守卫能力:objectstack lint 的 warning 不 gate(exit 0)——warning 级缺陷在 pnpm verify 里永远不会红,只能靠人眼或逐条复刻成测试

[yinlianghui]

源自 #715(PR #1007)实施过程,dev 记录、PM 立单。观察级(能力缺口),不指派。

现象(已实测)

pnpm exec objectstack lint --skip-i18n; echo $? 在 13 条 warning 下返回 0。pnpm verify 里的 lint 步骤因此对任何 warning 级缺陷都不可能失败——包括 #715 的 field-group-shadowed 本身:它在 lint 输出里躺了不知多久,直到 rc2 验收人工扫出来才立单。

现状缓解(仓内已形成的模式)

把重要规则逐条复刻成测试:test/field-groups-coverage.test.ts 现已复刻 validateSemanticRoles 全部三条规则(#715 补全第三条)。这解释了为什么这个文件存在——它就是 lint 不 gate 的代偿。

可选收口方向(不预判)

1. lint 配置升级部分规则为 error(动公共面,需要平台支持按规则设 severity);
2. verify 链里加 warning 计数基线(新增 warning 即红,存量豁免);
3. 维持"逐条复刻成测试"的现模式,接受选择性覆盖。

查重

• CI step name for the source hygiene gate lists only three of its four checks #753 是 CI step 命名漏列检查项(命名问题,非 gate 能力);
• [观察] 文档里的状态名/字段标签与语言包是否一致，无人检查——一周内同一缺陷类出现三次，全靠人眼撞见 #802 是文档↔语言包一致性无守卫(另一面);
• 均不同面,故新立。

Refs #715 · PR #1007

[huangyiirene]

New measured evidence for this finding, merged in from #1083 (filed today by me without a dedup search — my error; #1083 is closed as a duplicate of this card).

#1060 / PR #1080 produced a concrete, dated instance of exactly what this card describes, on the i18n rule family:

• objectstack lint exit code is 0 both with and without --skip-i18n — measured directly, it does not fail the process on warnings alone.
• PR chore(lint): drop --skip-i18n from lint script; fill es-ES/zh-CN page-component i18n keys #1080 merged (2026-08-11) carrying 25 known, enumerated i18n/missing-page warnings (the ja-JP third, deliberately held back for a concurrent claim), and the Quality Checks job was green. That is the failure mode of this card as a merged artifact rather than an argument: a branch with named, counted debt passed the gate.
• Consequence for how results get read: chore(lint): drop --skip-i18n from the lint script — i18n debt is zero, coverage should enter the gate (successor to #494) #1060's title asserts "coverage should enter the gate". After chore(lint): drop --skip-i18n from lint script; fill es-ES/zh-CN page-component i18n keys #1080 the coverage runs and prints; it does not block. Anyone reading the card would reasonably conclude a regression can no longer land. It can.

Two knock-on effects worth recording here, since this card is the root:

1. Acceptance criteria in this repo have to be stated as rule-hit counts, not exit codes. fix(i18n): fill 25 missing ja-JP page-component translation keys #1084 closed the debt and reported it correctly as i18n/missing-page 25 → 0 rather than "lint exits 0", precisely because the exit code proves nothing. Any future card whose acceptance is "lint passes" is under-specified while this finding stands.
2. The options for fixing are split by repo: making the rule error-severity is a platform-side change in packages/lint (upstream), while a CI step asserting a zero count is app-side and can be done here. Only the latter is in this repo's power.

No new card needed — this one already covers it, and now has a current instance attached.

Refs #1060 · PR #1080 · PR #1084 · #494

---

Generated by Claude Code

[huangyiirene]

PM triage sweep, 2026-08-11 — using this card as the anchor for a cluster, and promoting it to pm:queue.

A findings sweep over all 34 open finding cards shows they are not 34 independent observations. Eight of them say the same structural thing: a gate in this repo has a blind spot. Listing them together because the cluster is the finding, and each one read alone understates it:

Card	The blind spot
#1018 (this one)	objectstack lint warnings never fail the process — warning-class defects can never go red in pnpm verify
#838	The control-byte scan cannot see docs/, .github/, .claude/ or the root first-class files (including three root .ts configs)
#809	The "built-in sharing rules table" doctrine rule reads only the English page; the other two locales' tables are unchecked
#802	Nothing checks that status names / field labels in the docs match the locale packs — the same defect class surfaced three times in one week, all caught by eye
#1003	The retired-persona rule scans only content/docs; user-visible strings in src/ (title / label / description / help) are unchecked
#867	No anchor-existence or locale-prefix check on in-docs links (#749 / #764 / #866 each had a cause this would have caught)
#935	docs-drift's headingLabel() rejects explicit fumadocs heading ids, so the dashboards page can never carry [#id]
#753	The CI step name for the source-hygiene gate lists three of its four checks

Why this cluster deserves attention ahead of its individual members' severities. Today's work produced a card that states the pattern outright — #1091: nine object files carry an unused import, and typecheck, lint and hygiene are all green on it. Several of today's defects were caught by a person or an agent reading carefully, not by a gate:

• chore(lint): drop --skip-i18n from the lint script — i18n debt is zero, coverage should enter the gate (successor to #494) #1060 shipped 25 enumerated i18n/missing-page warnings through a green Quality Checks — this card's exact failure mode, as a merged artifact.
• fix(opportunity): retire the placeholder competitors picklist (#1061) #1090 removed a placeholder picklist that had sat in production metadata since the app was authored, translated into four locale packs.
• A hand-filled period_end walks around #1008's calendar-boundary fix — a row labelled Q3 2026 can still span ten months #1093 (filed today) reports that a hand-typed period_end walks around a boundary rule that shipped this morning.

An unguarded class does not stay one instance. That is the argument for treating these eight as one body of work rather than eight separately-scheduled small cards.

This card promoted to pm:queue as the cluster's first and cheapest member: the app-side half is a CI step asserting a zero rule-hit count, which is small, and it is the gate that would have caught #1060's 25 warnings. Note the split the card already records — making the rule error-severity is an upstream packages/lint change; only the count assertion is in this repo's power.

⛔ Not merging the other seven into this card. Each has its own file surface and its own measurement; folding them in would make an epic that never closes. They stay separate, now cross-referenced.

Sweep scope, stated honestly: this pass read the 34 cards' titles and labels, not their bodies. Premise-staleness — which of these were silently fixed by today's seven merged PRs — has not been checked, and remains the outstanding half of the triage round.

Refs #1091 · #1060 · #1093 · #494

---

Generated by Claude Code

[huangyiirene]

⚠️ Correction to my triage-sweep comment above — one of its three pieces of evidence was wrong.

That comment cited #1091 as proof that this repo's gates have blind spots:

Today's work produced a card that states the pattern outright — #1091: nine object files carry an unused import, and typecheck, lint and hygiene are all green on it.

#1091's premise has since been falsified (see PR #1095 and the dev report on that card). F and P are @objectstack/spec's tagged-template aliases, invoked as F`...` / P`...` and never with a dot. All nine files use them productively. The card's own reproduce command, grep -cE '\b[FP]\.', matches dot-notation only and so returns 0 regardless of real usage — the zero was the instrument's blind spot, not the code's.

So the gates were not blind to an unused import; there was no unused import. Strike that evidence line.

What the cluster's argument rests on now — and it still stands, on two measured instances rather than three:

• chore(lint): drop --skip-i18n from the lint script — i18n debt is zero, coverage should enter the gate (successor to #494) #1060 shipped 25 enumerated i18n/missing-page warnings through a green Quality Checks. Unchanged, and it is this card's exact failure mode as a merged artifact. This is the strongest instance and it is directly about this card.
• A hand-filled period_end walks around #1008's calendar-boundary fix — a row labelled Q3 2026 can still span ten months #1093 — a hand-typed period_end walks around a boundary rule that shipped the same morning. ⚠️ Note that card's own caveat: it is inferred from a measurement of the form, hook and rule set, not driven end-to-end, so it is weaker evidence than chore(lint): drop --skip-i18n from the lint script — i18n debt is zero, coverage should enter the gate (successor to #494) #1060 and should be reproduced before being counted as fact.

The other seven cards in the table are unaffected — each was measured on its own terms, and none of them depended on #1091.

The methodological point is worth keeping even though the instance evaporated, because it now cuts the other way too: #1091 was a finding whose measurement could not fail, and it survived in the backlog looking rigorous. That is an argument for the premise-recheck half of the triage round — the half that has not been done — rather than against the cluster.

---

Generated by Claude Code

[huangyiirene]

Claim: PM loop round 10

Session: session_01NM6o28jmBgsyTRQHutn7LC
Branch: claude/issue-1018-i18n-gate
Worktree: hotcrm-issue-1018
File surface: .github/workflows/**, package.json scripts, scripts/** if the assertion needs a helper, and test/. ⛔ Not src/translations/** — #597 is in flight on those. Stop on breach; explain in the report.
Container & model: S–M, model: sonnet, mode:subagent.
Serial constraints cleared: nothing else touches CI config; #597 (concurrent) is confined to src/, seeds and the locale packs.

⚠️ Read the correction comment on this card before designing. My cluster analysis above originally cited #1091 as evidence that the gates are blind; that evidence was falsified and I struck it. The argument still stands on #1060 (25 enumerated warnings merged green — this card's exact failure mode as a merged artifact), but the second supporting instance (#1093) carries its own "not reproduced end-to-end" caveat. ⛔ Do not repeat the original three-instance framing anywhere.

Scope is the app-side half only: a CI assertion that the rule-hit count is zero. Making the rule error-severity is an upstream packages/lint change and is not in this repo's power.

---

Generated by Claude Code