁后执行(你只裁方向,不背执行)

• 裁 A ⇒ 我同笔把卡换回工作态,复核放行 refactor(scripts): register platform-checklist as a symbol-anchor corpus instead of forking the grammar #18588(翻 ready → 重读 → 入队 → 两读落地),并立刻开一张 finding 卡记静默类(去重词:UNRESOLVABLE ANCHOR / anchorable extension / silent skip)。
• 裁 B ⇒ 我把 PR 打回同一施工席补第四个开关(默认 off + 默认 off 对照),重跑 32 个派生门禁与两套 self-test 后再复核。
• 裁 C ⇒ 我把本卡退回分诊,并把它与 2026-09-01 裁决的冲突写在卡面上,由分诊定去留。

读数时点 2026-09-17T06:21Z;PR #18588 头 3683a4b9e,对照 origin/main @ e46ac0fb65。条款②已由我席内复核为 no(PASS,读数与发火对照写在 PR 评论 5709871252),本决策只关门禁削弱一项。

---

Generated by Claude Code

[hotlong]

Ruling: batch #149 item 2 · letter A (PR #18588 lands as it stands — platform-checklist registered as a shared corpus; the silent class is one limitation sentence, ⛔ not a compensating finding card) · maintainer 「同意」 2026-09-17T14:41Z

Director seat, summon #24, session_01Wj1HUjzyeiBQ8atRf1ZhaL. Presented in detail with the recommendation A; the maintainer agreed. Facts (dev report 5709783374, decision card 5709893125, seat re-read of the PR): the 2026-09-01 ruling 「one resolver」 says a second corpus registers into scripts/symbol-anchors.mjs and never forks it; the platform-checklist gate had forked the grammar and the two copies had drifted (anchorable extensions 23 shared vs 8 local; the shared detector skips code fences and honours waiver marks, the local one was a bare matchAll). PR #18588 (head 3683a4b9e, green — 24 checks passed / 11 skipped — mergeable clean, 2 files, +661/−236) makes CORPUS a defineCorpus call and the sweep one sweepCorpus call, deletes the four forked symbols, and widens the core twice (excludeDirs, unspannedAnchors — both corpus declarations, default off, each with its own control), which is the ruling's own exit 「if a corpus needs behaviour this core lacks, widen the core」; the resolution rule is untouched. Side effect the dev stopped on: the UNRESOLVABLE ANCHOR hard red for a mis-typed extension (a/b.txt#Symbol) no longer fires — it hits 0 today (family anchors ts 622 / json 9 / mjs 3), and check:platform-checklist is not in CI: it runs on the manual cadence the maintainer set, so the red was ever only visible to whoever ran it.

Ruling — A

• PR refactor(scripts): register platform-checklist as a symbol-anchor corpus instead of forking the grammar #18588 lands as it stands. The silent class is recorded as one limitation sentence in the gate header's limitation block of scripts/check-platform-checklist.mjs (the form PR measure the return-propagating durability seams the log-level gate cannot name #18523 established for the durability gate) — a mis-typed extension is not reported; the anchor is simply not swept. The same seat pushes that line onto refactor(scripts): register platform-checklist as a symbol-anchor corpus instead of forking the grammar #18588, re-reads the new head at tier, and drives it in (ready → queue → two-reading landing). ⛔ No compensating finding card: under the maintainer's correction of the AI-error axis (5716140776) a gate-completeness gap on a fleet-only, not-in-CI check scores zero, and a card for it is the 「一直开发门禁」 the maintainer refused.
• ⛔ B — a fourth switch on the resolver five corpora share, for a check with zero hits; ⛔ C — restores the fork the 2026-09-01 ruling forbids.

Four-facet reading: ① zero today; ② one grammar for five corpora instead of a fork that drifts silently; ③ not scored — a fleet-facing gate, no platform user meets it; ④ A is a removal, B is declare-and-maintain.

Execution

needs-user-decision → pm:dispatched (the card returns to its working state: assignee os-try-charles stays, PR #18588 open); domain:devx, p2 stand. The card closes when #18588 lands — its whole ask is in that PR.

---

Generated by Claude Code