Desktop E2E: 'failed initial relay dial retries automatically' is order-dependent (passes only as the first test in the file) #6242

Summary

failed initial relay dial retries automatically in desktop/tests/e2e/relay-reconnect.spec.ts fails whenever any test precedes it in the file. It currently passes only because it is the first test in the file.

This surfaced while reviewing #3320, which adds a test above it. #3320 is not the cause — the defect reproduces on clean main with that PR entirely absent.

Reproduction (on main @ 5694e78de, PR absent)

Unmodified file:

14 passed

Insert a trivial no-op test immediately above failed initial relay dial retries automatically:

test("trivial preceding test", async ({ page }) => {
  await installMockBridge(page);
  await page.goto("/");
  await expect(page.getByTestId("channel-general")).toBeVisible();
});

Result:

✓ trivial preceding test
✘ failed initial relay dial retries automatically
  Error: page.evaluate: Error: Relay state seam is not installed.
    at relay-reconnect.spec.ts:207

Content of the preceding test is irrelevant — I reproduced with a trivial body, and with an unrelated existing test. Moving the added test below it restores green (15/15).

Why it matters

The failure mode is __BUZZ_E2E_GET_RELAY_CONNECTION_STATE__ not being installed when the poll first runs, i.e. the test races bridge installation. As the first test in the file it wins that race; anywhere else it loses.

This makes a load-bearing reconnect test silently order-dependent: any PR that adds a test above it inherits a confusing red that looks like it was caused by that PR. It also means the test's guarantee is weaker than it appears.

Suggested fix

Have the poll tolerate the seam not yet existing (return a sentinel rather than throwing) so expect.poll can retry, instead of throwing out of page.evaluate — the other call sites in the same file already use __BUZZ_E2E_GET_RELAY_CONNECTION_STATE__?.() for this reason. Line 207 is the only one that throws.

Environment note (separate, minor)

playwright.config.ts:186 uses reuseExistingServer: !process.env.CI against a fixed port 4173. Two agents/worktrees running desktop E2E on one machine serve each other's dist, producing unrelated cross-worktree failures. Worth a per-worktree port or at least a doc note.