fix(tts): OpenRouter-TTS bei transient leerer/5xx-Antwort wiederholen

Preview-Modelle liefern gelegentlich HTTP 200 mit leerem Body (oder 5xx) -> bisher
brach das die ganze Sprech-Runde ab ('empty audio content', 502). Jetzt bis zu 3
Versuche mit linearem Backoff; 4xx (z. B. ungueltige Stimme) wird sofort gemeldet
(kein Retry). Damit klappt das Durchprobieren von Stimmen zuverlaessig (Live-Test:
alle 20 getesteten Gemini-Stimmen inkl. Leda liefern Audio).
Tests: tests/test_openrouter_tts.py (leer->ok, 4xx sofort, alles-leer, 5xx->ok).
Doku-Hinweis in BEDIENUNGSANLEITUNG.

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

tests/test_openrouter_tts.py

@@ -0,0 +1,70 @@
+"""Tests fuer das Retry-Verhalten des OpenRouter-TTS (transiente leere/5xx-Antworten)."""
+
+import asyncio
+
+import httpx
+import pytest
+
+import app.providers.tts.openrouter as orm
+from app.providers.tts.openrouter import OpenRouterTTSProvider
+
+REQ = httpx.Request("POST", "https://openrouter.ai/api/v1/audio/speech")
+
+
+def _run(coro):
+    return asyncio.run(coro)
+
+
+def _patch(monkeypatch, responses):
+    """httpx-POST liefert nacheinander die vorgegebenen Antworten; kein echtes Sleep."""
+    it = iter(responses)
+    calls = {"n": 0}
+
+    async def fake_post(self, url, **kw):
+        calls["n"] += 1
+        return next(it)
+
+    async def no_sleep(*a, **k):
+        pass
+
+    monkeypatch.setattr(httpx.AsyncClient, "post", fake_post)
+    monkeypatch.setattr(orm.asyncio, "sleep", no_sleep)
+    return calls
+
+
+def _provider():
+    return OpenRouterTTSProvider("key", "some/model", "Zephyr")
+
+
+def test_retries_on_empty_then_succeeds(monkeypatch):
+    calls = _patch(monkeypatch, [
+        httpx.Response(200, content=b"", request=REQ),       # transienter Aussetzer
+        httpx.Response(200, content=b"AUDIO", request=REQ),  # dann echtes Audio
+    ])
+    assert _run(_provider().synthesize("Hallo")) == b"AUDIO"
+    assert calls["n"] == 2
+
+
+def test_4xx_raises_immediately_without_retry(monkeypatch):
+    calls = _patch(monkeypatch, [
+        httpx.Response(400, text="invalid voice", request=REQ),
+    ])
+    with pytest.raises(RuntimeError, match="400"):
+        _run(_provider().synthesize("Hallo"))
+    assert calls["n"] == 1  # kein Retry bei 4xx
+
+
+def test_all_empty_raises_after_attempts(monkeypatch):
+    calls = _patch(monkeypatch, [httpx.Response(200, content=b"", request=REQ)] * orm._MAX_ATTEMPTS)
+    with pytest.raises(RuntimeError, match="empty audio content"):
+        _run(_provider().synthesize("Hallo"))
+    assert calls["n"] == orm._MAX_ATTEMPTS
+
+
+def test_5xx_then_success(monkeypatch):
+    calls = _patch(monkeypatch, [
+        httpx.Response(503, text="upstream", request=REQ),
+        httpx.Response(200, content=b"OK", request=REQ),
+    ])
+    assert _run(_provider().synthesize("Hallo")) == b"OK"
+    assert calls["n"] == 2