fix(gateway): pool-route voice RPCs so STT lazy install can't block WS sends

voice.toggle (status) triggers check_voice_requirements() -> STT provider
auto-detect -> a synchronous faster-whisper lazy install (uv/pip subprocess
with a 300s timeout). Inline on the WS reader thread it stalls handle_ws
before it reads the next frame, so prompt.submit / session.list queued
behind a voice.toggle sit unread and the desktop 'send message' appears
dead for minutes while the install churns (reproduced: voice.toggle ->
session.list 40s+ timeout).

Route voice.toggle/voice.record/voice.tts to the RPC pool (same bug class
as #21123 / #50005) so a slow lazy install can't block message handling.
Adapt the voice handler tests to drive the handler inline via a small
_dispatch_sync helper (preserving transport-binding semantics) since
dispatch() now returns None for pool-routed methods, and add a regression
test asserting the voice RPCs stay pool-routed.

tests/test_tui_gateway_server.py

@@ -16,6 +16,24 @@ from hermes_cli.active_sessions import active_session_registry_snapshot
 from hermes_cli.browser_connect import ChromeDebugLaunch
 from tools import async_delegation as ad
 from tui_gateway import server
+from tui_gateway.transport import bind_transport, reset_transport
+
+
+def _dispatch_sync(req: dict, transport=None) -> dict | None:
+    """Run one RPC to completion synchronously, regardless of pool routing.
+
+    Voice RPCs (voice.toggle/record/tts) are in ``_LONG_HANDLERS`` so they run
+    on the RPC pool and ``server.dispatch`` returns None — the pool worker
+    writes the response via the bound transport. These tests exercise the
+    handler's business logic, not the routing, so they drive the handler
+    inline while preserving the transport-binding semantics ``dispatch``
+    applies around a real request.
+    """
+    token = bind_transport(transport)
+    try:
+        return server.handle_request(req)
+    finally:
+        reset_transport(token)
 
 
 @pytest.fixture(autouse=True)
@@ -1234,10 +1252,10 @@ def test_voice_toggle_returns_configured_record_key(monkeypatch):
     # review on #19835).
     monkeypatch.setenv("HERMES_VOICE", "0")
 
-    on_resp = server.dispatch(
+    on_resp = _dispatch_sync(
         {"id": "voice-on", "method": "voice.toggle", "params": {"action": "on"}}
     )
-    status_resp = server.dispatch(
+    status_resp = _dispatch_sync(
         {"id": "voice-status", "method": "voice.toggle", "params": {"action": "status"}}
     )
 
@@ -1260,7 +1278,7 @@ def test_voice_toggle_on_carries_stop_hint(monkeypatch):
     )
     monkeypatch.setenv("HERMES_VOICE", "0")
 
-    on_resp = server.dispatch(
+    on_resp = _dispatch_sync(
         {"id": "voice-on", "method": "voice.toggle", "params": {"action": "on"}}
     )
     assert on_resp["result"]["stop_hint"] == 'Say "halt" to end the voice chat.'
@@ -1274,13 +1292,13 @@ def test_voice_toggle_on_carries_stop_hint(monkeypatch):
             voice_stop_hint=lambda: "",
         ),
     )
-    on_resp = server.dispatch(
+    on_resp = _dispatch_sync(
         {"id": "voice-on2", "method": "voice.toggle", "params": {"action": "on"}}
     )
     assert on_resp["result"]["stop_hint"] == ""
 
     # off carries no hint text (mode is ending).
-    off_resp = server.dispatch(
+    off_resp = _dispatch_sync(
         {"id": "voice-off", "method": "voice.toggle", "params": {"action": "off"}}
     )
     assert off_resp["result"]["stop_hint"] == ""
@@ -1306,7 +1324,7 @@ def test_voice_toggle_handles_non_dict_voice_cfg(monkeypatch):
     for bad in (True, "cmd+b", None, 42, ["ctrl+b"]):
         monkeypatch.setattr(server, "_load_cfg", lambda b=bad: {"voice": b})
 
-        status_resp = server.dispatch(
+        status_resp = _dispatch_sync(
             {
                 "id": "voice-status",
                 "method": "voice.toggle",
@@ -1325,7 +1343,7 @@ def test_voice_toggle_handles_non_dict_voice_cfg(monkeypatch):
     for bad_root in (True, None, [], "ctrl+b", 42):
         monkeypatch.setattr(server, "_load_cfg", lambda r=bad_root: r)
 
-        status_resp = server.dispatch(
+        status_resp = _dispatch_sync(
             {
                 "id": "voice-status-root",
                 "method": "voice.toggle",
@@ -1366,7 +1384,7 @@ def test_voice_record_start_handles_non_dict_voice_cfg(monkeypatch):
         captured.clear()
         monkeypatch.setattr(server, "_load_cfg", lambda b=bad: {"voice": b})
 
-        resp = server.dispatch(
+        resp = _dispatch_sync(
             {
                 "id": "voice-record",
                 "method": "voice.record",
@@ -1395,7 +1413,7 @@ def test_voice_record_start_handles_non_dict_voice_cfg(monkeypatch):
         captured.clear()
         monkeypatch.setattr(server, "_load_cfg", lambda c=bad_bool_cfg: {"voice": c})
 
-        resp = server.dispatch(
+        resp = _dispatch_sync(
             {
                 "id": "voice-record-bool",
                 "method": "voice.record",
@@ -1609,7 +1627,7 @@ def test_wake_owner_is_sticky_and_routes_detection_to_first_transport(monkeypatc
             "method": "wake.stop",
             "params": {},
         }, transport=second)
-        denied_voice_stop = server.dispatch({
+        denied_voice_stop = _dispatch_sync({
             "id": "voice-stop-2",
             "method": "voice.record",
             "params": {"action": "stop"},
@@ -1640,7 +1658,7 @@ def test_wake_owner_is_sticky_and_routes_detection_to_first_transport(monkeypatc
         )]
         assert state["paused"] is True
 
-        voice_started = server.dispatch({
+        voice_started = _dispatch_sync({
             "id": "voice-start-1",
             "method": "voice.record",
             "params": {"action": "start", "session_id": "first-session"},
@@ -1863,7 +1881,7 @@ def test_voice_record_start_forwards_max_recording_seconds(monkeypatch):
         captured.clear()
         monkeypatch.setattr(server, "_load_cfg", lambda c=cfg: {"voice": c})
 
-        resp = server.dispatch(
+        resp = _dispatch_sync(
             {
                 "id": "voice-record-cap",
                 "method": "voice.record",
@@ -1893,7 +1911,7 @@ def test_voice_record_stop_forces_transcription(monkeypatch):
         ),
     )
 
-    resp = server.dispatch(
+    resp = _dispatch_sync(
         {
             "id": "voice-record-stop",
             "method": "voice.record",
@@ -1916,7 +1934,7 @@ def test_voice_record_stop_updates_event_session_id(monkeypatch):
     )
     monkeypatch.setattr(server, "_voice_event_sid", "old-session")
 
-    resp = server.dispatch(
+    resp = _dispatch_sync(
         {
             "id": "voice-record-stop-session",
             "method": "voice.record",
@@ -1940,7 +1958,7 @@ def test_voice_record_start_reports_busy_when_stop_is_in_progress(monkeypatch):
     monkeypatch.setenv("HERMES_VOICE", "1")
     monkeypatch.setattr(server, "_load_cfg", lambda: {"voice": {}})
 
-    resp = server.dispatch(
+    resp = _dispatch_sync(
         {
             "id": "voice-record-busy",
             "method": "voice.record",
@@ -1978,7 +1996,7 @@ def test_voice_toggle_tts_branch_also_carries_record_key(monkeypatch):
     # later test in the file (which now spins up the streaming TTS pipeline).
     monkeypatch.setenv("HERMES_VOICE_TTS", "0")
 
-    tts_resp = server.dispatch(
+    tts_resp = _dispatch_sync(
         {"id": "voice-tts", "method": "voice.toggle", "params": {"action": "tts"}}
     )
 

tests/tui_gateway/test_protocol.py

@@ -658,6 +658,20 @@ def test_completion_handlers_are_pool_routed(completion_method, server):
     assert completion_method in server._LONG_HANDLERS
 
 
+@pytest.mark.parametrize("voice_method", ["voice.toggle", "voice.record", "voice.tts"])
+def test_voice_handlers_are_pool_routed(voice_method, server):
+    """Voice RPCs must run on the pool, never the WS reader thread.
+
+    Regression: voice.toggle (status) triggers check_voice_requirements() →
+    STT provider auto-detect → a SYNCHRONOUS faster-whisper lazy install (uv/pip
+    subprocess, up to a 300s timeout). Inline on the WS reader loop it blocked
+    prompt.submit / session.list frames queued behind it — the desktop showed
+    sent messages that never reached the agent. Same bug class as #21123 /
+    #50005: anything that can stall for seconds must stay off the reader thread.
+    """
+    assert voice_method in server._LONG_HANDLERS
+
+
 def test_skin_live_switch_end_to_end(server, tmp_path, monkeypatch):
     """Real config + skin files: activating a skin (as `hermes config set` does)
     makes the per-tool reconcile broadcast skin.changed with the resolved palette.

tui_gateway/server.py

@@ -274,6 +274,17 @@ _LONG_HANDLERS = frozenset(
         # the WS read loop and causing false "needs setup" (#50005 family).
         "setup.runtime_check",
         "setup.status",
+        # Voice RPCs can trigger check_voice_requirements() → STT provider
+        # auto-detect → a SYNCHRONOUS faster-whisper lazy install (uv/pip
+        # subprocess with a 300s timeout). Inline they stall the WS reader
+        # loop (handle_ws awaits dispatch before reading the next frame), so
+        # prompt.submit / session.list queued behind a voice.toggle sit
+        # unread and the desktop "send message" appears dead for minutes
+        # (reproduced: voice.toggle → session.list 40s+ timeout). Route them
+        # to the pool so a slow lazy install can't block message handling.
+        "voice.toggle",
+        "voice.record",
+        "voice.tts",
         # Desktop also polls the in-memory live-session registry every 15s.
         # The handler is normally cheap, but under heavy agent GIL pressure it
         # can still stall for tens of seconds. Keep it off the WS reader thread