fix(agent): priorite a la generation de contenu dans le mock (contexte 'collection')

app/services/llm_client.py

@@ -214,6 +214,14 @@ class LLMClient:
                 model=model,
             )
 
+        # Inline content-generation ("Ask AI" / "AI meeting note") — answered
+        # before the tool-intent heuristics and scoped to the user objective
+        # only, so an injected "# Contexte" that happens to mention "collection"
+        # can't misroute a writing request into a create-collection action.
+        draft = self._draft_reply(user_content.split("\n# Contexte")[0])
+        if draft:
+            return LLMResponse(text=draft, model=model)
+
         # Exact keyword → tool intent resolution.
         for regex, builder in _CREATE_PATTERNS:
             m = regex.search(user_content)
@@ -247,10 +255,6 @@ class LLMClient:
                 )
 
         # Plain conversational objective → final answer (no tool).
-        draft = self._draft_reply(user_content)
-        if draft:
-            return LLMResponse(text=draft, model=model)
-
         return LLMResponse(
             text=self._summarize(user_content),
             model=model,