fix(theme): stop the dark scheme reading as a stock generated landing page

[sanity]

Ian flagged that freenet.org looked like cliché AI-generated design ("the glowing colors"). In light mode that turned out to be mostly wrong; in dark mode it was right, and this fixes it.

Root cause

The dark palette was a mechanical lightness inversion of the light one. Hue and saturation were preserved and only lightness flipped:

token	light	dark (before)
--text	#404654	#abb1bf
--surface	#ffffff	#14161a
heading accent	#0066CC	#4da6ff

That lands on a near-black ground (#121212) carrying a fully saturated #4da6ff on every card heading. Add a gradient-filled headline and a blue glow under the primary button and you have the recognisable generated-template look. A designed dark theme desaturates its accents and lifts the ground off pure black; a flipped one does neither.

Changes

• Accent #4da6ff → #79a7d6, chosen for the same relative luminance as the old value, so every contrast pair involving it is unchanged (~7.1:1 on the ground) while chroma is roughly halved. The neon read came from saturation, not brightness, so this is a chroma cut rather than a dimming.
• Ground #121212 → #15171b — off pure near-black, slight cool cast.
• Card headings use --text-strong instead of the accent, in both schemes. Blue is now reserved for things you can click.
• Cards get --surface (#1c1f26) and a stronger border. #1a1a1a on #121212 was a 1.03:1 difference and essentially invisible.
• Headline is solid ink in both schemes; gradient text fill removed.
• Primary CTA is a flat fill with a neutral shadow rather than a gradient with a coloured glow.
• Ring visualisation: teal halo band removed, peers and event hues desaturated (the five originals were Tailwind's default 400-level ramp), chord alpha raised (0.05 → 0.13 light, it was washing out entirely against the page), and the canvas ground now matches the page so there is no disc edge.

Two accessibility defects fixed on the way

Corrected after review. An earlier version of this description said the wordmark measured "1:1 on every page" and that this PR fixed "15 contrast failures". Both overstated. See the review-response comment.

• The "Freenet" wordmark was gradient-clipped text with color: transparent. It rendered fine in production, because background-clip: text painted the gradient into the glyphs — an automated checker reports 1:1 for it, but a human sees the word. The real defect is that transparent-fill text is fragile: it breaks under forced-colors, and it broke in light mode mid-way through this branch when the gradient was removed but the color: transparent was left behind. Now a plain solid colour.
• "Live network dashboard" was #888 at 0.6 opacity, compositing to 1.95:1 light / 2.67:1 dark. Now a real link at 5.32:1 / 7.1:1. This is the one unambiguous contrast fix, and it appears twice.

Verification

Playwright across home, quickstart, river, faq, news, donate, ghostkey in both schemes plus mobile: no page errors, no transparent-text, no horizontal overflow.

A contrast sweep was run against production as a control, so pre-existing failures could be separated from new ones. After the review round: 7 failures remain, all pre-existing on /river/, and none are introduced by this PR.

The first commit did introduce a regression (/quickstart/ showing two different accent blues, because the find-and-replace only covered freenet.css) and a contrast bug on the donate CTA hover at 2.53:1. Both are fixed in f4e0fed; details in the review-response comment.

Known gap

The light-mode hero screenshot (river-hero-light.webp) is not addressed. Its legible content shows users reporting that the tutorial is outdated and that there is no search engine. It needs a fresh capture from a real River session, which cannot be produced authentically here. See the review comment for detail.

[AI-assisted - Claude]

https://claude.ai/code/session_01Cwq3ZRYFqSW9rMtGJYqfo2

[sanity]

Review round 1 — findings addressed

An adversarial review pass found a regression I introduced plus several real defects. All fixed in f4e0fed, each verified in a browser rather than by inspection.

Correcting two claims I made in the PR description

The wordmark was not invisible. I wrote that it "measured 1:1 on every page in both schemes". That is what an automated checker reports for color: transparent, but background-clip: text was painting the gradient into the glyphs, so it rendered fine in production. The genuine issue is that transparent-fill text is fragile (it breaks under forced-colors, and it broke in light mode mid-way through this branch when the gradient was removed but the color: transparent was left behind). Real defect, but a much milder one than I described.

"15 contrast failures fixed" was inflated. 12 of those were the wordmark, and 1 more (Install Freenet) was a measurement artifact of walking up to a transparent background-color over a gradient. The honest count of real contrast fixes is 2 — both instances of the "Live network dashboard" link, which was #888 at 0.6 opacity compositing to 1.95:1 light / 2.67:1 dark.

The "0 introduced" half of the claim was also wrong; see below.

Regression I introduced

/quickstart/ rendered both accents at once. The find-and-replace only covered freenet.css; layouts/shortcodes/os-install.html kept 7 instances of #4da6ff. So the muted accent on "Try River in your browser" sat a few hundred pixels above the old neon on the active OS tab. Worse than either state alone. Also missed: river/baseof.html (--rv-brand), and donation-success.html + river-invite-button.html, whose #1a1a1a card surfaces the ground lift had made less visible (1.10 → 1.06) — the exact defect this PR set out to fix. All now on the shared #1c1f26 surface at 1.09.

Other fixes

• .funding-donate-button:hover set a bare accent fill under white text: 2.53:1, and lightened on hover while every other primary CTA darkens. The base rule already resolves correctly in dark, so the override was redundant and wrong. Deleted → 8.42:1 in both schemes, confirmed by actually hovering. A non-hover contrast sweep cannot catch this, which is why mine did not.
• Reduced-motion canvas never repainted on a scheme change. That path draws one frame and returns, so the per-frame scheme poll never runs. A reduced-motion visitor whose OS switches theme with the tab open kept the old scheme's ring permanently. Now redraws on the media query: mean canvas colour 105,148,190 → 58,104,157 on flip, where before it did not move.
• Hero link hover had been reduced to an underline-only change; colour shift restored. Plus a dead text-decoration: none overridden two lines later.

Verified after the fixes

/quickstart/ dark now reports a single blue accent (rgb(121,167,214)). Full sweep across home, river, faq, news, donate, ghostkey in both schemes plus mobile: no page errors, no transparent text, no horizontal overflow, and the contrast sweep shows the same 7 pre-existing /river/ failures and nothing new.

Knowingly not fixed here

• The 7 /river/ contrast failures (#6b7a8d small text at 4.19–4.38:1, and white-on-image hero text that may be a checker false positive). Pre-existing, outside these files.
• --accent-text is a half-applied abstraction: ~20 raw #79a7d6 literals remain in dark media blocks. That is precisely the fragility that produced the /quickstart/ miss, but finishing it is a mechanical refactor better done on its own.
• The light-mode hero screenshot, which needs a fresh capture from a live River session.

[AI-assisted - Claude]

[sanity]

Review round 2 — accessibility / cross-browser findings addressed

All fixed in 2d4568f, each verified by measuring the real state in a browser.

Blocking regression this branch caused

The focus ring stopped meeting SC 1.4.11. base.css drew it in --color-primary (#0066CC) — a fill colour chosen to carry white text, never checked against the surfaces it lands on. Against the old #1a1a1a card it scraped 3.13:1; raising --surface to #1c1f26 dropped it to 2.96:1, under the 3:1 required of a focus indicator. Six homepage card links affected.

Fixed by giving the ring its own --focus-ring token (defaulting to --color-primary, so nothing outside this theme changes) and resolving it to the accent in dark. Measured by focusing a real link: 2.96 → 6.53:1.

Pre-existing, but in scope and fixed here

• --navbar-item-hover was declared only in the light scheme. base.css therefore painted a near-white #f1f2f4 under near-white text on a keyboard-focused navbar item — the wordmark disappeared on Tab at 1.05:1. Now declared for dark: 12.26:1. Worth noting this directly contradicted this PR's own claim that the wordmark is "now solid"; it wasn't, in the focused state.
• os-install.html kept the coloured glow this branch strips from .cta-button, on a visually identical button — so Windows and macOS visitors still got it. Now flat fill + neutral shadow. Its :focus-visible ring stays coloured deliberately (indicator, not decoration) and uses the new accent.
• .home-browser was the last dark card on #1a1a1a; separation from the ground had fallen to 1.031:1. Now --surface, 1.09:1 like the rest.
• Four remaining rgba(77,166,255) tints migrated.
• color-scheme: light dark declared. base.css forces overflow-y: scroll, so a scrollbar is always present and was rendering light against the dark page.

Independent confirmation of the core claim

The reviewer recomputed the luminance argument: L(#4da6ff) = 0.360702, L(#79a7d6) = 0.365574 — the new accent is 1.35% lighter, so every pair where it is the foreground improves slightly. It also flagged my wording: I said contrast was "unchanged", but the ground moved too, so accent-on-ground is 7.33 → 7.10:1. The small drop comes from the lifted ground, not the new accent. Comment in the code corrected.

It also independently reproduced the honest accounting: of the "15 failures fixed", 13 were checker artifacts and 2 were real. Pixel-sampling showed the wordmark rendered at 7.3–20:1 in Chromium, Firefox and WebKit, and was fine under forced-colors too. Already corrected in the PR body.

Verified clean

• Chromium 145 / Firefox 146 / WebKit 26: no engine divergence, no page errors, canvas renders identically in all three.
• prefers-reduced-motion honoured (0 changed pixels over 1.8s), and the redraw fix confirmed against production, which kept stale dark-scheme colours on a white page indefinitely.
• forced-colors: no regression.
• Full sweep, both schemes, 10 pages, production as control: 0 new text-contrast failures.

Still knowingly not fixed

The pre-existing /river/ contrast failures, the half-applied --accent-text abstraction (~20 raw literals; mechanical, better on its own), and the light-mode hero screenshot, which needs a fresh capture from a live River session.

[AI-assisted - Claude]