test(mcp): work around Chromium losing WebMCP tools in same-origin iframes

[Dmitry Gozman (dgozman)]

Summary

• Chromium reuses the initial empty document's window, and its browser interface pipe, for an iframe's first same-origin navigation. Tools registered while parsing race the navigation commit in the browser process and are sometimes lost, with getTools() never resolving in that frame.
• The three iframe tests now reload the frame once on first load, so the second document registers its tools reliably.

Fixes #42869

[github-actions]

Test results for "MCP"

1 failed
❌ [firefox] › mcp/idle-timeout.spec.ts:19 › closes the browser after the idle timeout and relaunches it on the next call @mcp-windows-latest-firefox

8704 passed, 1474 skipped

---

Merge workflow run.

[github-actions]

🟢 The one failure is a pre-existing flake — this PR is clear

Hi, I'm the Playwright bot and I triaged the failing CI run.

[firefox] › mcp/idle-timeout.spec.ts:19 is a known flake on the mcp-windows-latest-firefox bot: it has failed 6 of 229 runs (2.6%) in the aggregated CI results, on five other PRs with the same error. This PR only edits tests/mcp/cli-webmcp.spec.ts, so it can't reach it.

Details

This PR changes a single test file — tests/mcp/cli-webmcp.spec.ts, +9/-3 — with no product code touched. Nothing it changes is shared with mcp/idle-timeout.spec.ts, and everything else in the report is green (8704 passed, 1474 skipped).

Pre-existing flake / infra

• [firefox] › mcp/idle-timeout.spec.ts:19 › closes the browser after the idle timeout and relaunches it on the next call (@mcp-windows-latest-firefox) — the same (project, file, title) on the same bot failed on five unrelated PRs, all with the identical signature:

  run	PR	sha	date
  34915545227	Fix spelling of restoring in credentials docs #42718	d979f021	2026-09-15
  35266013213	devops(bidi): publish BiDi reports to a git branch instead of Azure #42760	c78538ba	2026-09-17
  35275610916	feat(screencast): style action decorations, keep cursor visible #42758	b26a1c79	2026-09-17
  35406976347	chore(coverage): istanbul instrumentation for the trace viewer and html reporter #42785	fabe18a1	2026-09-18
  35655715523	fix(clock): make Temporal.Now follow page.clock #42838	dcdf8cf2	2026-09-21

  It's Windows-Firefox only — 0 failures across the other 12 MCP bots (mcp-ubuntu-latest-firefox 0/236, mcp-macos-latest-firefox 0/232), so this is bot-specific, not something a test-file edit would produce.

  The failure is the relaunch after the idle timeout losing the race with the browser teardown:

  - ObjectContaining { "snapshot": StringContaining "Hello, world!" }
  + Object {              "snapshot": undefined }
  
  ### Error
  Error: browserBackend.callTool: Target page, context or browser has been closed
  

  The second browser_navigate lands while the --idle-timeout=500 shutdown is still tearing the Firefox instance down, so it errors instead of relaunching. Worth a look under #42548, the issue this test already annotates — but it's independent of this PR.

_{Triaged by the Playwright bot - agent run}