Run the CI suite on stein2 next to the gate

Suite 759 went red on e2e 67a-ii (firstAt 0.083 s) while a job of
another project loaded the GPU on melektaus-atx. The suite measures
frame timings, so it should not share that GPU.

- The suite job now runs on `feuerwerk-ci`, the runner on stein2,
  like the gate. Its config.yaml already passes /dev/dri/renderD128
  and a 1 GB /dev/shm into the job container, and has capacity 2.
- The header of ci.yml says how to move the suite back to
  `feuerwerk-wsl`, and that GM_TEST_WORKERS now matters again,
  because gate and suite share the cores of one machine.

Tests: YAML parses, both jobs name `feuerwerk-ci`. Not run in CI yet,
the next push starts it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.gitea/workflows/ci.yml

@@ -1,19 +1,19 @@
 # Runs the whole proof on the Gitea runner: gate, the editor suite and the three example games.
 #
-# The suite runs on `feuerwerk-wsl`, the runner in the WSL2 distro feuerwerk-ci on
-# melektaus-atx (#283). Its label picks the image built from tools/ci/Dockerfile. The GPU
-# (/dev/dxg and the Windows driver libraries), /dev/shm and FEUERWERK_BROWSER=wsl come from that
-# runner's own config.yaml, not from here – without them chromium falls back to software WebGL and
-# the harness stops. The gate needs no GPU and runs on `feuerwerk-ci`, the runner on stein2 with
-# the same image. Pointing the suite at that label too moves the browser runs back to stein2.
+# Both jobs run on `feuerwerk-ci`, the runner on stein2. Its label picks the image built from
+# tools/ci/Dockerfile. The GPU (/dev/dri/renderD128) and /dev/shm come from that runner's own
+# config.yaml, not from here – without them chromium falls back to software WebGL and the harness
+# stops. The suite ran on `feuerwerk-wsl` in the WSL2 distro feuerwerk-ci on melektaus-atx (#283),
+# whose config adds /dev/dxg and FEUERWERK_BROWSER=wsl. Pointing the suite at that label again
+# moves the browser runs back there. It moved to stein2 because other jobs on melektaus-atx share
+# its GPU and made the suite's frame timings wobble.
 #
 # Two jobs without `needs:`, so they start together. The gate does not stand in front of the
 # suite: a lint error used to cost the full e2e run before it was reported, and the run reported
 # it after six minutes instead of one. Both still have to pass for the push to be green.
 #
-# On two runners the gate's vitest no longer shares cores with the suite's chromium, which
-# measures text widths and frame counts. GM_TEST_WORKERS still caps vitest at two workers, for the
-# day both jobs share one runner again.
+# On one runner the gate's vitest shares cores with the suite's chromium, which measures text
+# widths and frame counts. GM_TEST_WORKERS caps vitest at two workers for that.
 name: feuerwerk
 
 on:
@@ -51,7 +51,7 @@ jobs:
           php tests/php/rights.test.php
 
   suite:
-    runs-on: feuerwerk-wsl
+    runs-on: feuerwerk-ci
     timeout-minutes: 45
     steps:
       - uses: actions/checkout@v4