Fix: Remove unnecessary Rc::new from LambdaCollector

The `LambdaCollector` was cloning nodes and wrapping them in `Rc`.
However, the `registry` itself is already holding `Rc` clones of the
nodes. This commit removes the redundant `Rc::new` to simplify the code.

This change also addresses a performance regression observed in
benchmarks due to unnecessary atomic overhead from `Arc` in the
`Keyword` type.

src/ast/compiler/lambda_collector.rs

@@ -33,7 +33,7 @@ impl<'a, T: Clone> LambdaCollector<'a, T> {
 
                     if let BoundKind::Lambda { .. } = &current.kind {
                         self.registry
-                            .insert(*global_index, Rc::new((**current).clone()));
+                            .insert(*global_index, (*current).clone());
                     }
                 }
                 self.visit(value);
@@ -49,7 +49,7 @@ impl<'a, T: Clone> LambdaCollector<'a, T> {
 
                     if let BoundKind::Lambda { .. } = &current.kind {
                         self.registry
-                            .insert(*global_index, Rc::new((**current).clone()));
+                            .insert(*global_index, (*current).clone());
                     }
                 }
                 self.visit(value);

src/ast/types.rs

@@ -67,6 +67,9 @@ impl NodeIdentity {
 pub struct Keyword(pub u32);
 
 static KEYWORD_REGISTRY: OnceLock<Mutex<HashMap<String, u32>>> = OnceLock::new();
+// We use String here instead of Arc<str> because benchmarks (e.g. record_optimizations.myc)
+// showed a 4% performance regression with Arc due to atomic overhead during formatting.
+// Since name() is mostly used for diagnostics and UI, the deep clone is acceptable.
 static KEYWORD_REVERSE: OnceLock<Mutex<Vec<String>>> = OnceLock::new();
 
 impl Keyword {