fix(mcp): do not crash when download fails after browser close

[Yury Semikhatsky (yury-s)]

Summary

• Catch saveAs rejections from the MCP tab download handler; when the browser was closed mid-download the rejection fired after the context stopped listening for unhandled rejections and took down the server.

Fixes #42831

[github-actions]

Test results for "MCP"

3 failed
❌ [chrome] › mcp/annotate.spec.ts:291 › should enter annotate mode on fresh dashboard.tsx mount with -s --annotate @mcp-windows-latest-chrome
❌ [chromium] › mcp/cli-json.spec.ts:191 › tool error on bad navigation returns JSON error @mcp-macos-latest-chromium
❌ [webkit] › mcp/http.spec.ts:624 › should close session when heartbeat ping is not answered @mcp-macos-latest-webkit

8748 passed, 1480 skipped

---

Merge workflow run.

[github-actions]

Hi, I'm the Playwright bot and I took a first look at the CI failures here.

🟡 Nothing points at this PR, but two failures aren't proven flakes

The PR only changes the download listener in tools/backend/tab.ts to log a rejected _downloadStarted() instead of leaving it unhandled, and adds one test in mcp/files.spec.ts. None of the three failing tests start a download. The annotate.spec.ts failure is a confirmed flake. The other two, http.spec.ts:624 on [webkit] and cli-json.spec.ts:191 on [chromium], have no earlier failures of their own in the DB. The same errors have shown up on sibling projects on main, which points toward flakes.

Details

The failures come from the MCP merge report, which matches test run 36074277043 (sha 0fc10c0f). The new test, closing browser during download does not crash the server, passes in that run. It timed out on macOS in an earlier push (0d05cbad) and passed after it was switched to isolated: true.

Pre-existing flake / infra

• [chrome] › mcp/annotate.spec.ts:291 › should enter annotate mode on fresh dashboard.tsx mount with -s --annotate (Windows): getByRole('main', { name: 'Dashboard: annotate' }) not found. In the test-results DB this test flips verdict on every browser, with 19 failed runs and 13 runs rescued by a retry out of about 3,700. The chrome project alone failed 4 of 749 runs, on chore: introduce _TestInfoEx interface for builtin fixtures #42352, test: race between close() and navigation with RenderDocument #42251, fix(mcp): keep upload modal, skip short secrets and bad headers #42713 and main (0d3f1b47). The dashboard annotate flow doesn't touch the download handler.

Uncertain

• [webkit] › mcp/http.spec.ts:624 › should close session when heartbeat ping is not answered (macOS): Expected: 1, Received: undefined after the 5s poll. The only [webkit] failure in the DB is this PR's run. The identical error hit [firefox] on Windows on main (0f5a946a, run 33657515908) and passed on retry. The failure is in HTTP session heartbeat handling, which this PR doesn't touch.
• [chromium] › mcp/cli-json.spec.ts:191 › tool error on bad navigation returns JSON error (macOS): 30s timeout with SyntaxError: Unexpected end of JSON input. The [chromium] project has no other failure in the DB, so this is its first recorded one and I can't strictly call it a flake. The test is a bad goto and never fires a download event, so the changed code doesn't run. The identical error failed [chrome] on macOS on main (80ef9143, run 33186806849) and [webkit] on macOS on feat(reporter): report step params in the chrome://tracing report #42358, which strongly suggests a flake. A green rerun of the MCP macOS chromium job would settle it.

_{Triaged by the Playwright bot - agent run}