fix(#310): firefox-gen was red since #297 — and "renders the atlas" passed on a DEAD atlas

[TortoiseWolfe]

Closes #310. Found by the owner spotting a red shard on main.

main has been red for four runs

firefox-gen 6/6 has failed on every main run since a8ee9b4 — #297's merge, when the
atlas became the default. PR CI is chromium-only (firefox/webkit only run on push-to-main), so
#297 went in green and four more merges landed on top of a red main without anyone looking.

run	sha	merged	firefox-gen 6/6
c6b7021	—		✅
a8ee9b4	#297 — atlas becomes default		❌
c2cadd0	#304		❌
77d464f	#306		❌
8910cc1	#308		❌

The cause is the environment, not the atlas

Probed with Playwright's Firefox against the real page:

webgl: false
hudText: "Atlas — chatt   error: The browser supports WebGL, but initialization failed."
typeChip: false
console: "Error constructing CesiumWidget. Visit http://get.webgl.org ..."

Headless Firefox has no WebGL → Cesium's Viewer ctor throws → AtlasViewer catches it into its
error state → everything behind {ready && …} (chips, count, legend) never renders. Real
Firefox has WebGL. The tests were right; two of them just needed the WebGL gate this file
already uses for the R3F canvas.

The worse bug this exposed

test('/chatt/ renders the atlas by default', ...)
  await expect(page.getByText('Atlas —')).toBeVisible()

Atlas — {slug} renders unconditionally — the status line beside it is a ternary whose first
branch is error: ${error}. The probe shows Atlas — chatt sitting directly beside the error.

So the test whose whole job was proving the atlas renders was green, on Firefox, on a
completely dead atlas. It could not fail when the product was broken. Same family as
toBeVisible() ignoring occlusion (#299) and the placeholder client_id a test asserted
toBeTruthy() on (#288).

The fix: the mount and the scene are different claims, so they're different tests

test	gate	proves
the atlas MODULE mounts	none — runs everywhere	the Cesium chunk loaded and the module evaluated. Deliberately weak. This is what catches a #294-class dead vendor chunk (which took prod down for five hours), on every browser, WebGL or not.
the atlas SCENE initialises without erroring	WebGL	the HUD contains no error: and the chips exist. Cannot pass on a dead atlas.

Verified RED for the right reason — skip removed, run on Firefox's genuinely dead atlas:

Expected string: not "error:"
Received string: "Atlas — chatt error: The browser supports WebGL, but initialization failed."

It fails on the exact atlas the old test called green.

The gate is deliberately narrow

test.skip() is #288's false-green failure mode, so only the two specs that need ready are
gated (the building count, the type chip). Everything else — cannot-scroll, HUD width, cookie
banner/PWA, contentinfo, the #299 wordmark — passes on Firefox without WebGL and stays
ungated, keeping its signal on all three browsers. Chromium has swiftshader, so every real
assertion still runs on every PR.

The building count test is not loosened: it's the one that catches "0 buildings" over a
correct 8031-building legend, which has shipped once.

Verified on a CI-identical root build

firefox-gen    13 passed,  5 skipped,  0 failed      (was: 2 failed)
chromium-gen   20 passed,  0 skipped