feat(voice): start replies from stable streaming partials

tests/test_ai_voice_media_runtime_streaming_stable.py

@@ -36,7 +36,7 @@ class _StubTTSProvider(TTSProvider):
 
 def test_media_runtime_voice_v2_uses_streaming_sidecar_for_partial_and_final_asr_stable():
     reply_starts: list[tuple[str, str | None, float]] = []
-    turns: list[str] = []
+    turns: list[tuple[str, str]] = []
 
     class _BatchASRProvider(ASRProvider):
         name = "batch-asr"
@@ -131,7 +131,7 @@ def test_media_runtime_voice_v2_uses_streaming_sidecar_for_partial_and_final_asr
         plan_reply=lambda session_id, text, metadata, kind: None,
         record_latency=lambda session_id, metric, latency_ms: None,
         process_turn=lambda session_id, transcript_text, language, barge_in, metadata: (
-            turns.append(transcript_text)
+            turns.append((str((metadata or {}).get("reply_phase") or "final"), transcript_text))
             or VoiceAITurnDecisionOut(
                 language=language or "ru",
                 intent="schedule",
@@ -191,7 +191,7 @@ def test_media_runtime_voice_v2_uses_streaming_sidecar_for_partial_and_final_asr
 
     phases = [phase for _, phase, _ in reply_starts]
     finalize_started_at = next(ts for name, ts in streaming_provider.events if name == "finalize")
-    assert turns == ["need work schedule"]
+    assert [text for phase, text in turns if phase == "final"] == ["need work schedule"]
     assert "open" in [name for name, _ in streaming_provider.events]
     assert "close" in [name for name, _ in streaming_provider.events]
     assert phases[:2] == ["ack", "main"]