✦ Audit & remediation record

The perfect score was hiding four Level A failures.

Before this audit, automated tooling said naxatar.com was in fine shape: Lighthouse averaged 97.2 across seventeen routes and gave the homepage — the most complex page on the site — a clean 100/100, with zero axe-core violations. That homepage carried four Level A failures. This is the record of what was actually wrong, how it was found, and the evidence that it is fixed — on the live site, not a local build.

Before → after

axe-core violation nodes: ~~214~~ 0 17 routes × 2 themes = 34 scans · 0 on production too
Lighthouse accessibility, mean: ~~97.2~~ 100.0 all 17 routes at 100
Success criteria failing: 11 0 16 defects total — some criteria failed in more than one place
Homepage tab stops on invisible content: ~~18 / 35~~ 0 / 36 focus landed on content painted at opacity 0
Controls hidden behind the fixed header: 40 0 measured over 181 controls on 6 pages
Looping animations after “background off”: 39 0 the toggle previously stopped only the WebGL layer
Pages with a working skip link: ~~0 / 17~~ 16 / 17 the standalone visual essay has no shared layout
Live regions announcing async results: 1 7 form results previously changed silently

A keyboard user could black out the homepage.

The homepage is a locked full-page deck: the document never scrolls, the deck slides. Nothing in it followed keyboard focus. Pressing Tab moved focus into slides the pager had never revealed — their content sat at opacity: 0 — while the browser silently scrolled the pager's shell as far as 3,482 pixels to compensate. The visitor ended up typing into an invisible contact form over an empty starfield, and one arrow key later the page showed nothing at all.

No scanner can see this. It is not a missing attribute or a bad colour — it is an interaction between a locked pager, a CSS reveal, and the browser's scroll-into-view behaviour. Finding it took tabbing the page and asking a question no tool asks: is the thing I just focused actually being painted?

The naxatar homepage showing only a starfield and the header — no visible content — while keyboard focus is inside the contact form — Focus is in the contact form's Name field on both screenshots. Before, the form is invisible and the position readout still claims slide 1 of 5. After, the deck follows focus: the form is painted and every indicator agrees.

The same keyboard position after the fix: the contact slide is displayed with the form fully visible and the position readout showing slide 5 of 5 — Focus is in the contact form's Name field on both screenshots. Before, the form is invisible and the position readout still claims slide 1 of 5. After, the deck follows focus: the form is painted and every indicator agrees.

The homepage rendered as a completely blank starfield with no content, after a single arrow-key press — One arrow key later. Before, the deck advanced while the shell stayed scrolled — composing to a position past the end of all content: a blank page. After, the pager holds its place.

The homepage after the fix showing the contact slide fully rendered following the same keystrokes — One arrow key later. Before, the deck advanced while the shell stayed scrolled — composing to a position past the end of all content: a blank page. After, the pager holds its place.

The same defect had a second entrance: VoiceOver's reading cursor moves without moving DOM focus, so the first fix — following focus — didn't cover it, and a live VoiceOver pass caught the screen going blank while the speech stayed correct. The durable rule was the invariant, not the event: the pager's shell is overflow: hidden, so any scroll of it is illegitimate, whoever caused it — and is now converted into real navigation.

Sixteen defects, eleven criteria — the scanners caught three.

axe found two of the eleven failing criteria and Lighthouse a third. The rest came from keyboard traversal, computing ratios from the design tokens, applying WCAG's text-spacing metrics, hover-content checks, and a screen reader. All sixteen are closed.

Criterion	Level	Found by	The defect
2.1.1 Keyboard	A	Manual	The homepage deck ignored keyboard focus — 18 of 35 tab stops put focus on unpainted content, including every contact-form field
2.4.3 Focus Order	A	Manual	The page’s position indicator read “slide 1 of 5” while focus sat in the form on slide 5
2.4.7 Focus Visible	AA	Manual	An indicator cannot be visible when the control it surrounds is painted at opacity 0
2.1.1 Keyboard	A	axe	A screenshot strip with nothing focusable inside — 4 of its 5 images unreachable without a mouse
2.2.2 Pause, Stop, Hide	A	Manual	39 infinitely looping animations; the one background control stopped only the WebGL galaxy
2.4.1 Bypass Blocks	A	Manual	No skip link on any page; nine header controls before the content, every time
1.4.1 Use of Color	A	Manual	A link inside prose distinguished from its surrounding text by lightness alone — 2.92:1, under the 3:1 required
1.3.1 Info and Relationships	A	Lighthouse	Heading levels jumped h1 → h3 on both product pages; one page had no heading at all
1.4.3 Contrast (Minimum)	AA	axe	214 failing nodes across 9 routes, from two colour tokens — the site’s quiet grey measured 2.24:1
1.4.11 Non-text Contrast	AA	Manual	Every form field’s only visual boundary was a 1.32:1 hairline
1.4.12 Text Spacing	AA	Manual	Applying WCAG’s own text metrics cut 55px of content away inside an overflow:hidden slide, unreachably
1.4.13 Content on Hover or Focus	AA	Manual	The hover tooltip was neither dismissable (Escape did nothing) nor hoverable (pointer-events: none), and obscured text
2.4.11 Focus Not Obscured	AA · new in 2.2	Manual	40 controls came to rest entirely behind the fixed header after anchor jumps
3.1.2 Language of Parts	AA	Manual	Devanagari (नक्षत्र) in nine places inside lang="en" — pronounced with English rules
4.1.2 Name, Role, Value	A	Manual	A role="radiogroup" where each radio was its own tab stop and the arrow keys did nothing
4.1.3 Status Messages	AA	Manual	One live region on the whole site; contact-form results announced nothing to a screen reader

Why the tools couldn't see it — in both directions.

On the homepage, axe returned fourteen contrast checks as “incomplete” — neither pass nor fail — because it cannot resolve a backdrop made of translucent glass over a WebGL canvas: “element's background color could not be determined because element contains an image node.” Those fourteen were settled by computing ratios directly from the design tokens. The two scanners also disagreed with each other: Lighthouse reported contrast failures on two routes where axe recorded none, because a different viewport renders a different set of elements.

Careful measurement cuts the other way too. A size-threshold reading of target size (2.5.8) flags 21 failures on this site. Applying the criterion's actual spacing and inline exceptions — computing whether a 24-pixel circle on one target's centre reaches another's — gives zero. All 21 would have been false findings in the conformance report.

Fixed without touching the design.

The remediation is +260/−48 lines across 22 files — no design-system refactor, no new UI. Every colour correction was solved for the smallest change that crosses the WCAG threshold, hue held, so the site's quiet visual register survived. The pair below carries a 2.24:1 → 4.85:1 correction on the eyebrow, the back link and the closing line. Nothing else moved.

The support page before remediation, its small uppercase labels rendered in a grey that fails contrast requirements

The support page after remediation, visually identical except the small labels are now a legible darker grey

Verified where a stranger can check it.

Re-audited against the deployed production site, not a local build: 0 axe violations across 34 scans, 0 unpainted focus stops, 0 controls behind the header.
Repeated in all three browser engines a screen reader can sit on — Chromium, WebKit and Gecko: the exposed role, name and state of 21 controls agreed across all three, with the same live regions and identical keyboard behaviour from the custom radio group.
Screen-reader testing: VoiceOver on macOS, which found a defect nothing else had. NVDA and JAWS were not tested — the conformance report says so in those words rather than implying broader coverage.
Three of the fixes were themselves defective on the first attempt — a focus ring that would have squared the pill buttons, a skip link revealed on the wrong pseudo-class, and the VoiceOver path above. Each was caught by re-verification and is documented, because remediation that audits itself is the difference between a fix and a claim.
The whole harness — axe and Lighthouse runners plus purpose-written probes for keyboard traversal, target-size exceptions, focus obscuring, motion, and text spacing — is published in the site's repository, so every number above can be reproduced.

✦ Full detail: the conformance report · the accessibility statement · audited August 2026 by the site's developer