ci: retry once when vitest loses its browser page

[obiot]

The browser suite aborts at a random point with:

Error: Failed to run the test <spec>.
Caused by: [vitest] Browser connection was closed while running tests.
Caused by: [birpc] rpc is closed, cannot call "createTesters"
 Test Files  <N> passed (263)

A different victim spec every time — six distinct ones observed — and zero
failed tests in every occurrence. Measured at 4 deaths in 21 local runs; on CI
it took three attempts to land a PR that changed nothing but Markdown.

Cause

Vitest 4.1.11 runs each worker as one long-lived orchestrator page and dispatches
every spec over a single websocket. Its ws.on("close") handler destroys the
session and rejects the pending createTesters, so losing that socket kills the
whole run — the in-page client does attempt to reconnect, but the session is
already gone. That is why no test ever fails and the named spec is just whichever
was in flight.

Underneath, the browser intermittently severs a renderer's network-service
channels: every websocket belonging to that page dies within ~3 ms, the server
receives a graceful 1001 while the page sees 1006 wasClean=false, and both
processes stay alive with healthy memory.

Ruled out by direct measurement:

hypothesis	result
memory / OOM	2.46 GB peak against 64 GB
CPU starvation	5/5 clean under 14 competing hogs
browser crash	no crash event, no signal in stderr
page navigation / reload	zero in ~90 instrumented runs
worker concurrency	single worker was worse, 2/6
drawmesh_bench (#1612)	skips in ~30 ms, no stall in any log
CI-specific	reproduces locally

Why a retry rather than a fix

Launching the full Chromium build instead of the headless shell fixes it
locally — 20 consecutive clean runs against a baseline that died 4 times in
21. It does nothing here: 0 of 4 CI attempts passed with it applied. So it
was dropped, and this retries instead.

The condition is deliberately narrow — the page must have been lost and no
test may have failed:

grep "Browser connection was closed"  &&  ! grep "Test Files.*N failed"

A genuine failure prints N failed in the summary and exits immediately with no
second attempt. Both branches were verified against real captured logs from this
investigation: the flake log retries, a 1 failed | 262 passed log does not.

texture.spec.js

A latent bug found on the way, independent of the retry. The app is built with
video.AUTO while the test asserts boundTextures and currentTextureUnit —
WebGL texture-unit bookkeeping that no WebGPU batcher keeps. It passes only
because the headless browser has no GPU and falls back to WebGL; on a machine
where WebGPU is available it fails. Now gated on the renderer rather than on the
fields being present, so a WebGL regression that drops them still fails rather
than silently skipping.

Worth reporting upstream

Vitest destroying the session on the first socket close while its own client is
still retrying, and @vitest/browser-playwright silently ignoring an
instance-level launch: key (it only reads launchOptions from the
playwright({...}) factory call — setting the channel on the instance looks
correct, does nothing, and quietly re-measures the default).

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.