llm: clean chain (Qwen gated off, Cloudflare off)

app/llm/router.py

@@ -129,8 +129,12 @@ def build_router_from_settings(settings: Any) -> LLMRouter:
         # model id; sits as a FALLBACK after gpt-oss-120b (which still benchmarks
         # best at instruction-following), giving model-family diversity in the
         # failover chain. The output guards protect regardless of which answers.
+        # OFF by default: Qwen3-235B 404s on this Cerebras account (the model isn't
+        # enabled there — verified live on both id forms). gpt-oss-120b is the best
+        # free instruction-follower anyway. Flip CEREBRAS_QWEN_ENABLED=1 (and enable
+        # the model in the Cerebras dashboard) to add it as a fallback later.
         qwen_model = getattr(settings, "cerebras_qwen_model", "qwen-3-235b-a22b")
-        if qwen_model:
+        if qwen_model and getattr(settings, "cerebras_qwen_enabled", False):
             providers["cerebras_qwen"] = providers["cerebras"]
             models["cerebras_qwen"] = {"small": qwen_model, "large": qwen_model}
             order.append("cerebras_qwen")

tests/llm/test_router.py

@@ -88,9 +88,24 @@ def test_build_router_with_three_providers():
         cerebras_api_key="cb", cerebras_base_url="https://api.cerebras.ai/v1", cerebras_model="gpt-oss-120b",
     )
     router = build_router_from_settings(s)
-    # cerebras now also exposes Qwen3 as an extra fallback (same client, Qwen model)
-    assert router.order == ["groq0", "cloudflare", "cerebras", "cerebras_qwen"]
-    assert len(router.providers) == 4
+    assert router.order == ["groq0", "cloudflare", "cerebras"]  # Qwen off by default
+    assert len(router.providers) == 3
+
+
+def test_build_router_adds_qwen_when_enabled():
+    from types import SimpleNamespace
+
+    from app.llm.router import build_router_from_settings
+
+    s = SimpleNamespace(
+        groq_api_key="k1", groq_api_keys=[], groq_base_url="u",
+        model_small="s", model_large="l",
+        cloudflare_account_id="", cloudflare_api_token="",
+        cerebras_api_key="cb", cerebras_base_url="https://api.cerebras.ai/v1",
+        cerebras_model="gpt-oss-120b", cerebras_qwen_enabled=True,
+    )
+    router = build_router_from_settings(s)
+    assert router.order == ["groq0", "cerebras", "cerebras_qwen"]
     assert router.models["cerebras_qwen"]["large"] == "qwen-3-235b-a22b"
 
 
@@ -110,9 +125,8 @@ def test_build_router_full_cascade_with_extra_free_providers():
         openrouter_api_key="ok", openrouter_base_url="https://openrouter.ai/api/v1", openrouter_model="x:free",
     )
     router = build_router_from_settings(s)
-    # full failover cascade: gpt-oss (groq, cerebras) -> Qwen3 fallback -> extras
-    assert router.order == ["groq0", "cloudflare", "cerebras", "cerebras_qwen",
-                            "sambanova", "mistral", "openrouter"]
+    # full failover cascade (Qwen3 fallback off by default; CF enabled here)
+    assert router.order == ["groq0", "cloudflare", "cerebras", "sambanova", "mistral", "openrouter"]
     assert router.models["mistral"]["large"] == "mistral-small-latest"
     assert router.models["sambanova"]["large"] == "Meta-Llama-3.3-70B-Instruct"