[pbakaus]

Round 13 of the detector corpus, branch E: low-contrast on text a visitor sees in another state. Under taste call r13-t1, a stepper item dimmed beside the active one is advisory; text a script parked at the start of a reveal is not at rest and is scored at its declared colour.

What the row was

On the craft run (1009), low-contrast scored four shapes of dimmed text as if the page sat that way:

• arc.net's feature stepper: every article.feature sits at opacity 0.3 by stylesheet with transition: opacity .3s, and the active one, article.feature.is-active, is at 1. The visitor scrolls the active slot along, so each item is read lit. 21 findings, all marked false positive.
• ramp.com: a card's copy held at inline opacity: max(.2, var(--card-animation-1)), a value a scroll script scrubs from 0.2 upward. The box is a sole child. 4 findings, all false positive.
• raycast.com: every key of a keyboard parked at inline opacity: 0.2 with an opacity transition; the same component further down the page carries one more class (glowing) at 1. 10 findings in the pixel pass.
• framer.com: a card parked at its start values, inline will-change: transform; opacity: 0.2; transform: translateX(-19.99px). 2 findings in the pixel pass.

The fix

All in the shared element colour pass, crates/core/src/browser/element_checks.rs.

A state (arc, advisory). dimmed_by_a_state: a box between the text and its surface below opacity 0.5, with an opacity transition, beside a same-tag sibling at opacity 1 whose class list is the dimmed box's plus exactly one token. The finding keeps its place and becomes advisory, like the decorative path. The pixel pass marks such an item's candidate advisory the same way, so a stepper over a photo stays advisory there too. It stands down when the dimmed box is a control or the text reaches it through one: button, field, tab, a[href], aria-selected, link/option/menu/tree roles, and a short box whose computed cursor is pointer where its parent's is not.

A reveal (ramp, raycast, framer; not at rest). parked_inline_for_a_reveal, read where the pass decides whether a fade is at rest. The box needs an inline opacity, and one of:

• framer: inline transform values that displace the box, plus a will-change naming opacity or a transform. All four signals together; a stylesheet value, opacity alone, or no displacement stays at rest.
• ramp: opacity below 0.5, an opacity transition, and a scrubbed value: var( inside max(, min(, clamp( or calc(, or a custom property set in a style attribute on the box or an ancestor.
• raycast: opacity below 0.5, an opacity transition, at least three boxes of its tag and class list parked alike at the same inline opacity, and a peer elsewhere on the page at 1 whose class list is theirs plus one token, also with an opacity transition.

The ramp and raycast arms stand down when the box is a control, sits in one or holds one, including a clickable child with a short label and no role. The style attribute is read for the declaration that wins (the last, unless an earlier one is !important), and custom property names match case-sensitively.

An inline-parked box with a bright sibling is a state, not a reveal: it goes advisory, never dropped.

The snapshot capture now reads computed cursor (STYLE_PROPS, docs/CLI-CONTRACT.md). Stored corpus captures predate that and read it as empty, so the replay cannot exercise the cursor guard; the fixture and browser test do.

The static engine reads no transition, transform or will-change, so its output does not change. The new fixture's static goldens pin that.

Accepted as advisory

Paul accepted two shapes that match arc's exactly and become advisory with it: a checkout stepper's future steps beside the current one, and a footer list whose current item is at 1 and the rest at 0.4 with a hover transition. Nothing in a snapshot separates "the visitor moves the active slot" from "this is how the page sits". The fixture pins the checkout stepper as advisory. Footer links all faded alike, or list items with no class, still warn.

Run 1009 (craft), against main's replay

Only low-contrast moves: 1112 on main, 1108 here. 4 removed, 0 added, 21 now advisory.

Verdict	Removed	Now advisory
Legit	0	0
Harmless	0	0
False positive	4 (ramp 10134263, 10134264, 10134402, 10134403)	21 (arc 10133793-99, 10133824-30, 10133857-63)
Unclear	0	0

No legit mark is removed or demoted. arc.net's mobile captures do not move: that grey timeline copy is at full opacity.

framer (10133521-22) and raycast (10134381-84, 10134970-75) are pixel-pass findings, which a replay cannot run. On the stored snapshots both are caught mid reveal, so the pixel pass skips them: raycast's parked keyboard (76 of 152 desktop primaries, which holds every finding rect, and 4 of 4 on mobile) and both framer headings. They will show on the next cloud craft run.

Corpus runs, against main's replay

Run	Captures	Violations (main / branch)	Demoted harmful (main / branch)	Rule-line changes
1006	243	16 / 16	0 / 0	none
1004	177	8 / 8	0 / 0	none
47	226	14 / 14	2 / 2	none

The violations are main's own (replay against recorded), with identical id sets on both sides, so this branch loses or demotes no finding both judges called harmful. The first version demoted partner.co.il's inactive tab buttons on 1006 (10089911, 10089926, both harmful); the control guard keeps them, and 1006 now matches main.

Hook cost

The new tests sit in the shared element colour pass, which the static HTML engine also runs (color-rules). rule_cost --runs 9 over main's tests/fixtures/antipatterns plus the impeccable.style site source (803 inputs, cold per edit), main (c51e529) and this branch built from their own trees, in the order main, branch, branch, main:

Run	static-html per edit, mean / p50 / p95 ms	color-rules, mean / p50 / p95 ms
main, 1st	36.62 / 31.91 / 72.55	3.977 / 3.005 / 9.747
branch, 2nd	31.27 / 30.36 / 62.03	3.602 / 2.852 / 9.035
branch, 3rd	32.86 / 30.28 / 66.19	3.749 / 2.857 / 9.497
main, 4th	31.34 / 30.35 / 61.60	3.575 / 2.839 / 9.028

No difference above run-to-run noise on a loaded machine. On the hook path the reveal test reads the inline style attribute and stops at the transition check, which the static engine never passes (it reads no transitions); the state test runs only for an element that already failed contrast. The text engine is untouched.

Tests

• Fixture tests/fixtures/antipatterns/low-contrast-resting-states.html. Should flag: the arc stepper and a checkout stepper (both advisory), footer links faded alike, tabs with and without a transition, an inline number, a lone parked badge with a lit namesake in another section, parked nav links, an inline design token (opacity: var(--muted)), a faint modifier on the dimmed box, a clickable plan option with no role, identical class lists. Should pass (URL engine): the framer, ramp and raycast shapes.
• Unit tests in element_checks.rs for each arm and its twins, and crates/browser/tests/transient_opacity.rs, which asserts on computed colours and the advisory flag only, so it does not depend on host fonts.
• Oracle: new cases detect-fixture-json-low-contrast-resting-states-html and detect-fixture-text-low-contrast-resting-states-html (static engine: fourteen warnings). No existing golden moved; the directory sweeps gain the same findings but are accepted deltas and not compared (DELTAS entry).
• Review round (d39ef14): twins for the pixel-pass advisory, a clickable child, declaration order and !important, and custom property case. Runs 1009, 1006, 1004 and 47 re-measured on this head: identical to the numbers above.
• On the merged head (05a180e, main ea69c18 merged in): cargo build --release -p impeccable, cargo xtask bundle --check, node oracle, cargo test --workspace, bun run build (no tracked changes), bun run test with this binary. All green.

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

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes when low-contrast fires and its severity across element and pixel passes; behavior is nuanced but heavily fixture-tested.

Overview
Refines low-contrast so dimmed text is not always scored as if the page permanently sits that way.

Resting UI state (advisory): Text behind a faded ancestor that has an opacity transition and a same-component sibling at full opacity (e.g. inactive stepper/tab beside the active one) still reports low-contrast but is advisory, with the same treatment in the pixel pass. Controls on the path (tabs, links, cursor: pointer chips, etc.) keep full severity.

Parked reveals (not at rest): Fades driven by inline style at animation start are excluded from “at rest” blending—Framer-style transform+will-change, scroll-scrubbed opacity: max(.2, var(...)), and groups of keys parked alike when a lit peer exists elsewhere on the page.

Snapshot capture now records computed cursor so pointer-based control detection works in the browser engine. Fixture, unit, and URL-engine tests cover the matrix.

^{Reviewed by Cursor Bugbot for commit 42bd9dc. 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

• _{Not sent. low-contrast detector pass isn’t a page to walk.}
• _{Design system. This pull request adds --muted, --Muted, --delay; it is judged against the system as it leaves it.}
• _{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, 22:05 UTC · Run page}

_{Reviews (10) · Last reviewed commit: Merge origin/main into corpus/round13-resting}

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The change reaches browser-based low-contrast findings across scanned pages.

Fix the shadow-tree control check before merging so slotted button labels keep their contrast warnings.

Findings

1.  Slotted labels lose contrast warnings ▶

Summary

This PR makes dimmed stepper text advisory and treats supported inline reveal shapes as temporary states.

• Dimmed stepper text keeps its finding but becomes advisory.
• Text in a parked reveal is scored at its declared color.

Earlier fixes cover pointer-only children, visual advice, declaration order and custom-property case, summary, menuitemcheckbox, and hidden bright siblings. pbakaus accepted the repeated page-walk cost because the checks narrowly limit when it runs and the static hook never reaches it. pbakaus also accepted deferring the control-above-faded-box gap because no current corpus or craft replay contains that shape.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Text under a faded box] --> B{Recognized inline reveal?}
  B -->|Yes| C[Ignore fade in colour check]
  B -->|Yes| D[Skip visual check]
  B -->|No| E{Matching bright sibling and no control?}
  E -->|Yes| F[Keep contrast finding as advice]
  E -->|No| G[Keep normal contrast warning]

Loading

_{Reviews (10) · Last reviewed commit: "Merge origin/main into corpus/round13-re..." · 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 30b0e03. Configure here.}

[linear-code]

REN-1219