fix(agent): URLs OpenAI-compatibles validees pour les fournisseurs LLM

Cohere -> /compatibility/v1, Google Gemini -> /v1beta/openai, Perplexity -> host racine, Chutes -> llm.chutes.ai/v1, SenseNova -> compatible-mode/v1. Mise a jour des modeles par defaut (xAI grok-4.6, Fireworks deepseek-v4-pro-0813, Cohere command-a-plus).

Retrait de LTX Studio (aucune API chat-completions) et MemTensor/MemOS (API memoire non OpenAI-compatible, auth Token + /chat). Google utilise desormais l'auth Bearer pour lister les modeles. Test de non-regression ajoute.

tests/test_agent.py

@@ -88,6 +88,24 @@ def test_llm_client_providers_listed(client):
         assert name in PROVIDERS
 
 
+def test_llm_client_openai_compatible_bases(client):
+    """Providers that need an OpenAI-compatible surface must point at it.
+
+    Anthropic/Google/Cohere native APIs are not chat-completions endpoints:
+    the client always speaks `/chat/completions` with a Bearer token, so the
+    base URLs must target the compatibility surfaces (validated live).
+    """
+    from app.services.llm_client import PROVIDERS
+    assert PROVIDERS["cohere"][0] == "https://api.cohere.ai/compatibility/v1"
+    assert PROVIDERS["google"][0] == (
+        "https://generativelanguage.googleapis.com/v1beta/openai")
+    # Perplexity serves /chat/completions at the host root (no /v1 prefix).
+    assert PROVIDERS["perplexity"][0] == "https://api.perplexity.ai"
+    # Chutes' OpenAI-compatible inference host.
+    assert PROVIDERS["chutes"][0] == "https://llm.chutes.ai/v1"
+    assert PROVIDERS["sensenova"][0].endswith("/compatible-mode/v1")
+
+
 def test_llm_client_offline_no_credentials(client):
     from app.services.llm_client import LLMClient
     llm = LLMClient(provider="openai", api_key="")