Summary

_run_stdio() is an async MCP startup path, but it was calling the OSV malware preflight synchronously. That preflight eventually reaches urllib.request.urlopen(), so an intermittent OSV SSL/connect hang can block the entire MCP event loop during server discovery. In TUI startup this can exceed the 15s gateway readiness window even though MCP discovery itself is waiting on a 120s future.

This PR now covers both layers:

1. move the blocking OSV lookup into asyncio.to_thread() so MCP discovery keeps the event loop responsive
2. wrap that threaded check with an explicit startup guard so stdio MCP startup fails open if the OSV check itself never returns promptly

This preserves the existing pre-spawn malware check and error handling while avoiding a global socket.setdefaulttimeout() mutation.

The bug

tools/mcp_tool.py::_run_stdio() runs inside the MCP asyncio loop:

malware_error = check_package_for_malware(command, args)

check_package_for_malware() calls into tools/osv_check.py, which uses synchronous urllib I/O. When the OSV request stalls in SSL handshake or network setup, the MCP event loop cannot schedule other work. That makes MCP startup appear hung and can cause the TUI gateway startup timeout described in #29184.

Even after moving the work into a thread, a lower-level OSV stall can still leave startup waiting on the thread result. Because the OSV malware check is already fail-open for network errors, the bounded async wrapper should also fail open on timeout.

The fix

• Keep the malware check in the same pre-spawn position, before constructing/starting the stdio MCP server.
• Run it via asyncio.to_thread(check_package_for_malware, command, args) so blocking urllib work happens off the event-loop thread.
• Wrap the threaded check in asyncio.wait_for(..., timeout=_OSV_MALWARE_CHECK_TIMEOUT).
• On timeout, log at debug level and continue with malware_error = None, matching the existing fail-open policy for OSV/network failures.
• Preserve the existing behavior when OSV flags a package: _run_stdio() still raises ValueError("MCP server '<name>': ...") before spawning.

Regression coverage

Added coverage in tests/tools/test_mcp_tool_issue_948.py:

• test_run_stdio_malware_check_does_not_block_event_loop
  • patches the malware check to sleep for 200ms, simulating a slow/blocking OSV lookup
  • starts a stdio MCP server task with mocked MCP session/client objects
  • asserts that await asyncio.sleep(0.05) resumes promptly instead of being delayed by the malware check
  • verified RED first: before the fix the test failed with elapsed ~= 0.20s; after the fix it passes
• test_run_stdio_malware_check_timeout_fails_open
  • lowers _OSV_MALWARE_CHECK_TIMEOUT to 50ms
  • simulates a malware check that would return only after 200ms
  • asserts stdio MCP startup proceeds before the hung check returns

Validation

Closes #29184.