ci(e2e): fail loud when wait-on times out, capture serve diagnostics

[TortoiseWolfe]

Summary

CI run 24970006226 on commit `00edd23` had 14/26 jobs fail. All 7 firefox shards and all 6 webkit-gen shards failed with `NS_ERROR_CONNECTION_REFUSED` to `http://localhost:3000/account\` in their first `beforeEach`. The static server (`npx serve out -l 3000`) wasn't responding when Playwright tried to connect, but the workflow proceeded to run tests anyway — producing 50 minutes of cascade ECONNREFUSED per shard with no actionable signal.

The bug is in the workflow: `Start server` ran `npx serve ... &; sleep 5; npx wait-on ... --timeout 60000`, and wait-on's failure was not reliably propagating as a step error (likely `-e` interacting badly with the backgrounded job).

Fix

Replace all 6 identical `Start server` blocks with a defensive version that:

1. Captures serve's PID and tees output to `serve.log`
2. Verifies serve didn't exit immediately (port in use, missing `out/`, etc.)
3. Explicitly checks wait-on's exit code via `if ! ... ; then ...`
4. On failure: dumps `serve.log`, listening sockets (`ss`/`netstat`), serve process state (`ps`), and a direct `curl` probe — then `exit 1`
5. On success: prints `serve is responding on http://localhost:3000\` for grep-friendly logs

Affected jobs (one block each): `smoke`, `rate-limiting`, `auth-setup`, `e2e`, `e2e-firefox`, `e2e-webkit`. All 6 jobs preserve their existing `env: CI: true` blocks (where present on `e2e`, `e2e-firefox`, `e2e-webkit`).

What this gives us

After merge, the next failing run produces clear diagnostics at the moment of failure, instead of cascading 100s of test failures downstream. Total wall time on a serve-bind failure goes from ~50 minutes per shard to ~90 seconds. That signal will tell us why serve isn't binding (timeout too short? serve crash? port collision?) — currently the cascade noise drowns it out.

What this doesn't do

• No timeout bump. We have no evidence yet that 60s is too short — bump speculatively only after diagnostics confirm. Follow-up plan, not this PR.
• No source-code changes. Workflow only.
• Doesn't address #50/#57 (cross-shard messaging test-user collision) — different root cause, different fix.

Test plan

• YAML parses cleanly (`python3 -c "import yaml; yaml.safe_load(open('.github/workflows/e2e.yml'))"`); 8 jobs intact.
• All 6 `Start server` blocks contain `serve started, PID=` and `wait-on timed out` strings (`grep -c` returns 6).
• The 3 `env: CI: true` blocks (e2e/firefox/webkit jobs) are still attached.
• Real CI verification on this PR's run: smoke, rate-limiting, auth-setup, and chromium-gen jobs should produce `serve is responding on http://localhost:3000\` log line. Firefox/webkit jobs may either pass or fail loudly (≤90s) with diagnostics; either is the signal we want.

🤖 Generated with Claude Code