[krazymir]

What

Three e2e fixes from Jenkins build 191, which failed on one spec and flagged five as flaky. Test only, so there's no version bump and no changelog entry.

Why

Every failure lined up with two windows where the next dev server stalled. Single requests took 15–32 s, the watchdog's 5 s health probe timed out, and Turbopack logged a 10 s "filesystem cache database compaction". Three tests failed in the same second at the end of the first stall. On top of that, one test had a genuine race.

The noise in the server log is not related. The many /api/users/:id/photo 404s are the gateway working as specified: dev personas have no Entra photo. The uncaughtException: Error: aborted (ECONNRESET) blocks are a known Next.js bug (vercel/next.js#84649, #99278), and the server only logs them and keeps running.

Changes

• e2e/discovery.spec.ts: the profile's e-mail switch flips optimistically before its PATCH /api/profile lands. The test reloaded right after the click, which aborted the save (the server logged client disconnected for /api/profile), so the reload read the old value. It now waits for the PATCH response before reloading.
• e2e/helpers/auth.ts (signIn): the whole sign-in is now retried, for up to 100 s, not just the panel load.
  • Each attempt reloads Home and waits for whichever renders first: the account menu or the dev panel.
  • If the menu is already there, an earlier attempt's slow callback landed. The persona is not posted again, because that would re-seed it (for example, reset fresh onboarding).
  • The 30 s wait for .user-trigger was the most common failure point, since every test signs in.
• e2e/public-submissions.spec.ts: the two end-to-end flows (submit → confirm → accept, and Specific people) call test.slow(), the same convention the directory-card and admin-settings specs use.
  • Specific people does three sign-ins, a proof-of-work wait, confirm, accept, a triage move and an inbox check. It failed all three attempts in build 191.
  • It runs in the serial platform-state project, so its first retry was skipped behind the preceding test's timeout and its second landed in the next stall.

Reviewer notes

• Not run against a dev server yet: the next CI run is the real check. The three files type-check (tsc --strict, checked separately) and playwright test --list loads every affected spec.
• Coverage gap, not fixed here: pnpm typecheck does not see e2e/ (the web tsconfig includes only src), and ESLint ignores e2e/ too. A type error in a spec would only show up when Playwright transpiles it.
• The stalls themselves look like Turbopack persistent-cache compaction and memory eviction on the agent. The archived e2e-host-stats.log from build 191 would confirm that.