fix(chatFormats): suppress content when it duplicates reasoning_content

Some upstream providers (e.g. certain OpenAI-compatible aggregators)
emit thinking/reasoning tokens with the same text written to both
delta.content and delta.reasoning_content simultaneously.

When metapi forwarded these chunks as-is, downstream clients would read
delta.content and display the internal reasoning text to end users —
appearing as word-by-word "thinking narration" in chat interfaces.

Fix: after resolving both contentDelta and reasoningDelta, suppress
contentDelta when it is non-empty and identical to reasoningDelta.
This covers the duplicate-write pattern without affecting any legitimate
case (a provider would never intentionally set content == reasoning).

src/server/routes/proxy/chatFormats.ts

@@ -736,7 +736,7 @@ export function normalizeUpstreamStreamEvent(
     const delta = isRecord(choice?.delta) ? choice.delta : {};
     const deltaParsed = extractTextAndReasoning(delta.content ?? delta);
 
-    const contentDelta =
+    const rawContentDelta =
       deltaParsed.content
       || (typeof choice?.message?.content === 'string' ? choice.message.content : '')
       || '';
@@ -747,6 +747,14 @@ export function normalizeUpstreamStreamEvent(
       || deltaParsed.reasoning
       || '';
 
+    // Some upstream providers (e.g. certain OpenAI-compatible aggregators) emit thinking
+    // tokens with the same text duplicated in both delta.content and delta.reasoning_content.
+    // When the two values are identical it means content is just echoing the reasoning —
+    // suppress it so internal thinking is never leaked to downstream consumers.
+    const contentDelta = (reasoningDelta && rawContentDelta === reasoningDelta)
+      ? ''
+      : rawContentDelta;
+
     const rawToolCalls = Array.isArray((delta as any).tool_calls)
       ? ((delta as any).tool_calls as unknown[])
       : [];