fix(coach): use nova-2-general for ar/tr STT to avoid 400 errors

backend/server/api/coach/transcribe.post.ts

@@ -83,31 +83,39 @@ export default defineEventHandler(async (event) => {
     buffer.length,
   );
 
-  // Deepgram language mapping. WICHTIG: model=nova-3, NICHT nova-2 —
-  // nova-2 unterstützt weder Arabisch (ar) noch Türkisch (tr) und antwortet
-  // dafür mit 400 "No such model/language/tier combination" → leeres
-  // Transcript / Fehler beim Client ("kein Text nach Speech"). nova-3 deckt
-  // alle hier gelisteten Sprachen als diskrete Codes ab (inkl. ar + tr),
-  // ohne Regression für de/en/fr/es/pt/it. Verifiziert gg. Deepgram-Doc
-  // models-languages-overview (2026-05).
+  // Deepgram language mapping.
+  // Live-Diagnose (2026-05-30): nova-3 lehnt language=ar (und tr) mit
+  // 400 "No such model/language/tier combination found" ab — entgegen
+  // der vorherigen Annahme. Fallback für ar/tr: nova-2-general
+  // (multilingual auto-detect). Für alle anderen Sprachen bleibt nova-3
+  // (bessere Genauigkeit, diskrete language-codes).
   const deepgramLang =
     language &&
     ["de", "en", "tr", "ar", "fr", "es", "pt", "it"].includes(language)
       ? language
       : "de";
 
+  const needsGeneralModel = ["ar", "tr"].includes(deepgramLang);
+  const deepgramUrl = needsGeneralModel
+    ? "https://api.deepgram.com/v1/listen?model=nova-2-general"
+    : `https://api.deepgram.com/v1/listen?language=${deepgramLang}&model=nova-3`;
+
+  console.log(
+    "[transcribe] language:",
+    deepgramLang,
+    "model:",
+    needsGeneralModel ? "nova-2-general" : "nova-3",
+  );
+
   try {
-    const response = await fetch(
-      `https://api.deepgram.com/v1/listen?language=${deepgramLang}&model=nova-3`,
-      {
+    const response = await fetch(deepgramUrl, {
       method: "POST",
       headers: {
         Authorization: `Token ${config.deepgramApiKey}`,
         "Content-Type": blobType,
       },
       body: buffer,
-      },
-    );
+    });
 
     const result = await response.json();