agent: code-atomiser-fix — test_llm_model respects caller GPU preference (Race-8)

Add an Option<bool> use_gpu parameter to test_llm_model with the same
default-true semantics as load_llm_model. Hard-coding true triggered an
engine tear-down/rebuild when a parallel load_llm_model(use_gpu=false)
was in flight (LlmEngine::load_model triples on id+path+use_gpu) and
silently flipped the user's GPU mode underneath the test.

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

src-tauri/src/commands/llm.rs

@@ -188,6 +188,7 @@ pub async fn test_llm_model(
     window: tauri::WebviewWindow,
     state: State<'_, AppState>,
     model_id: String,
+    use_gpu: Option<bool>,
 ) -> Result<LlmTestResult, String> {
     ensure_main_window(&window)?;
     let id = parse_model_id(model_id)?;
@@ -238,12 +239,19 @@ pub async fn test_llm_model(
         });
     }
 
-    // Not currently loaded. Attempt a real load (with GPU by default
-    // — matches load_llm_model's default) and classify any failure.
+    // Not currently loaded. Attempt a real load with the caller's
+    // GPU preference (matching load_llm_model's default of `true` when
+    // unspecified). Hard-coding `true` here used to race with a parallel
+    // load_llm_model(use_gpu=false) — LlmEngine::load_model decides
+    // "same model?" via (id, path, use_gpu) triple-equality and tears
+    // the engine down on mismatch, silently flipping the user's GPU
+    // mode underneath the in-flight test.
+    let resolved_use_gpu = use_gpu.unwrap_or(true);
     let engine = state.llm_engine.clone();
-    let load_result = tokio::task::spawn_blocking(move || engine.load_model(id, &path, true))
-        .await
-        .map_err(|e| e.to_string())?;
+    let load_result =
+        tokio::task::spawn_blocking(move || engine.load_model(id, &path, resolved_use_gpu))
+            .await
+            .map_err(|e| e.to_string())?;
 
     match load_result {
         Ok(()) => Ok(LlmTestResult {