test: Coverage für Gender-/Stimm-System + hermetische Suite

Schließt die Test-Lücken der Gender-/Stimm-Arbeit der letzten Tage:
- test_routing.py: voice_for_route über alle Sprachen × Geschlechter
  (inkl. ES Multi-Speaker #0/#1, "any"→weiblich, Gegengeschlecht-Fallback,
  AR/PT/ZH Einzelstimmen, unbekannte Sprache → None).
- test_piper_tts.py: _parse_speaker (#id) + Gender-Code-Auflösung in
  synthesize() + speaker_id-Durchreichung für Multi-Speaker.
- test_cartesia_tts.py (neu): Voice-Map-Parsing, _resolve_voice_id-Präzedenz,
  nicht unterstützte Sprachen, direkte-UUID-Durchreichung, Cartesia→Piper-Fallback.

Hermetik-Fix in conftest.py: Deployment-Fallback-Ketten (stt/llm/tts_fallback)
werden pro Test auf "" gesetzt. Vorher schlugen 3 WS-Streaming-Tests fehl, weil
llm_fallback="openrouter" aus der TOML durchsickerte und Stub-LLMs ohne stream()
auf den echten OpenRouter durchfielen ("OPENROUTER_API_KEY is empty").

Ergebnis: 213 passed (vorher 164 passed / 3 failed). DEPLOYMENT.md Teil 4.0
dokumentiert den Pflicht-Testlauf vor jedem Deploy.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

tests/test_routing.py

@@ -64,3 +64,71 @@ def test_registry_unknown_provider_raises():
 
 def test_registry_known_provider():
     assert type(get_tts_provider("piper")).__name__ == "PiperTTSProvider"
+
+
+# ---------------------------------------------------------------------------
+# voice_for_route: Stimm-/Geschlechts-Auswahl je Provider und Sprache.
+# ---------------------------------------------------------------------------
+from app.dependencies import voice_for_route, PIPER_VOICE_GENDERED
+
+
+@pytest.mark.parametrize("lang,gender,expected", [
+    ("de", "f", "de_DE-kerstin-low"),
+    ("de", "m", "de_DE-thorsten-high"),
+    ("en", "f", "en_US-amy-medium"),
+    ("en", "m", "en_US-lessac-high"),
+    ("fr", "f", "fr_FR-siwis-medium"),
+    ("fr", "m", "fr_FR-tom-medium"),
+    ("es", "f", "es_ES-sharvard-medium#1"),   # Multi-Speaker: #1 = weiblich
+    ("es", "m", "es_ES-sharvard-medium#0"),   # Multi-Speaker: #0 = maennlich
+    ("it", "f", "it_IT-paola-medium"),
+    ("it", "m", "it_IT-riccardo-x_low"),
+    ("ru", "f", "ru_RU-irina-medium"),
+    ("ru", "m", "ru_RU-ruslan-medium"),
+    ("pl", "f", "pl_PL-gosia-medium"),
+    ("pl", "m", "pl_PL-darkman-medium"),
+])
+def test_piper_gendered_voice_exact(lang, gender, expected):
+    assert voice_for_route("piper", lang, gender) == expected
+
+
+@pytest.mark.parametrize("lang,expected", [
+    ("ar", "ar_JO-kareem-medium"),
+    ("pt", "pt_BR-faber-medium"),
+    ("zh", "zh_CN-huayan-medium"),
+])
+def test_piper_single_voice_languages_ignore_gender(lang, expected):
+    # Diese Sprachen haben keinen Geschlechtspartner -> selbe Stimme fuer m und f.
+    assert voice_for_route("piper", lang, "f") == expected
+    assert voice_for_route("piper", lang, "m") == expected
+
+
+def test_piper_any_prefers_female():
+    # 'any' (und unbekannte Codes) bevorzugen weiblich, konsistent mit Cartesia.
+    assert voice_for_route("piper", "de", "any") == "de_DE-kerstin-low"
+    assert voice_for_route("piper", "fr", "any") == "fr_FR-siwis-medium"
+
+
+def test_piper_unknown_language_returns_none():
+    assert voice_for_route("piper", "xx", "f") is None
+
+
+def test_piper_without_language_returns_none():
+    assert voice_for_route("piper", None, "f") is None
+
+
+def test_cartesia_passes_gender_code_through():
+    # Cartesia loest die UUID selbst auf -> nur der Gender-Code wird durchgereicht.
+    assert voice_for_route("cartesia", "de", "f") == "f"
+    assert voice_for_route("cartesia", "de", "any") == "any"
+
+
+def test_other_provider_returns_none():
+    assert voice_for_route("openrouter", "de", "f") is None
+
+
+def test_piper_falls_back_to_opposite_gender(monkeypatch):
+    # Fehlt die angeforderte Richtung, wird die vorhandene Gegenrichtung genommen
+    # (vor dem gender-agnostischen LANG_TO_PIPER_VOICE-Fallback).
+    monkeypatch.delitem(PIPER_VOICE_GENDERED, ("de", "m"), raising=False)
+    assert voice_for_route("piper", "de", "m") == "de_DE-kerstin-low"