fix(daemon): let POSIX stop() see the daemon exit instead of always SIGKILLing

POSIX stop() set should_stop_, which made the monitor thread return
without reaping the child, and then polled the state-based isRunning(),
which stays true while the state is Stopping. So every Linux/macOS stop
waited out its whole grace period, sent SIGTERM, waited 10 s more and
SIGKILLed, even when the node had exited seconds earlier. Closing the GUI
with an idle node took ~30 s; a node still flushing at 30 s was killed.

- stop() joins the monitor first, then polls waitpid(pid, WNOHANG) on
  a cached pid, so it is the only reaper and sees the exit at once.
- markStopRequested() (set by App::stopEmbeddedDaemon before the RPC
  stop, and by stop() itself; cleared by start()) makes an exit that
  lands before stop() runs a clean stop instead of a crash, on POSIX
  and in the Windows monitor. crash_count_ drives auto-restart, and a
  crash recorded before the request keeps its count.
- The SIGTERM -> SIGKILL escalation stays for now, but only fires for
  a daemon that has really not exited. Replacing it with a wait policy
  and an explicit user choice is follow-up work.

Verified against the fake daemon: an idle node that exits 3.4 s after
stop used to be SIGTERMed at +20 s and SIGKILLed at +30 s (GUI closed
in 30.6 s); it is now reaped as "exited cleanly" ~50 ms after it exits
(GUI closes in 4.0 s). New POSIX tests cover the clean exit, the early
requested exit and a graceful SIGTERM, each time-bounded so the old
code fails fast.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/daemon/daemon_controller.cpp

@@ -52,6 +52,11 @@ void DaemonController::stop(int waitMs)
     daemon_->stop(waitMs);
 }
 
+void DaemonController::markStopRequested()
+{
+    daemon_->markStopRequested();
+}
+
 bool DaemonController::isRunning() const
 {
     return daemon_->isRunning();