chore: no need to arc wrap client (#4741)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

lib/bindings/c/src/lib.rs

@@ -1031,7 +1031,7 @@ pub async fn create_worker_selection_pipeline_chat(
 
     // Create worker monitor if busy_threshold is set
     // Note: C bindings don't register with ModelManager, so HTTP endpoint won't see this
-    let worker_monitor = busy_threshold.map(|t| KvWorkerMonitor::new(Arc::new(client.clone()), t));
+    let worker_monitor = busy_threshold.map(|t| KvWorkerMonitor::new(client.clone(), t));
 
     let engine = build_routed_pipeline::<
         NvCreateChatCompletionRequest,

lib/llm/src/discovery/model_manager.rs

@@ -531,7 +531,7 @@ impl ModelManager {
     pub fn get_or_create_worker_monitor(
         &self,
         model: &str,
-        client: Arc<Client>,
+        client: Client,
         threshold: f64,
     ) -> KvWorkerMonitor {
         let mut monitors = self.worker_monitors.write();

lib/llm/src/discovery/watcher.rs

@@ -405,11 +405,8 @@ impl ModelWatcher {
             // Get or create the worker monitor for this model
             // This allows dynamic threshold updates via the ModelManager
             let worker_monitor = self.router_config.busy_threshold.map(|threshold| {
-                self.manager.get_or_create_worker_monitor(
-                    card.name(),
-                    Arc::new(client.clone()),
-                    threshold,
-                )
+                self.manager
+                    .get_or_create_worker_monitor(card.name(), client.clone(), threshold)
             });
 
             // Add chat engine only if the model supports chat

lib/llm/src/discovery/worker_monitor.rs

@@ -55,11 +55,11 @@ impl WorkerLoadState {
 
 /// Worker monitor for tracking KV cache usage and busy states.
 ///
-/// All fields are `Arc`, so cloning shares state. This allows multiple pipelines
+/// Cloning shares state via internal Arc-wrapped fields. This allows multiple pipelines
 /// (e.g., chat and completions) to share the same monitor instance.
 #[derive(Clone)]
 pub struct KvWorkerMonitor {
-    client: Arc<Client>,
+    client: Client,
     worker_load_states: Arc<RwLock<HashMap<u64, WorkerLoadState>>>,
     /// Threshold stored as parts-per-10000 (e.g., 8500 = 0.85)
     busy_threshold: Arc<AtomicU32>,
@@ -72,7 +72,7 @@ impl KvWorkerMonitor {
     ///
     /// The threshold (0.0-1.0) controls when workers are considered busy based on
     /// KV cache utilization. It can be dynamically updated via `set_threshold()`.
-    pub fn new(client: Arc<Client>, threshold: f64) -> Self {
+    pub fn new(client: Client, threshold: f64) -> Self {
         Self {
             client,
             worker_load_states: Arc::new(RwLock::new(HashMap::new())),