fix: feature QA observation captures JSON-LD structured data blocks

[sanrai]

The bug

The Feature QA review reported FAIL on PR #567 twice, claiming no script[data-caas-jsonld] tag was present, while the same build provably injects it (green run-e2e-tests asserting the tag parses, plus manual DOM inspection on localhost).

Trace: the Step 5 observation extracts only card-level data, .consonant-Card nodes with title, text, and links. Script tags, meta tags, and container-level attributes never enter the observation. Step 6 then hands the judge only that card extraction, so any assertion about non-card DOM is structurally unverifiable and yields an automatic FAIL. The judge's own wording confirmed it: "not present in the rendered card/container data provided".

The fix

The observation now also captures up to four script[type="application/ld+json"] blocks, each with its parent element, attribute names, and up to 1500 chars of content. They are passed to the validation judge as a separate labeled section and the PR comment reports the block count.

• Observation returns { cards, jsonLd } instead of a bare card array; both downstream consumers (judge prompt, comment renderer) updated to match.
• Verified with a static consumer-consistency check (all consumers read the new shape, no stale references) and node --check.

Scope note

This fixes the one blind spot that produced a false FAIL today. The general class, non-visual DOM assertions (meta tags, aria attributes, dataset attributes), still shares the limitation; a follow-up could let the planner emit an observeSelector the harness evaluates directly. Kept out of this PR to stay minimal.

Expected effect

Re-running Feature QA on PR #567 after this merges should flip its verdict to PASS, since the judge will finally see the block it was asked about.

[github-actions]

Agent QA review — interactive + visual diff (advisory, non-blocking)

Last updated Aug 5, 2026, 1:59 AM PDT · PR opened · commit 90ebbd6 · 2 files changed.

0 open · 0 resolved · visual diff 0.00% · verdict PASS

Open findings

• Smoke test passed ✅

What the agent checked

Tested PR #569 on the live business.adobe.com/resources/main.html page with the PR's CaaS build injected.

Context: This PR only touches QA tooling (.github/qa/feature-review.mjs) and adds a new test fixture file (html/e2e/verify-local.html). It does not change any CaaS runtime/rendering code. Consistent with that, the pixel diff of PR vs stable showed 0.00% changed pixels, and the loaded diff.png showed no magenta/highlighted regions — confirming no visual change was introduced by this PR.

What I verified on the live page:

1. Initial page load: Resource Center renders normally, header, breadcrumb, filter panel, card grid, and chat widget all display correctly with no layout issues.
2. JSON-LD structured data: confirmed via evaluate() that two <script type="application/ld+json"> tags exist in (lengths 134 and 278 chars) — exactly the kind of non-card DOM artifact this PR's observation step is meant to capture for QA judging. This confirms the underlying premise of the PR (that JSON-LD is genuinely present) is correct.
3. Filter panel interaction: expanded "Products" filter group, checked "Acrobat" — filter chip appeared ("Acrobat x"), checkbox count badge updated to "1", result count updated from 3771 to 181, and card grid correctly refreshed to Acrobat-related content (Forrester Acrobat Studio study, Acrobat PDF accessibility). Unchecking the box worked cleanly (clear function fired without error).
4. Search input: typed "photoshop" into the left-panel search field without errors.
5. Console errors: none observed after all interactions.
6. Accessibility: ran axe-core scoped to the filter panel — 0 violations, 23 passes, 0 incomplete.

No broken, misaligned, truncated, or low-contrast elements were found. All interactive elements (filter checkboxes, search input, clear buttons) behaved as expected. Since this PR is purely a change to the internal QA review script and a new local verification fixture (not shipped to the live site), there is no user-facing regression risk, and testing confirms the live page renders and behaves identically to stable with no defects.

Verdict: PASS — no regressions found; page and PR changes are QA-tooling only and functioning as intended.

PR / stable / diff screenshots + console + axe artifacts in the workflow run.

Review history (1 run)

• 90ebbd6 · Aug 5, 2026, 1:59 AM PDT · PR opened · passing, no regressions — fix: feature QA observation captures JSON-LD structured data

[sanrai]

Folding this into #567 instead: the Feature QA workflow preserves and runs the reviewer from the PR's own merge ref, so carrying the observation fix inside #567 means the next review of #567 already uses it. Same change, cherry-picked there.

[github-actions]

🧪 Feature QA review — injected feature test (advisory, non-blocking)

Last updated Aug 5, 2026, 2:01 AM PT · commit 7e72275 · pull_request.

Overall: SKIPPED

Scenario mapping was not proven -- skipped instead of guessing.

The changed unit test in this PR (.github/qa/feature-review.mjs) is a test-harness/tooling change to the QA observation script itself, not a product feature under test in the CaaS card grid. It captures script[type=application/ld+json] tags for the judge's benefit, but the research explicitly found no production consumer (showJsonLd, JsonLd, structuredData) anywhere in react/src/js, and no evidence that any card config field causes the app to render a script[type=application/ld+json] tag. The verify-local.html example uses a 'showJsonLd':true flag in its state payload, but source search across react/src/js turned up zero matches for that flag being read/consumed, so there is no proven injection path from card/config JSON to an actual rendered ld+json script tag. Fabricating a card style or

Code searches performed: 8.

Screenshot in the workflow run.

Review history (1 run)

• Aug 5, 2026, 2:01 AM PT · SKIPPED · 7e72275 · pull_request