fix: sanitize response based on sourceFormat, not targetFormat

open-sse/handlers/chatCore.ts

@@ -1690,7 +1690,7 @@ export async function handleChatCore({
     // Strips non-standard fields (x_groq, usage_breakdown, service_tier, etc.)
     // Extracts <think> and <thinking> tags into reasoning_content
     // Target format determines output shape. If we are outputting OpenAI shape or pseudo-OpenAI shape, sanitize.
-    if (targetFormat === FORMATS.OPENAI || targetFormat === FORMATS.OPENAI_RESPONSES) {
+    if (sourceFormat === FORMATS.OPENAI || sourceFormat === FORMATS.OPENAI_RESPONSES) {
       translatedResponse = sanitizeOpenAIResponse(translatedResponse);
     }