feat: Audio-Streaming (chunked TTS) über WebSocket (#4 Ausbau)

- SentenceChunker (pipeline/sentence_chunker.py): inkrementelle Satzsegmentierung
- Orchestrator.chat_stream(on_audio): satzweise TTS, Audio-Chunk pro fertigem Satz;
  Gesamtaudio zusaetzlich an den Output-Endpunkt
- WS /ws/chat {"audio_stream":true}: audio-Events (json seq + binaerer Frame) live,
  kein finales Vollaudio; mit stream kombinierbar
- Tests: 45 gruen (+2: Sentence-Chunker, Audio-Streaming)
- Doku aktualisiert (README, BEDIENUNGSANLEITUNG, Architektur)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_streaming.py

@@ -5,10 +5,21 @@ from fastapi.testclient import TestClient
 import app.dependencies as deps
 from app.main import app
 from app.providers.llm.base import sse_delta, LLMProvider
+from app.pipeline.sentence_chunker import SentenceChunker
 
 client = TestClient(app)
 
 
+def test_sentence_chunker_incremental():
+    ch = SentenceChunker()
+    emitted = []
+    for tok in ["Hallo", " Anna", ". ", "Wie", " geht", " es", "? ", "Tschuess"]:
+        emitted += ch.feed(tok)
+    assert emitted == ["Hallo Anna.", "Wie geht es?"]
+    assert ch.flush() == "Tschuess"
+    assert SentenceChunker().feed("Eins. Zwei! Drei? Vier") == ["Eins.", "Zwei!", "Drei?"]
+
+
 def test_sse_delta_parsing():
     assert sse_delta('data: {"choices":[{"delta":{"content":"Hal"}}]}') == "Hal"
     assert sse_delta("data: [DONE]") is None
@@ -92,3 +103,42 @@ def test_ws_stream_fallback_for_nonstreaming_llm(monkeypatch):
         token = ws.receive_json()
         assert token["type"] == "token" and token["text"] == "komplett"
         assert ws.receive_json()["type"] == "semantic"
+
+
+def test_ws_audio_stream_sends_chunks_per_sentence(monkeypatch):
+    tts_calls = []
+
+    class StreamLLM:
+        async def complete(self, text, history=None, session_id=None):
+            return "Satz eins. Satz zwei."
+
+        async def stream(self, text, history=None, session_id=None):
+            for tok in ["Satz ", "eins. ", "Satz ", "zwei."]:
+                yield tok
+
+    class CountTTS:
+        async def synthesize(self, text, voice=None, audio_format="pcm"):
+            tts_calls.append(text)
+            return b"A" * len(tts_calls)
+
+    monkeypatch.setitem(deps.LLM_REGISTRY, "stream", lambda s: StreamLLM())
+    monkeypatch.setitem(deps.TTS_REGISTRY, "cnt", lambda s: CountTTS())
+    base = {"llm_provider": "stream", "tts_provider": "cnt", "output_endpoint": "loopback"}
+
+    with client.websocket_connect("/ws/chat") as ws:
+        ws.send_json({"text": "x", "audio_stream": True, **base})
+        assert ws.receive_json()["type"] == "ack"
+
+        audio_events = 0
+        event = ws.receive_json()
+        while event["type"] != "semantic":
+            assert event["type"] == "audio"
+            assert ws.receive_bytes()  # binärer Audio-Chunk folgt
+            audio_events += 1
+            event = ws.receive_json()
+
+        assert audio_events == 2  # zwei Sätze -> zwei Chunks
+        # Kein finales Vollaudio mehr -> direkt done.
+        assert ws.receive_json()["type"] == "done"
+
+    assert len(tts_calls) == 2  # TTS pro Satz