fix(#301, #299): the scene is the page, and the chrome floats on glass

[TortoiseWolfe]

Closes #301. Closes #299.

The bug was not z-index — the page scrolled

The type chip the atlas now defaults to (#304) was unreachable. Measured live on /chatt:

nav 65 + h-screen 630 + pb-14 56 + footer ~120  =  871   vs a 630 viewport
scrollHeight 871   innerHeight 630   canScroll TRUE

scrollY	chip top	elementFromPoint at the chip's own centre
0	156	the chip itself
100	56	NAV — "ScriptHammer Home Blog Docs 3D Map"

It looks perfect at scroll 0 — which is why it read as "renders fine now, randomly glitches
later" (the variable was scroll position, not time) and why every screenshot taken without
scrolling looked fine.

The footer was the absurdity at the centre: 297px below the fold, unreachable, serving no
one, existing only to generate the scroll that hid the HUD. One root layout renders it for
every route, so a 3D atlas inherited a document shape.

So the twin routes stop pretending to be documents

• The scene IS the page — fixed; inset:0, rendering behind the chrome. Out of flow, the
  flex-1 wrapper collapses to free space, the body is exactly the viewport, and the page
  cannot scroll. The HUD cannot hide because there is no scroll to hide it. Measured:
  scrollHeight 871 → 630. Also un-clips the globe's bottom — the same bug's other half — and
  retires the 100dvh question, since inset:0 computes no length.
• The nav floats on glass (base-100/90 + blur) — the atlas's own answer for its critical
  panels, applied to the nav.
• The footer becomes a compact 44px glass strip pinned over the scene: visible and useful
  instead of dead below the fold.

#299 closes with it

The diorama's Hud was inset:0, putting "Chattanooga Mini" at y=16 — permanently under the
nav. Nobody had ever seen the page's own title, and toBeVisible() asserted it was
visible for months, because it checks CSS and not occlusion. It now lands at y=92 and hit-tests
to itself.

Chrome ownership moves to the route

The twin-fullscreen effect lived in TwinCanvas (the diorama), so when #292 made the atlas
the default, /chatt — the most-shared route — got no chrome suppression at all, which is
why the PWA popup collided with the target card. A renderer cannot own what the route needs.

Two classes, because the renderers genuinely want different things:

class	who	effect
twin-fullscreen	both	scene is the page; hides the PWA popup
twin-diorama	diorama only	also hides the cookie banner, whose ~66px buries a dock at bottom:16

The atlas keeps its cookie banner (owner's call): /chatt is the link we promote, and a
first-time visitor must still be able to consent. The banner covers the strip until dismissed —
accepted, and it's dismissible.

--nav-h is a CSS var, not a TS constant

This spans two styling idioms — AtlasViewer uses Tailwind classes, stage/Hud.tsx uses inline
style objects — and a var is the only thing both can share. It's calc(4rem + 1px): the nav's
real height. Every existing hard-coded site says 64 and is 1px short; migrating them is out
of scope and the debt is commented.

Verification — 19/19 against a CI-identical root-path build

• the page cannot scroll; scrollTo(0, 500) is a no-op (structural, cannot flake)
• elementFromPoint at the chip's own centre hits the chip and not the nav, after
  attempting to scroll. toBeVisible() is CSS-only and would pass with the HUD buried — which is
  exactly how The diorama's 'Chattanooga Mini' wordmark is invisible behind the nav — and toBeVisible() passed on it for months #299 hid for months
• the wordmark probe that proved The diorama's 'Chattanooga Mini' wordmark is invisible behind the nav — and toBeVisible() passed on it for months #299 now runs as the fix's assertion
• atlas keeps the banner / diorama hides it; exactly one contentinfo, carrying FOOTER_LINKS
• glass contrast: the nav's computed background composited over #000 and over #fff in a
  2D canvas (the canvas does the real OKLCH→sRGB the browser paints with; a regex reads L/C/H as
  R/G/B and collapses the ratio — this repo has measured a real 4.13:1 mis-read as 1.03:1).
  Those endpoints bracket every pixel the scene can produce, and blur() moves a backdrop
  toward its local mean — strictly inside the bracket. Asserting equal sign at both endpoints
  makes the minimum provably an endpoint rather than two lucky samples. Measured: dark 10.33:1,
  light 10.37:1.
  Plus an honesty guard — alpha < 1 on /chatt and == 1 on /, or an opaque nav (which has
  superb contrast) would sail through while shipping nothing.
• unit: 3954 passed; type-check + lint clean; pre-push validate-ci.sh --quick green including
  the Adding cesium inlines a WASM module as a template literal → vendor chunk fails to parse → client JS dead app-wide #294 chunk-parse check

Screenshots reviewed for both renderers; the diorama's R3F canvas grows 64px and its framing was
checked by eye, not just by a green test.

Follow-ups filed, not fixed: #307 (the HUD's ~302px — PR #297's row split is ~52px of it).