ci(playwright): make the e2e job concurrent, told-where-to-seed, and readable

[rubenvdlinde]

Four defects in the shared E2E Tests (Playwright) job, all found while making it pass end to end on a repo for the first time — ConductionNL/opencatalogi#787, now green at 82 passed / 0 failed, with a bundle-move-aside control confirming the specs genuinely load the app. All four are in this job, so no caller can work around them.

1. php -S runs one worker

Single-worker by default, so every request a Nextcloud SPA fires on boot serialises. Measured: whichever spec ran first blew its 60s timeout waiting for the index page, then passed in 9.1s on retry, while every later spec ran in 4–7s. A guaranteed retry on every run — and a plain failure for any repo not using retries. → PHP_CLI_SERVER_WORKERS: 8.

2. The seed step isn't told which instance to seed

Seed test data declared no env:, while Run Playwright tests directly below exports BASE_URL/ADMIN_USER/ADMIN_PASSWORD. A seed script normally drives the app's own admin API, so the only way to write one was to hardcode http://localhost:8080 — the shared dev container on a developer box. The journeydoc capture job already exports the full set to its own seed step; this matches it.

3. The test step exports only BASE_URL

Surveying the 21 fleet repos: 15 resolve the target as process.env.NEXTCLOUD_URL || 'http://localhost:8080' and never read BASE_URL. They don't fail — they fall through to the literal, which happens to be this runner's own php -S. So the suite runs against the right host for the wrong reason, and CI silently certifies a config that points at the shared dev container the moment anyone runs it locally. Repos should still drop the literal (openregister, doriath, shillinq, planix, scholiq, zaakafhandelapp have the resolver that throws); this just stops CI depending on it.

4. Both upload steps matched nothing for most repos

They list only server/apps/<app>/playwright-report/ and .../test-results/, but the scaffolded config in most fleet repos writes under tests/e2e/. With if-no-files-found: ignore that failed quietly and uploaded an empty artifact — a red run left no report and no traces, exactly when you need them.

---

Not fixed here — needs a new input and a decision

The run step passes no --project, so every project in the chosen config runs, including visual (pixel baselines a CI Linux runner cannot byte-match — the configs say so themselves) and docs-capture (screenshot re-shoots with their own dedicated job). opencatalogi worked around it with a CI-only config at tests/e2e/playwright.config.ts, which the existing lookup finds before the repo root. A playwright-project input would let callers say this directly — happy to add it if you'd prefer that shape.

Also worth knowing (not a change here)

A newly created branch is skipped entirely by the caller guard github.event.created != true. The run appears in the list and reports skipped, so a first push looks like it ran CI and ran nothing. It takes a second push to trigger.