subprocess_seconds was being applied per hop instead of as a total trace
budget, so a path with several non-responding hops could block for up to
max_hops * subprocess_seconds. The blocking call also ran on a
ThreadPoolExecutor worker via asyncio.to_thread, which concurrent.futures
joins at process exit - so an orphaned hop stuck past its timeout kept the
whole CLI alive long after diagnose() had already written its report.
Divide the budget across hops (capped at 4s per hop) with an early exit
once the total deadline passes, and dispatch the call on a raw daemon
thread instead so it can never hold the process open even in the worst
case.
