[pbakaus]

Fixes the contrast residue from corpus observations-1004 branch 3 and the contrast misses in walkthroughs-1004.

URL engine:

• When the pixel pass refuses copy because hiding it repainted most of the box (a wide dark halo under light copy over a photo), it takes a second reading with the text-shadow kept. It does this only when every shadow layer stands apart from the ink (2:1 or more over the ink colour) and the text is not outlined. Its first reading still hides the shadow, so a glow in the ink's own colour or a drop shadow under grey copy reads as before.
• A backdrop-filter under a fill at least 0.8 opaque no longer blocks the pixel pass (new reason backdrop filter under a near-opaque fill). The sampled pass still refuses it.
• The hit-test stacks treat a fill at least 0.95 opaque as the surface the text sits on, so nav links on a 0.95 white pill are scored.
• On the link and span path, a same-colour pair reports 1.0:1 only for the whole visible label of a link or button on a box of the page's own (a white "Start Chat" on its white card). A hidden label beside an icon, a span in no control, and the page ground are still dropped. Styled controls and the static engine are unchanged.
• A box a timed one-shot entrance holds at opacity 0 at capture counts at the entrance's end frame.
• A box a paused loop holds part way through is not at rest when the loop fades to 0.1 or below. A paused breathing loop that never vanishes is read at its frame. That split is a judgment.
• The site's own name set huge and faint is a decorative wordmark and reports as advisory (taste call r3-04). It applies at 120px, or at 3x the h1 and at least 80px, with 0.25 or less of its ink showing, and its words must match the host, og:site_name, application-name, or the title's first or last part.

Corpus:

• No new ratchet violations on runs 1004, 47, 42 and 35. The ratchet replays the element pass only, so it does not cover the halo and backdrop changes.
• On run 1004 the element pass adds 16 findings and moves 9 ghost wordmarks to advisory; run 35 moves 6 more. The 16 are missed nav links, two white-on-white buttons, and an entrance-held CTA reported under three rules.
• A side-by-side live recapture covered the 7 evidence sites plus 10 sites with text-shadow glows and halos. The pixel side gains counted verdicts where the pass used to refuse or block: smartbi.ca 11, imageto3d.studio 2, a grey count badge on fastsocial.co, klimaschutz-niedersachsen.de 3.4:1, rhem.ai 3.0:1. It loses iubcr.vercel.app's four paused-loop readings.
• Nothing new fails on the glow and halo sites. Every other difference reproduces with both engines on the same capture.
• Known leftover: an incoming crossfade slide caught at opacity 0 is now counted while the outgoing one still paints.

Tests: unit tests per part, two browser fixtures with should-flag and should-pass twins (including a neon glow, a grey drop shadow, a hidden menu label and an "opens in a new tab" span), new oracle goldens, cargo test --workspace, bun run test, bun run build. The in-page bundle is regenerated.

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

🤖 Generated with Claude Code

https://claude.ai/code/session_01FTbrWaYuBu9QB8gE99M2xi

---

Note

Medium Risk
Touches core contrast routing, pixel screenshot logic, and opacity/animation heuristics; corpus behavior shifts (e.g. paused-loop findings dropped, new pixel verdicts) despite broad test coverage.

Overview
Improves low-contrast detection for halos, frosted glass, animation timing, same-color controls, and faint footer wordmarks.

The pixel contrast pass can retry after a “box repainted” refusal: it hides glyphs but keeps text-shadow when the core marks a separating halo (haloSeparates), not outlined text. backdrop-filter under fills ≥0.8 opaque gets a dedicated reason so pixels can measure it while the sampled pass still refuses. Candidates gain haloSeparates via new halo_separates_ink (each shadow layer ≥2:1 from ink).

Element-pass contrast uses settled opacity for timed entrances caught at 0 (entrance_end_opacity in folds), treats paused infinite loops that fade to nothing as mid-motion (paused_mid_loop), and picks ≥0.95 opaque overlays as the hit-test surface host for links on glass pills. On link/span safe-tag paths, ink-matching-background 1.0:1 only applies when the text is the sole visible label of a control on the page’s own surface; hidden menu text and decorative spans stay suppressed. The browser safe-tag guard now respects same_color_surface_is_unread.

Decorative text adds ghost wordmarks: huge faint site-name type (size/opacity thresholds, host/title/meta matching) → advisory, with browser-side settled_opacity so mid-entrance captures are not misclassified.

New browser fixture tests cover halo/glass pixel reads and settled-state contrast; extensive unit tests accompany each behavior.

^{Reviewed by Cursor Bugbot for commit 8b76f98. Bugbot is set up for automated code reviews on this repo. Configure here.}

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the numbered previous finding is fixed, and no new actionable issue was established.

Summary

Improves browser contrast checks for halos, near-opaque glass, hidden labels, and settled animation states.

• The screenshot pass can score text over separating halos and near-opaque glass.
• Browser checks score same-color text when it labels a control on its own surface.
• Browser contrast checks use the opacity a timed entrance settles at.
• Faint oversized versions of a site’s name count as decorative wordmarks.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Browser text] --> B[Read surface and settled opacity]
  B --> C[Element contrast check]
  B --> D[Uncertain background]
  D --> E[Compare screenshots with text hidden]
  E --> F{Separating halo and box repainted?}
  F -->|Yes| G[Retry with halo kept]
  F -->|No| H[Pixel result]
  G --> H
  C --> I[Warning or decorative advisory]
  H --> I

Loading

_{Reviews (9) · Last reviewed commit: "Judge a ghost wordmark at the opacity it..." · Reviewed by Greptile}

[Copilot]

🟡 Changes recommended

Incorrect contrast edge cases, repeated document traversal, and stale aggregate oracle expectations remain unresolved.

6 open findings

• Add new fixtures to aggregate oracle goldens · New
• Use settled opacity for decorative text detection · New
• Avoid repeated full-document heading queries · New
• Exclude hidden or zero-area media from visible paint checks · New
• Inspect descendant decorations when the control is the label · New
• Apply entrance end opacity during opacity folding · New

What changed in this PR

Improves browser contrast detection for shadowed text, translucent surfaces, animation states, and decorative wordmarks.

Changes:

• Adds halo-preserving pixel retries and near-opaque glass support.
• Refines surface, control-label, animation, and wordmark classification.
• Adds regression fixtures, browser tests, oracle expectations, and contract documentation.

File	Description
tests/​oracle/​golden/​detect-fixture-text-low-contrast-settled-states-html.json	Records settled-state text output.
tests/​oracle/​golden/​detect-fixture-text-low-contrast-pixel-reads-html.json	Records pixel-fixture text output.
tests/​oracle/​golden/​detect-fixture-json-low-contrast-settled-states-html.json	Records settled-state JSON output.
tests/​oracle/​golden/​detect-fixture-json-low-contrast-pixel-reads-html.json	Records pixel-fixture JSON output.
tests/​fixtures/​antipatterns/​low-contrast-settled-states.html	Adds animation, label, and wordmark cases.
tests/​fixtures/​antipatterns/​low-contrast-pixel-reads.html	Adds halo, glow, and glass cases.
docs/​CLI-CONTRACT.md	Documents revised contrast behavior.
crates/​html/​src/​decorative_text.rs	Keeps layout-dependent wordmarks browser-only.
crates/​core/​src/​checks/​rules.rs	Refines same-colour safe-tag scoring.
crates/​core/​src/​checks/​decorative_text.rs	Adds decorative wordmark classification.
crates/​core/​src/​browser/​visual.rs	Adds halo eligibility and glass routing.
crates/​core/​src/​browser/​painted.rs	Recognizes entrance endpoints and paused loops.
crates/​core/​src/​browser/​element_checks.rs	Updates surfaces, labels, and opacity handling.
crates/​core/​src/​browser/​decorative_text.rs	Collects browser wordmark evidence.
crates/​browser/​tests/​low_contrast_settled_states.rs	Tests settled-state contrast behavior.
crates/​browser/​tests/​low_contrast_pixel_reads.rs	Tests halo and glass pixel readings.
crates/​browser/​src/​screenshot_contrast.rs	Retries eligible captures with halos preserved.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cursor]

Stale Bugbot comment from a previous run.

[impeccable-reviews]

Nothing to fix.

5/5 · Ready: no major issues found
_{This push leaves 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 19 Oct 2026, 00:46 UTC · Run page}

_{Reviews (7) · Last reviewed commit: Judge a ghost wordmark at the opacity it settles at}

[cursor]

Stale Bugbot comment from a previous run.

[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 8b7434e. Configure here.}

[linear-code]

REN-905