feat(voice): start replies from stable streaming partials

tests/test_ai_orchestrator_service.py

@@ -1451,6 +1451,37 @@ def test_voice_v2_streaming_duplex_early_plan_returns_fast_safe_reply_without_ll
     assert "оператор" in decision["reply_text"].lower()
 
 
+def test_voice_v2_streaming_duplex_early_plan_returns_domain_followup_without_llm(monkeypatch):
+    monkeypatch.setenv("AI_VOICE_POLICY_MODE", "v2_streaming_duplex")
+
+    def _unexpected_llm(messages):
+        raise AssertionError(f"LLM should not be called for early plan: {messages!r}")
+
+    monkeypatch.setattr(ai_module, "_request_structured_model_decision", _unexpected_llm)
+
+    decision = voice_module._voice_decision(
+        language="ru",
+        customer=None,
+        interaction=SimpleNamespace(interaction_id="int_voice_early_schedule", status="new", queue_id="que_voice", subject="unknown"),
+        transcript_text="Мне надо узнать график работы",
+        transcript_window=[],
+        kb_results=[],
+        disclosure_required=False,
+        request_metadata={
+            "voice_v2_enabled": True,
+            "reply_phase": "early_plan",
+            "response_plan_id": "rsp_early_schedule",
+            "early_intent": "schedule",
+        },
+    )
+
+    assert decision["model"] == "voice_early_plan_domain"
+    assert decision["reply_text"]
+    assert decision["needs_handoff"] is False
+    assert decision["metadata"]["reply_phase"] == "early_plan"
+    assert decision["metadata"]["early_intent"] == "schedule"
+
+
 def test_voice_decision_hearing_check_keeps_active_topic_without_llm(monkeypatch):
     def _unexpected_llm(messages):
         raise AssertionError(f"LLM should not be called for hearing check: {messages!r}")