[pbakaus]

Round 13 of the detector corpus, branch F: the evidence a URL scan hands a reviewer. Nothing here changes what the engine detects. It changes the pictures and anchors that go with the findings, so a judge sees what the visitor sees.

Row 21: a fixed sheet over every element shot

On theguardian.com's phone pages, a fixed "This is the catch" support sheet covers the lower two thirds of the viewport. Every element shot past the screenshot's cut showed the sheet instead of the flagged element: 9 of 9 tiles on the international front, all 27 on an article.

Element shots now hide the fixed layers painted over the flagged element, for that shot only, and put them back afterwards. At 9 points over the element's visible part the engine walks document.elementsFromPoint down to the element, and hides each position: fixed layer it passes. Guards:

• A point counts only when the walk reaches the element or something inside it. A layer painted under the element (a fixed backdrop, a WebGL canvas behind content) is never hidden.
• The element is made hit-testable for the walk (pointer-events: auto, restored right after), so content that takes no pointer events, such as card text under a link stretched over the card, is listed at its true paint position.
• Never the layer that holds the element, and never a fixed box under a transformed, filtered or contained ancestor (it scrolls with the page, the same containment list stitch uses).
• An open dialog or an aria-modal layer stays in the shot, and so does a fixed backdrop that wraps one: the visitor sees it.
• A tile is reused for a later element only when no layer it left showing covers that element.

Each shot records what was hidden in ElementShot::hidden_covers, so a judge can tell "chrome removed for this shot" from "what the visitor sees".

Live, on the revised head: theguardian.com/international on a phone, 32 of 32 element shots hide the sheet and no tile shows it (base: every tile showed it). The article on desktop: 15 of 15. The article on a phone: 61 of 61. The international front on desktop has no fixed sheet and records none. The hit-testable step was added after review: without it, 6 of the 32 phone shots (card labels under a stretched link) kept the sheet on 5 of 11 tiles.

Row 23: a page that re-lays itself for the beyond-viewport capture

developer.apple.com's guideline pages answer the capture's resize (Chromium dispatches it at 1x1, then at the real size) by folding the navigator and sliding the column back over about 300ms, and the capture paints that state: the column at x 309 while the rects say 390.

The engine now measures a sample of layout boxes (a 6x6 grid of elements in the first viewport, offset geometry, so transforms do not count) before and right after the capture. When at least 2 boxes and a quarter of the sample moved sideways, it waits up to 2 seconds for the boxes that moved to come back to where they were measured, twice in a row 100ms apart, and then stitches the page from viewport tiles. A page that keeps moving (a marquee on margin-left) never settles and keeps the beyond-viewport capture, and so does any page where a box that moved at the capture keeps its new geometry. Vertical moves are ignored: an image loading during the capture pushes content down without a relayout.

This does not fix developer.apple.com itself. On the live page five navigator link labels change width at the capture (96px to 84px and similar) and keep their new widths, while the column comes back within 100ms. The engine cannot tell whether the image or the rects hold those labels' true place, so the page fails safe to the beyond-viewport capture, as on main, after a 2 second wait. Widening the rule to let it through would also let through pages whose picture really is wrong. The fallback is proven by a browser fixture that folds on resize and slides back; the never-ending ticker and the slide-only page are its twins.

The stitched fallback also needed a fix to stitch: every tile after the first re-tests sticky boxes and hides one stuck at its own top offset once its own place in the document is above that tile, so an earlier tile already painted it. A section bar is no longer repeated at the top of every tile, and a bar whose section starts on the current tile (or exactly on its top edge) stays.

Row 24: page-level anchors for organic clip paths and bounce easing

page_form_anchor now anchors organic-clip-path (an element whose computed clipPath is a polygon passing the scan's own off-grid and not-rectilinear tests, through a shared css_scan::organic_polygon_vertices) and bounce-easing (a computed timing function with the same bezier, or an animation the snippet names). None of the corpus's listed findings gain an anchor from it today: apple.com's polygon is on a media-controls pseudo-element, and framer's bounce curve is declared but computed by nothing at capture. A local page with all three page forms gets anchors for the halo and the polygon.

Corpus numbers

Replays on main's runs, against main's replay of the same run (c51e529):

Run	Captures	Violations (branch / main)	Rule table and every other field
1009 (craft)	132	0 / 0	identical
1006 (corpus)	243	16 / 16	identical
1004 (corpus)	177	8 / 8	identical
47 (corpus)	226	14 / 14	identical

The violations are main's own against those runs' judged labels; this branch adds none and removes none.

No craft mark is removed on run 1009 (legit 0, harmless 0, false positive 0, unclear 0). Replays exercise detection, not the evidence step, so identical tables are the expected result; the evidence changes were checked live on the pages above.

Tests

• Browser tests in crates/browser/tests/fullpage_screenshot.rs: the sheet is hidden (an_element_shot_hides_the_fixed_sheet_over_the_element, ..._over_content_the_hit_test_misses), with twins that keep a fixed backdrop, a backdrop under content that takes no pointer events, a fixed box that scrolls along, an open modal, a modal wrapped in a fixed backdrop, and a shared tile only for elements no shown layer covers. The relayout fixture is stitched, with sticky bars that stick on a later tile, stick on the tile their section starts on, and start on a tile's top edge; a ticker that never stops and a page that only slides keep the beyond-viewport capture. Assertions read solid colour blocks, never text pixels.
• Unit tests for layout_moved, moved_boxes and settled, including the apple shape (labels that move at the capture and stay keep the page unsettled), and for the anchors, with a zigzag polygon decoy.
• No oracle case changed. No text or static-HTML check changed, so no hook cost is owed.

Gates on the merged head 1437fd2 (origin/main 5cb95ee merged in, no conflicts), before the review-round commit 8772352, which touches only fullpage.rs and its browser tests (all 21 pass, and the bundle check holds): cargo build --release -p impeccable, cargo xtask bundle --check (both live assets up to date), the node oracle (no case changed, no re-record, no DELTAS entry), cargo test --workspace (2257 passed, 0 failed), bun run build with no tracked changes, and bun run test with the branch binary.

AI assistance: prepared by Claude Code, under instructions from @pbakaus.

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes screenshot capture timing (up to 2s layout wait), in-page DOM opacity during shots, and new evidence fields; detection tables are unchanged but evidence output behavior shifts on real sites.

Overview
Improves evidence screenshots without changing detection rules: reviewers get clearer full-page and per-element images, plus better anchors for some page-level CSS findings.

Element viewport shots now temporarily hide position: fixed layers that paint over the flagged element (support appeals, cookie sheets), using a 9-point elementsFromPoint walk with guards for backdrops, modals, transformed “fixed” content, and pointer-events edge cases. Shots record what was hidden in ElementShot::hidden_covers, and tile reuse only applies when the previous capture left no covering layer.

Full-page capture probes layout before/after a beyond-viewport screenshot; when enough boxes shift horizontally and then settle back, it stitches viewport tiles instead of keeping a capture painted during a resize-driven relayout (e.g. folding side nav). Continuous motion or permanent geometry changes still keep beyond-viewport. Stitching also hides sticky bars that were already painted on earlier tiles.

Evidence anchoring extends page_form_anchor to organic-clip-path and bounce-easing via shared organic_polygon_vertices in CSS scan logic.

Browser and unit tests cover sheets, modals, shared tiles, relayout/sticky stitching, and layout probe helpers.

^{Reviewed by Cursor Bugbot for commit 0d0276f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[impeccable-reviews]

5/5 · Ready: no major issues found
_{This push leaves nothing to fix.}

Nothing to fix.

About this review

• _{crates/live/assets/detect-antipatterns-browser.js is too large to read.}
• _{Read from the code alone; no page was opened.}

_{Reviewed 1 file on 23 Oct 2026, 18:25 UTC · Run page}

_{Reviews (4) · Last reviewed commit: Merge origin/main into corpus/round13-harness}

[cursor]

Stale Bugbot comment from a previous run.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The PR changes evidence capture for URL scans, not the findings themselves.

Fix cover removal for sheets with opacity transitions before merging.

Findings

1.  Fading sheets remain in screenshots ▶

Summary

Improves screenshot evidence and adds page-level anchors without changing detection rules.

• Element shots show the flagged element instead of fixed layers painted over it.
• Full-page captures stitch pages that settle after a capture resize.
• Page-level anchors point to elements carrying organic clips and bounce easing.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Capture full page] --> B{Layout moved and returned?}
  B -->|Yes| C[Stitch viewport tiles]
  B -->|No| D[Keep beyond-viewport image]
  C --> E[Find flagged elements outside image]
  D --> E
  E --> F[Scroll element into view]
  F --> G[Hide fixed covers]
  G --> H[Capture element picture]
  H --> I[Restore covers and record hidden layers]

Loading

_{Reviews (4) · Last reviewed commit: "Merge origin/main into corpus/round13-ha..." · Reviewed by Greptile}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 0d0276f. Configure here.}

[linear-code]

REN-1212