test: add llamacpp model-slot consistency and content-null tests

Cover the properties that prevent mid-turn model swaps in llama-swap
exclusive mode: vision_model defaults to primary, cloned local client
mirrors the user-selected model, embeddings stay on their own slot.
Also test the content:null serialization for tool-calling messages.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/ai/llamacpp.rs

@@ -985,4 +985,91 @@ mod tests {
         assert!(vision.has_vision);
         assert!(other.has_vision);
     }
+
+    #[test]
+    fn vision_model_defaults_to_primary() {
+        let c = LlamaCppClient::new(None, Some("qwen3:32b".into()));
+        assert_eq!(c.primary_model, "qwen3:32b");
+        assert_eq!(c.vision_model, "qwen3:32b");
+    }
+
+    #[test]
+    fn vision_model_explicit_override_diverges_from_primary() {
+        let mut c = LlamaCppClient::new(None, Some("qwen3:32b".into()));
+        c.set_vision_model("minicpm-v".into());
+        assert_eq!(c.primary_model, "qwen3:32b");
+        assert_eq!(c.vision_model, "minicpm-v");
+    }
+
+    #[test]
+    fn cloned_local_with_model_override_keeps_all_slots_consistent() {
+        // Simulates what resolve_backend does for the `local` client:
+        // clone the configured client, then override primary + vision
+        // to match the user-selected chat model. This prevents mid-turn
+        // model swaps in llama-swap exclusive mode.
+        let mut base = LlamaCppClient::new(None, Some("chat".into()));
+        base.set_vision_model("vision".into());
+        base.set_embedding_model("embed".into());
+
+        let mut local = base.clone();
+        let user_selected = "qwen3:32b";
+        local.primary_model = user_selected.to_string();
+        local.set_vision_model(user_selected.to_string());
+
+        // Chat generation (rerank) routes through primary_model.
+        assert_eq!(local.primary_model, user_selected);
+        // describe_image routes through vision_model.
+        assert_eq!(local.vision_model, user_selected);
+        // Embeddings stay on the dedicated slot — separate endpoint,
+        // no model swap conflict.
+        assert_eq!(local.embedding_model, "embed");
+    }
+
+    #[test]
+    fn assistant_tool_calls_emit_null_content() {
+        let msg = ChatMessage {
+            role: "assistant".into(),
+            content: String::new(),
+            tool_calls: Some(vec![ToolCall {
+                id: Some("call_1".into()),
+                function: ToolCallFunction {
+                    name: "search".into(),
+                    arguments: json!({}),
+                },
+            }]),
+            images: None,
+        };
+        let wire = LlamaCppClient::messages_to_openai(&[msg]);
+        assert!(wire[0]["content"].is_null(), "empty content + tool_calls should emit null");
+    }
+
+    #[test]
+    fn assistant_with_content_and_tool_calls_preserves_content() {
+        let msg = ChatMessage {
+            role: "assistant".into(),
+            content: "Let me search for that.".into(),
+            tool_calls: Some(vec![ToolCall {
+                id: Some("call_1".into()),
+                function: ToolCallFunction {
+                    name: "search".into(),
+                    arguments: json!({}),
+                },
+            }]),
+            images: None,
+        };
+        let wire = LlamaCppClient::messages_to_openai(&[msg]);
+        assert_eq!(wire[0]["content"], "Let me search for that.");
+    }
+
+    #[test]
+    fn assistant_without_tool_calls_keeps_empty_string_content() {
+        let msg = ChatMessage {
+            role: "assistant".into(),
+            content: String::new(),
+            tool_calls: None,
+            images: None,
+        };
+        let wire = LlamaCppClient::messages_to_openai(&[msg]);
+        assert_eq!(wire[0]["content"], "");
+    }
 }