fix(gateway): name the display-wake stall in the launchd wrapper docstring; gate the macOS wrapper test with the platforms marker

The do shell script wait also stalls external-display wake on macOS 27
(#123595), not only idle CPU; the JXA system() wait removes both. The new
lifecycle test used a bare sys.platform skipif, which the OS lanes never
import; use @pytest.mark.platforms("macos") so the macOS lane runs it.

hermes_cli/gateway_launchd.py

@@ -226,7 +226,8 @@ def launchd_program_arguments(command: list[str], stdout_log: Path, stderr_log:
     executable).
 
     Standard Additions' ``do shell script`` polls WindowServer for a user-cancel event while it waits.
-    That is appropriate for a short interactive script but burns CPU for the gateway's process lifetime.
+    That is appropriate for a short interactive script but, for the gateway's process lifetime, burns CPU
+    and keeps a WindowServer event connection busy (external-display wake stalls ~10s on macOS 27, #123595).
     JXA calling libc ``system()`` waits in the kernel instead while retaining osascript as the responsible
     process. The shell's ``exec`` keeps the gateway in the launchd job's process group, so ``launchctl
     bootout`` / ``kickstart -k`` still deliver SIGTERM to it. stdout/stderr are appended inside the shell

tests/hermes_cli/test_gateway_service.py

@@ -1782,7 +1782,7 @@ class TestProfileArg:
         # The wrapper's own ps line must never be taken for the gateway (stop/status would signal osascript).
         assert status.looks_like_gateway_command_line(" ".join(program_args)) is False
 
-    @pytest.mark.skipif(sys.platform != "darwin", reason="osascript is macOS-only")
+    @pytest.mark.platforms("macos")
     def test_launchd_osascript_wrapper_preserves_process_group_and_exit_status(self, tmp_path):
         """The non-polling JXA wait keeps lifecycle signals and KeepAlive failure semantics intact."""
         stdout_log = tmp_path / "stdout.log"