lint: 视图容器阶梯遍历在 packages/lint 内已有三份实现,彼此按不同判据取舍

[hotlong]

Filed unassigned from #6251 / PR(见下)。本单只记录发现,不含修法承诺。 观察类:今天没有任何用户会踩到它,三份实现在「可能装 section 的每一级」上结论一致 —— 记下来是因为它们各自按不同判据取舍,而这类分歧的成本是下一次有人只改其中一份。

事实

packages/lint/src/ 里「从一个 views[] 条目下降到它真正的表单/视图站点」这套遍历,现在有三份各自独立的实现:

实现	位置	走的阶梯
formViewSites	validate-visibility-predicates.ts（PR #6248 落地）	自身 + form + formViews.*
collectViewSites	validate-translatable-sections.ts	自身 + form.sections + listViews.*.sections + formViews.*.sections
formViewSites（同名,独立副本）	validate-form-layout.ts（PR for #6251,照抄第一份并加了绑定继承）	自身 + form + formViews.*

另有第四处形状相近但职责不同的 collectViewRecord（validate-translation-references.ts），以及 lint-view-refs.ts 走的是 spec 官方的 expandViewContainerWithDiagnostics。

已知的一处判据分歧（当前无害）

第二份额外走 listViews.*.sections。按 schema，ObjectListViewSchema = ListViewSchema.omit({ userFilters }).extend(…) 不声明 sections，所以那一级只可能读到 undefined；spec 的 expandViewContainerWithDiagnostics 也把 list / listViews.* 划为 list 族。即：当前三份结论等价，分歧只体现为一级读不到东西的多余下降。

绑定解析也是三套写法：objectName → object → data.object（两份）、name → id → object → list.data.object → form.data.object（lint-view-refs.ts，因为它要的是对象名而非表单绑定）。后者的差异是有理由的，不是漂移。

为什么现在不动

合并这三份必然要改 validate-visibility-predicates.ts —— #6248 刚验收的面；而只为一份新建 helper 文件会把三份变四份。属于「等第三个消费者出现，或等一次可以一并动这几个文件的窗口」的整理，不该由某个具体缺陷单顺手做。

值得关注的信号

这条遍历已经连续两单在两个不同文件里被发现是错的（#6128 / PR #6248，然后 #6251）。每次都是「只读容器自身键」这同一个错法。复制正确实现是当下最省的做法，但复制次数本身就是这条阶梯该有一个单一来源的证据。

Refs：#6251（本次发现处）、#6128 / PR #6248（同族前一例）、#4984 / #5009（幽灵检查族）。

[os-project-manager]

Findings sweep (maintainer-authorized one-off, 2026-08-07 — registered on #6015): promoted to the queue. Three independent ladder traversals that agree today by luck, diverging by criteria — the cost lands on whoever edits one of three next. Converge to one shared walker while they still agree (the #6268 loader-convergence precedent, same shape). finding → pm:queue.

---

Generated by Claude Code

[os-project-manager]

Deferred to a later round by the devx seat (round 1, 2026-08-08) — not shelved, and this card is not up for grabs.

Reason: #5957 is in flight and lands in packages/lint. The two file surfaces are disjoint — #5957 touches only validate-flow-template-paths.test.ts / reference-integrity-suite.test.ts fixtures, while this card converges validate-visibility-predicates.ts / validate-translatable-sections.ts / validate-form-layout.ts — but step 3 serializes on same-package doubt rather than betting on it. This card is first in line once #5957 merges.

Two things worth re-measuring before this is dispatched, because both are load-bearing for the card's own argument and both have a shelf life:

1. The card's central claim is that all three traversals agree today. That is what makes this a safe convergence rather than a behaviour change. It was true when filed; packages/lint has moved since. Whoever takes this re-measures the three ladders' agreement first — if they have already diverged, this stops being a refactor and becomes a defect card with a different scope.
2. The listViews.*.sections step in collectViewSites. The card argues it is a harmless descent that can only read undefined, because ObjectListViewSchema does not declare sections. That reasoning is sound but it is a schema claim, and the schema is owned by another lane — re-check it against packages/spec at dispatch time rather than inheriting it.

Also noted for whoever picks this up: the card's own "why not now" section says merging the three necessarily edits validate-visibility-predicates.ts, the surface #6248 had just verified. #6248 has since landed, so that objection is weaker now than when written — but the file-churn check still runs at dispatch.

---

Generated by Claude Code

[os-project-manager]

Claim: PM loop round 3 (devx seat)
Session: session_01F8q5J1MQyocgtNspb15fSn
Branch: claude/issue-6381-view-ladder-convergence
Worktree: objectstack-issue-6381
Domain: domain:devx
File surface: packages/lint/src/validate-visibility-predicates.ts, packages/lint/src/validate-translatable-sections.ts, packages/lint/src/validate-form-layout.ts, plus the shared walker's new home and its tests.
Container: M — mode:subagent, shared container.
Serial constraints cleared: the packages/lint hold this card was under is released — #6636 (#5957) merged at 487a197b5. That PR touched only *.test.ts fixture literals, disjoint from these three source files. No other in-lane claim on packages/lint.

Re-priced at the unblocking ref rather than inherited (the deferral note on this card asked for exactly this):

• The card's central claim survives: three independent implementations still exist and are still separate — formViewSites at validate-visibility-predicates.ts:680, collectViewSites at validate-translatable-sections.ts:190, formViewSites at validate-form-layout.ts:142. The fourth, differently-purposed collectViewRecord is still its own thing at validate-translation-references.ts:242. So this stays a convergence, not a behaviour change.
• ⚠️ One thing has changed since filing, and it narrows the work. The card treats collectViewSites's extra listViews.*.sections descent as a harmless step that can only read undefined. That step now carries a documented rationale in the file's own docblock, citing lint: translation-target-unknown misses a view container's DEFAULT form.sections — correctly translating a rendered heading is reported as a stale key #5415 — the default form.sections anchor that is "neither a formViews.* entry nor the record's own". A convergence that flattens the three ladders to the narrowest common shape would silently delete a step someone established on purpose. The union, not the intersection, is the safe target — and lint: translation-target-unknown misses a view container's DEFAULT form.sections — correctly translating a rendered heading is reported as a stale key #5415's reason has to be read before any step is dropped.
• The card's own "why not now" objection (that merging necessarily edits validate-visibility-predicates.ts, then-freshly-verified by feat(lint): view/page 谓词裸标识符构建期闸门 —— 坏谓词发不出去 (#6128) #6248) has decayed: feat(lint): view/page 谓词裸标识符构建期闸门 —— 坏谓词发不出去 (#6128) #6248 landed some time ago. The objection is weaker now, as the deferral note predicted.

⚠️ Adjacency, flagged so the dev does not wander into it: #6422 concerns collectViewRecord's listViews/formViews spelling and is held in the maintainer decision box (⛔ not dispatchable). It is a different file and is explicitly out of scope for this card. Do not "helpfully" fold the fourth implementation in.

---

Generated by Claude Code

[os-project-manager]

ACCEPT — PR #6657 (devx seat review, round 3).

What shipped: one shared walker (packages/lint/src/view-walk.ts, internal, following the page-walk.ts / flow-walk.ts precedent) yielding typed sites, consumed by all three validators — viewContainerSites() for the full ladder, formViewSites() as a filter over the same ladder rather than a second ladder. +440/−169 across exactly the five expected files.

Review notes:

• The union constraint held, and with better reasoning than the dispatch demanded. The listViews.* rung was kept; the schema proof (ObjectListViewSchema declares no sections — cited to view.zod.ts lines) is recorded in the walker's docblock as the reason the filter exists, not used as licence to delete the rung — with the additional observation that os lint runs over the normalized stack, so the schema argument alone was never sufficient. lint: translation-target-unknown misses a view container's DEFAULT form.sections — correctly translating a rendered heading is reported as a stale key #5415's default-form anchor reasoning now lives in one place instead of three.
• The reverse-verification table's third row is the load-bearing evidence: dropping the listViews rung reds only validate-translatable-sections (6 failures) while both form rules stay green — proving in one measurement that the filter is a filter, the rung is live for exactly one consumer, and the convergence actually happened (rows one and two red all three suites from one edit, 22 and 15 failures).
• Refactor-grade differential: 888 rule runs / 3295 findings over 222 generated stacks, compared with JSON.stringify so order counts — byte-identical against origin/main baselines. This is what "a refactor that changes verdicts is a failed refactor" looks like when it is actually checked.
• Binding resolution deliberately not folded — the three files' different fallback compositions stay put; the one equivalence that WAS collapsed (entry-site vs sub-container formula in translatable-sections) ships with its proof written next to it.
• Boundaries held: collectViewRecord (lint: collectViewRecord 的 listViews/formViews 分支收「map key + 内层 name」两种拼写,而组装器只认 map key —— 冲突改名时两者恰好相反 #6422, decision box) untouched; lint-view-refs.ts untouched; packages/spec read-only. Same-day churn from fix(cli,lint): 分区标题只读声明拼写 label,删掉两处 ?? title 容忍读 (#5730) #6616 read and confirmed disjoint before starting.
• Consumption radius: the CLI suite (94 files, 978 tests) run in addition to lint's own — the CLI is what actually executes these rules.
• skip-changeset measured rather than assumed (no observable change; walker not exported), label applied as union and read back intact (size/l + skip-changeset).

Ready + auto-merge armed 08:58Z. The card's "signal" paragraph — two consecutive issues fixing the same missing rung in one copy each — is now structurally closed: a broken rung breaks every consumer at once, which is the property the reverse-verification table demonstrates.

---

Generated by Claude Code