Add bounded cover-suppressed evidence to Sites critique
Sites visual review currently asks whether the design extends beyond the release cover without supplying a cover-suppressed render. This adds bounded supplemental mobile/desktop captures to the existing review path and existing metered critique call, with explicit image indices, canonical candidate/world hashes, base-revision scope, qualitative reasons and targeted improvements.
Only selected Context artwork already present as a candidate image asset can be masked. Browser masks cover exact-source inline image bounding boxes without changing the site; rectangular overlap, unsupported representations and fresh-initial-state scope are disclosed. Missing or changed captures remain unverified. Structured cover-hidden issues are gated before repair admission; ordinary visual defects and objective runtime failures remain actionable. Model evidence classification remains an assessment, not deterministic proof.
Normal journey/opening images and the original byte-pinned journey runner and render preparation are unchanged. Supplemental capture uses the same Sandbox/lifetime, with a 30-second cap per invocation and no internal retry or extra model call. The existing single journey correction can invoke it again. Publication algorithms, database sources, repair limits, credits and cancellation remain unchanged. The offline bridge explicitly repins only the optional TypeScript review metadata addition and must qualify anew.
Validation at 2f28d70b57acc93d5ce0e7c8e4555f5dd3c45ed7: 483 Sites tests pass; 4 live tests deliberately skipped. Synthetic tests cover actual workflow/render plumbing, iframe input, mask identity, ancestor opacity/depth, changing DOM bounds, long signed URLs, canonical hashes, image ordering, both journey attempts and repair evidence gates. Scoped types and lint pass. The production type check has five diagnostics identical to untouched base 30c4d442 under the same local dependency setup. Independent spec and standards reviews passed the frozen candidate and predecessor deltas. Test-only successor adds schema-valid ordinary/mixed issue fixtures. All eight exact-head CI checks passed; Cubic completed with zero issues in the final test delta and the human-review condition below; Codex review completed with no findings, both independent reviews passed, and all 16 review threads are resolved. CodeRabbit retains an advisory maintainability warning with a documented disposition; it is not represented as a passed check.
Implements the bounded rendered-cover critique slice of recoupable/app#2094, not the complete issue. Source-only qualification: no live browser, provider, Vercel Sandbox, account or production operation ran. Native/rendered acceptance remains separate. Coordinator owns final merge.
Release qualification: Cubic's exact-head review reports zero issues in the final test delta but explicitly “Requires human review”: rollout and evidence-gating tradeoffs need human sign-off. Automated review completion is not rollout approval. The coordinator retains the existing final merge/release decision; this PR does not record that decision or introduce an artist-facing approval form. The tradeoffs are fresh initial-state rectangular masks (including overlapping pixels), unsupported captures remaining unverified, and model-assigned issue provenance feeding the existing bounded repair path only when its evidence gate admits it. Native/rendered acceptance remains open: #2094 still requires contrasting-artwork mobile/desktop renders and interaction evidence tied to the saved brief/revision. No new source defect was identified by this review condition; source qualification alone does not complete #2094 or establish customer-facing release readiness.
Summary by cubic
Adds bounded cover-suppressed mobile/desktop captures to the Sites visual review so the existing critique can verify the design extends beyond the release cover instead of judging it without a cover-masked render. Implements a slice of recoupable/app#2094, not the complete issue.
- Only an existing selected cover URL that is already a candidate image asset can be masked; source-exact browser masks resist author CSS without changing the site.
- Rectangular overlap and unsupported representations (CSS backgrounds, picture/srcset, canvas, transparent or unbounded ancestors, later states, dynamic DOM growth) are disclosed; captures are rejected when the mask target set fails to match the inspected visible covers or geometry changes during capture, so missing, changed, or failed captures stay unverified.
- Unverified cover evidence cannot promote a cover-specific repair: a needs-improvement critique triggers
reviseonly when all viewports captured, structured cover-only issues are discarded when evidence is missing, and a model-claimed needs-improvement with discarded evidence is forced topass, never sent to the repair loop. - Supplemental images are appended to the existing metered critique call with explicit image indices, canonicalized candidate/selected-world hashes, and base-revision scope — no extra model call, no retry, 30-second hard cap. Runner inputs, collected reports, and tagged review issues (ordinary or cover-hidden) are schema-checked against explicit bounds.
- Normal journey/opening evidence, the byte-pinned journey runner, and existing repair, publication, and cancellation behavior are unchanged.