field-groups-coverage's shadowed check can produce a verdict the running app contradicts — for the three objects that author their own record page #1704

Restart-when: epic #1583 对 test/field-groups-coverage.test.ts 的归属作出复判 —— 该行补上 group A only 标注(或该文件被明确改判 mixed),使「哪一半可动、哪一半属 epic」有了可查的答案。

Filed by the repo:hotcrm PM seat, from the browser measurement delivered on #1679 (PR #1703). The dev reported it rather than filing, because the dispatch reserved the filing to the PM — this card is that filing.

⭐ This is the finding, not the one #1679 was dispatched for. #1679 corrected a failure-message string. What the measurement also established is that the assertion's own reach is narrower than it reads.

What was measured (browser, @objectstack/console 17.3.0)

test/field-groups-coverage.test.ts's shadowed check walks every object and computes the highlight-strip hoist from obj.highlightFields. A double dissociation shows that computation is only valid on the synthesized path:

⇒ The check fires in both cases, but the harm it names exists only on the synthesized path. On an authored page it emits a verdict the running application directly contradicts.

The mechanism, also measured: an authored page's body is governed by that page's own record:highlights, not by the object's highlightFields. Baseline crm_lead drops exactly the six fields its own record:highlights lists, while company — which is in the object's highlightFields and inside this check's first-four window — renders.

⚠️ The null was made trustworthy rather than assumed: every leg required GET /api/v1/meta/object/NAME to serve the intended highlightFields before any DOM was read (that gate caught and voided two earlier runs), and the synthesized leg carried a positive control — crm_account, dropped from highlightFields by the same edit, reappeared as a rendered row. A field appearing and a section vanishing in one run is what proves the authored leg's "nothing changed" is the page's null and not a dead instrument.

The three affected objects

crm_lead, crm_case, crm_opportunity — the objects in this repo that author their own record page. For these three the shadowed check can report a group as "entirely title-or-strip" when the page renders those fields perfectly well.

⚠️ It is green today, so nothing is failing. The defect is latent: it is a check that can produce a false positive, and a false positive on a coverage guard is expensive in a specific way — the next person to hit it will either weaken the assertion to make it pass, or spend a round proving the app is fine.

What needs deciding — ⛔ not a dev-agent call

The honest fix is to narrow the check's reach so it only judges objects whose detail page is synthesized. But that changes what the assertion asserts, and:

• test/field-groups-coverage.test.ts sits in epic Epic: bring this repo's test farm back under the 2026-08-31 ruling — platform-first, os lint --strict first, then retire the local re-implementations by family #1579 family F2's retirement table (Retire the local metadata reference-integrity tests the platform's page / view / nav / action / field-group rules already enforce (epic #1579, step 3, family F2) #1583), and — as this seat observed on 5558354861 — it carries no group A only marker, so by that table's letter the whole file retires;
• this seat's standing fence on the file is comment-only, ⛔ zero assertion changes, which is exactly what let field-groups-coverage.test.ts justifies itself with the fallback mechanism #1521 measured false: fieldGroups sections a detail page only on the SYNTHESIZED path #1674 and test/field-groups-coverage.test.ts:188: the shadowed-group failure message says "detail pages" unscoped — and its correct wording depends on an UNMEASURED question #1679 touch it at all.

⇒ ⛔ Not dispatchable under the current fence. It needs the epic PM's re-judgement first.

Routed

This card is input for epic #1583's re-judgement, alongside this seat's earlier observation (5558354861) that the census row buckets all of this file as a platform-lint candidate while its own synopsis covers only the integrity half. ⭐ The two now point the same way: this file is mixed, and its two halves need different treatment — the coverage half is a business-fact pin no named platform rule covers, and the integrity half contains a check whose reach is wrong.

⛔ This seat does not dispatch it and does not touch #1583's labels.

Refs #1679 / PR #1703 · #1674 / PR #1680 · #1583 (F2) · #1579 · #806 · #1521

Generated by Claude Code