metadata-admin DOM tests: two surfaces that never settle under the light `dom` project, with no diagnostic

[yinlianghui]

Observation-class finding, measured while writing objectui#4132's pins (PR #4248). Nothing a user hits — this is a test-harness characteristic. Recording it because both specimens present identically (the test file produces no output at all and never exits), which is the most expensive failure shape to diagnose from scratch: there is no timeout message, no failing assertion, no partial reporter output, and --reporter=verbose does not help because it flushes per file.

Specimen 1 — a component that opens a Radix Sheet

Rendering MetadataDetailDrawer (which wraps its body in Sheet / SheetContent) inside the dom project ran past 540s with zero output. The component it hosts renders on its own in ~6s:

render	result
< MetadataDetailDrawer target={embedded} … >	no output, killed at 540s
< EmbeddedItemEditor … > (the entirety of the drawer's embedded branch)	8 tests, 6.7s

Consistent with the repo's own layout: the only test that renders an open Sheet — layout/__tests__/ChatDock.test.tsx — is in heavyDomTests (i.e. the dom-heavy project with the full setup). The two other light-project files that mention Sheet do not render one: preview/__tests__/CommitTimeline.test.tsx names it in a comment, and views/studio-design/__tests__/studioChatDockPersistence.test.tsx mocks it away (ChatDockMobileSheet: () => null).

Specimen 2 — a SchemaForm whose schema declares a condition property

SchemaForm.tsx:344 (CONDITION_FIELD_NAMES) routes visible / hidden / disabled / visibleOn / condition / predicate / *When to the CEL ConditionWidget, whose formula loader is a dynamic import (celAuthoring.__setCelFormulaLoader). Bisected one property at a time against an otherwise identical fixture:

schema property added	result
name	renders, 43ms
label	renders, 44ms
condition	no output, killed

This is the unbounded-module-load trap AGENTS.md §测试纪律 already describes, reached through a schema key name rather than through an obvious React.lazy in the component under test — which is why it is not obvious from the test's own source that a dynamic import is involved at all.

Why it is worth writing down rather than fixing here

Both are avoidable once known, and #4132 avoided both (mount the hosted component directly; keep the routed key out of the fixture's schema — the value is still asserted, on the preview that reads it off the draft). Neither is a defect in the product code, and neither should be "fixed" by parking a file in heavyDomTests — AGENTS.md restricts that list to registry <type> not registered failures and explicitly rejects it as a flake remedy.

Possible dispositions, none obviously right, hence finding and not pm:queue:

• leave as-is and let this issue be the searchable record;
• teach the light setup enough for Radix Dialog/Sheet to settle (has a cost — the dom/dom-heavy split in perf(test): make the DOM projects hermetic so they can run isolate:false (~3-4x suite speedup) #2648 exists precisely to keep heavy graphs out of ~290 files);
• have ConditionWidget no-op its loader when no loader is registered, so a fixture that names condition degrades to a plain input in tests instead of hanging.

Provenance

Measured on branch claude/issue-4132-retire-standalone-validation at 733935e3, container-shared runner, pnpm exec vitest run <file> --maxWorkers=2 from the repo root. Related: #2648 (closed — the DOM project split these two specimens land on).

---

Generated by Claude Code

[os-zhuang]

Findings triage (first grading) @ origin/main 6d01319: HOLD as finding — the searchable record is the deliverable.

Premises verified: SchemaForm.tsx:344 carries CONDITION_FIELD_NAMES verbatim (with the /When$/ route at :356), and the heavyDomTests escape hatch plus its AGENTS.md restriction ("only for registry not-registered failures, never a flake remedy") is live at AGENTS.md:193 — so both the trap and the reason not to park files around it stand exactly as filed.

Graded hold rather than queue: nothing user-facing, #4132/PR #4248 already routed around both specimens, and the filer's three dispositions genuinely trade off (teaching the light setup Radix costs the #2648 split its point; a no-op loader fallback in ConditionWidget changes widget behavior under test and deserves the owning lane's judgment, not a triage default). This card is the diagnostic that turns a 540s silent hang into a known shape — it earns its keep as a record today.

Restart condition: promote to pm:queue (likely the ConditionWidget loader-fallback direction, smallest blast radius) when a third silent-hang specimen lands on either trap, or when anyone opens the dom/dom-heavy setup for another reason.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code

[os-zhuang]

Concentrate round (state transition, no re-grade): finding → pm:on-hold, executing the standing HOLD verdict (2026-08-11): the searchable record is the deliverable; the three dispositions genuinely trade off and the ConditionWidget loader-fallback direction belongs to the owning lane's judgment, not a triage default.

Restart-when (per the standing verdict): a third silent-hang specimen lands on either trap, or anyone opens the dom/dom-heavy setup for another reason — promote then (likely the loader-fallback direction, smallest blast radius). Trigger files: packages/app-shell/src/views/metadata-admin/SchemaForm.tsx (CONDITION_FIELD_NAMES route), the celAuthoring loader seam, and the vitest dom project config — dispatchers touching these read this card first.

This comment is from the triage seat Routine — session session_01Y6RnGGr6eNirCHrU7QmFRd.

---

Generated by Claude Code

[os-zhuang]

Triage: domain:devx (lane only — pm:on-hold untouched). The defect is in the DOM test harness — two surfaces that never settle under it. Test infrastructure, so the engineering stream, even though the surfaces under test are product code.

⚠️ Same harness family as #5194 (await act(async …) never settling). Worth pricing together.

Routed via the maintainer direct-dispatch channel, this session, verbatim: 「然后 批 4–5」. PM session session_0124Qg8rLvpXnQDwCmpKUmaJ. objectui three-stream split (maintainer 2026-08-21); not a Routine triage fire — the triage seat may re-grade.

---

Generated by Claude Code