fix(#757): /game/cod-skeleton mounted a WebGL scene nobody had asked for

[TortoiseWolfe]

/game/cod-skeleton mounted a WebGL scene during page load. On a GPU-less runner that took
most of the mobile-layout sweep's 30s budget, so No horizontal overflow on /game/cod-skeleton
timed out three times and blocked two merges — most recently #791, which changes an edge
function and a migration and cannot touch a WebGL route.

The full finding lives in #757, rewritten with the measurement. This is the change.

It was not a flake

Measured from the results.json artifacts of the *-gen 5/6 shards across 14 e2e-local runs:

browser	/game/cod-skeleton	every other route
chromium	22.6–30.2s pass, 31.3–32.0s timeout — 12 pass / 6 timeout	median 2.7s, p95 3.7s
firefox	3.0–3.3s — 14 pass / 0 timeout	median 3.1s, p95 3.3s
webkit	12.7–20.0s pass, 30.3–41.7s timeout — 14 pass / 14 timeout	median 3.4s, p95 4.7s

The route needed ~93% of the cap, so variance decided the outcome. Every green run of it was
luck, and webkit had been retry-masking a 50% per-attempt failure rate the whole time — the run
went green, so nobody looked.

Why the page, not the test

The <Canvas> mount builds a MaterialSystem and bakes its 1K texture sets synchronously
before the page is measurable. Firefox — which has no WebGL and never mounts it — completed the
identical layout measurement in 2.8s. The whole gap is that mount.

A software rasteriser stands in for a low-end visitor, and they pay the same cost on a route
they have not asked to run anything on. Raising the timeout or suppressing the render loop from
the test would each have made a true report stop arriving. So the scene now waits to be asked —
what games do anyway, and what this page already told you to do ("Click to capture the mouse").

Measured effect

Same box, root build (DISABLE_BASE_PATH), SwiftShader forced via the flags
playwright.visual.config.ts already uses:

before   /game/cod-skeleton  7.6s   canvas=1   time-to-webgl-decided 5.3s
after    /game/cod-skeleton  2.6s   canvas=0   time-to-webgl-decided 0.2s
control  /blog/seo          ~3.5s

It is now faster than the DOM-only control. The previously failing test passes locally in 6.7s.

The gate opens, and the box does not move

Driven as the real user gesture in a real browser, not asserted from unit tests alone:

before click: canvas=0  wrapperBox=358x201
after  click: canvas=1  wrapperBox=358x201  canvasBox=300x150

The placeholder lives inside the same wrapper, so aspect-video w-full gives it the
identical box and the sweep measures the geometry it always measured.

Guards

Three, each mutation-verified — and the mutant confirmed present in the file before believing
the red, because a sed whose | collided with || once produced a green "proof" here:

• no canvas before the gate opens — killed by defaulting started to true (9 failures)
• a canvas with a real box after it opens — a gate that never opened would ship a dead route
• both branches carry the same wrapper class — killed by adding h-40 to one branch (1
  failure, precisely that test)

Plus the a11y suite now audits the placeholder state, which is what every visitor sees first,
and pins min-h-11 min-w-11 on the start control — mobile-touch-targets cannot see a control
that only exists after a click it never performs.

Checked, not assumed

• cod-skeleton-hud-contrast.spec.ts denies WebGL on purpose and asserts
  [data-webgl-ok="false"], so it takes the fallback branch and is unaffected. I had expected
  to have to update it.
• color-contrast.spec.ts excludes the route; that exclusion's stated reason is now stale —
  filed as color-contrast still excludes /game/cod-skeleton for a WebGL render loop that no longer runs at load #793 rather than widened here.
• Nothing else in tests/ or src/ expects a canvas on this route.

Not fixed here

#792 — firefox has no WebGL, so the sweep measures <FallbackPanel> on one browser and a
canvas on the other two, silently. This makes both branches cheap but does not close that
divergence: the WebGL probe still runs before the start gate. Confirmed and written up there.

Closes #757
Refs #792, #793