ci: run browser journeys on three shards per engine

.github/workflows/ci.yml

@@ -129,16 +129,18 @@ jobs:
 
   browser:
     if: github.event_name != 'workflow_dispatch'
-    name: Browser journeys (${{ matrix.engine }}, ${{ matrix.shard }}/2)
+    name: Browser journeys (${{ matrix.engine }}, ${{ matrix.shard }}/3)
     # Measurements stay isolated on their own runner, not on our critical path.
     # CI required still requires their success independently of these jobs.
     runs-on: ubuntu-24.04
+    # Keep enough shards that slow runner setup cannot push a passing shard
+    # past this limit; two shards ran 11-14 minutes and were often cancelled.
     timeout-minutes: 15
     strategy:
       fail-fast: false
       matrix:
         engine: [chromium, webkit]
-        shard: [1, 2]
+        shard: [1, 2, 3]
     steps:
       - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
         with:
@@ -161,7 +163,7 @@ jobs:
       - name: Functional journeys
         env:
           PLAYWRIGHT_JSON_OUTPUT_FILE: test-results/browser/ci-report.json
-        run: node scripts/ci-test-report.mjs kind=playwright report=test-results/browser/ci-report.json evidence=test-results/browser/ci-timing.json title="Browser journeys (${{ matrix.engine }}, ${{ matrix.shard }}/2)" -- pnpm test:browser:ci --project ${{ matrix.engine }} --no-deps --shard=${{ matrix.shard }}/2 --reporter=list,json
+        run: node scripts/ci-test-report.mjs kind=playwright report=test-results/browser/ci-report.json evidence=test-results/browser/ci-timing.json title="Browser journeys (${{ matrix.engine }}, ${{ matrix.shard }}/3)" -- pnpm test:browser:ci --project ${{ matrix.engine }} --no-deps --shard=${{ matrix.shard }}/3 --reporter=list,json
       - name: Browser evidence
         if: always()
         uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4

docs/browser-testing.md

@@ -3,8 +3,8 @@
 `pnpm test` and therefore `just scan` now run the checked-in Chromium and WebKit
 journeys in `tests/browser/`. `just iterate` remains the fast, browser-free loop.
 [GitHub Actions](../.github/workflows/ci.yml) runs these gates on PRs and main
-pushes: an isolated serial measurement job runs alongside four functional jobs
-(two file-level shards per engine, two workers per runner). Both engines remain
+pushes: an isolated serial measurement job runs alongside six functional jobs
+(three file-level shards per engine, two workers per runner). Both engines remain
 mandatory for the [CI-selected coverage](#ci-coverage-and-local-only-webkit-checks);
 `CI required` fails if any lane or shard fails, is cancelled, or is skipped.
 Owner-specific tests are colocated in `src/` and `dev/`; separate manual diagnostic
@@ -106,9 +106,9 @@ instances and every untagged WebKit case remain required. Both engines, serial
 measurement order, zero retries, the documented functional assertions and
 measurement ceilings, and the strict `CI required` aggregate remain in place.
 
-CI shards each functional engine across two runners, without waiting for the
+CI shards each functional engine across three runners, without waiting for the
 separate measurement runner. Each job selects its engine with `--no-deps` and
-`--shard=N/2`; measurement success is enforced by `CI required`, not job ordering.
+`--shard=N/3`; measurement success is enforced by `CI required`, not job ordering.
 This preserves measurement isolation while spending more setup/runner minutes,
 including when measurements fail. Local same-runner dependencies remain unchanged.
 Artifacts include engine and shard so parallel jobs never overwrite one another.

docs/contributing.md

@@ -320,7 +320,7 @@ the complete suite still runs with `pnpm test` / `just scan`:
   tests build Rust and install scaffold dependencies; they are intentionally CI-only
   rather than part of pre-push.
 - **Browser measurements:** Chromium then WebKit, serially on an isolated runner.
-- **Browser journeys:** four runners (Chromium and WebKit, two file-level shards
+- **Browser journeys:** six runners (Chromium and WebKit, three file-level shards
   per engine), each with two workers. They start alongside measurements on separate
   runners; `CI required` still requires both lanes. Each runner builds the native
   plugin-manager fixture in a separately logged setup step before starting

tests/integration/browser-ci.test.mjs

@@ -35,9 +35,9 @@ const matrixValues = (key) => {
   return values.split(",").map((value) => value.trim());
 };
 
-test("four independent browser jobs retain isolated measurements and native setup", () => {
+test("six independent browser jobs retain isolated measurements and native setup", () => {
   assert.deepEqual(matrixValues("engine"), ["chromium", "webkit"]);
-  assert.deepEqual(matrixValues("shard"), ["1", "2"]);
+  assert.deepEqual(matrixValues("shard"), ["1", "2", "3"]);
   assert.doesNotMatch(browser, /^ {4}(needs|continue-on-error):/m);
   assert.doesNotMatch(browser, /^ {8}(include|exclude):/m);
   assert.match(browser, /^ {6}fail-fast: false$/m);