docs(api): correct adaptive chat completion cache doc (keyed by endpoint+model) (#1492)

The doc called the compatibility cache "model-independent", but cacheKey is
`${normalizedEndpoint}\n${model.trim()}` — learned compatibility is keyed by
endpoint AND model, not shared across models. The tests confirm it ("isolates
learned compatibility by endpoint and model"; LRU evicts after N
endpoint-model keys). Describe the actual behavior.

Author: PratikDhanave

apps/api/src/adaptive-chat-completions.ts

@@ -150,8 +150,9 @@ async function negotiate<TResponse extends ChatCompletionHttpResponse>(
 }
 
 /**
- * Sends a chat completion using a process-local, model-independent compatibility
- * cache. Only structured unsupported-parameter responses trigger negotiation.
+ * Sends a chat completion using a process-local compatibility cache keyed by
+ * endpoint and model. Only structured unsupported-parameter responses trigger
+ * negotiation.
  */
 export async function postAdaptiveChatCompletion<
   TResponse extends ChatCompletionHttpResponse,