test(infra): harden E2E harness against recurring workerd broken-pipe crashes

[tiankaima-copilot]

Evidence

• Server currently uses Wrangler 4.125.0.
• In PR fix(ui): harden mobile layout and accessibility targets #945 run https://github.com/Life-USTC/server/actions/runs/33071505032/job/98517104269, workerd repeatedly logged kj/async-io-unix.c++:186 Broken pipe while E2E requests were still running.
• The first E2E attempt on PR perf(section): make section homework lists summary-only #949 showed the same signature, then the local Worker became unreachable and 56 subsequent tests failed with ECONNREFUSED. A rerun passed without product-code changes.
• These occurred on unrelated branches, so this is distinct from the deterministic moderation-dialog assertion fixed in PR fix(ui): harden mobile layout and accessibility targets #945.

Expected

The E2E harness should detect an unexpected Wrangler/workerd exit or failed health check and classify it as infrastructure failure. Product assertions must remain visible and must not be broadly retried.

Acceptance criteria

• Capture the Wrangler child exit status and preserve its logs in CI artifacts.
• Health-check the local Worker before and during Playwright execution.
• Retry only a confirmed infrastructure startup/crash failure, with a small fixed bound.
• Do not retry deterministic application/test assertion failures.
• Add a regression test for Worker termination during a shard.

[github-actions]

🎉 This issue has been resolved in version 1.95.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀