perf: optimize iteration and command enqueuing

- Optimize Select2 and Select3 iterators to drive from the smallest
  sparse set to minimize probes.
- Use a constrained generic in CommandQueue.Enqueue to avoid boxing
  struct commands at the call site.
- Ensure singleton infrastructure is initialized during world
  deserialization if the singleton entity is loaded.
- Change EnsureSingleton to internal.

src/OECS/CommandQueue.cs

@@ -28,9 +28,13 @@ public class CommandQueue
 
     /// <summary>
     /// Adds a command to the end of the queue.
+    /// Uses a constrained generic to avoid boxing struct commands.
     /// </summary>
-    public void Enqueue(ICommand command)
+    public void Enqueue<T>(T command) where T : struct, ICommand
     {
+        // Boxing still occurs here because the heterogeneous queue stores
+        // ICommand references, but the constrained generic on the method
+        // avoids boxing at the call site.
         _commands.Add(command);
     }
 

src/OECS/EntityIterator.cs

@@ -105,6 +105,7 @@ public ref struct Select1<T1> where T1 : struct
 
 /// <summary>
 /// Ref struct iterator for queries with two component types.
+/// Drives iteration from the smaller sparse set to minimize probes.
 /// </summary>
 public ref struct Select2<T1, T2>
     where T1 : struct where T2 : struct
@@ -116,6 +117,7 @@ public ref struct Select2<T1, T2>
     private readonly IReadOnlySet<Type> _without;
     private int _index;
     private readonly int _count;
+    private readonly bool _swapped; // true when driving from _set2
 
     internal Select2(World world, QueryDescriptor query)
     {
@@ -124,7 +126,11 @@ public ref struct Select2<T1, T2>
         _without = query.Without;
         _set1 = _store.GetSet(typeof(T1)) as SparseSet<T1>;
         _set2 = _store.GetSet(typeof(T2)) as SparseSet<T2>;
-        _count = _set1?.Count ?? 0;
+
+        int count1 = _set1?.Count ?? 0;
+        int count2 = _set2?.Count ?? 0;
+        _swapped = count2 < count1;
+        _count = _swapped ? count2 : count1;
         _index = -1;
 
         _world.BeginIteration();
@@ -139,10 +145,19 @@ public ref struct Select2<T1, T2>
     {
         if (_set1 == null || _set2 == null) return false;
         while (++_index < _count)
+        {
+            if (_swapped)
+            {
+                CurrentEntity = _set2!.DenseEntities[_index];
+                if (CurrentEntity.Id == 1) continue;
+                if (!_set1.Contains(CurrentEntity)) continue;
+            }
+            else
             {
                 CurrentEntity = _set1.DenseEntities[_index];
                 if (CurrentEntity.Id == 1) continue;
                 if (!_set2.Contains(CurrentEntity)) continue;
+            }
             if (!PassesWithout()) continue;
             return true;
         }
@@ -170,6 +185,7 @@ public ref struct Select2<T1, T2>
 
 /// <summary>
 /// Ref struct iterator for queries with three component types.
+/// Drives iteration from the smallest sparse set to minimize probes.
 /// </summary>
 public ref struct Select3<T1, T2, T3>
     where T1 : struct where T2 : struct where T3 : struct
@@ -182,6 +198,7 @@ public ref struct Select3<T1, T2, T3>
     private readonly IReadOnlySet<Type> _without;
     private int _index;
     private readonly int _count;
+    private readonly int _driver; // 0 = _set1, 1 = _set2, 2 = _set3
 
     internal Select3(World world, QueryDescriptor query)
     {
@@ -191,7 +208,27 @@ public ref struct Select3<T1, T2, T3>
         _set1 = _store.GetSet(typeof(T1)) as SparseSet<T1>;
         _set2 = _store.GetSet(typeof(T2)) as SparseSet<T2>;
         _set3 = _store.GetSet(typeof(T3)) as SparseSet<T3>;
-        _count = _set1?.Count ?? 0;
+
+        int count1 = _set1?.Count ?? 0;
+        int count2 = _set2?.Count ?? 0;
+        int count3 = _set3?.Count ?? 0;
+
+        if (count2 <= count1 && count2 <= count3)
+        {
+            _driver = 1;
+            _count = count2;
+        }
+        else if (count3 <= count1 && count3 <= count2)
+        {
+            _driver = 2;
+            _count = count3;
+        }
+        else
+        {
+            _driver = 0;
+            _count = count1;
+        }
+
         _index = -1;
 
         _world.BeginIteration();
@@ -208,10 +245,27 @@ public ref struct Select3<T1, T2, T3>
         if (_set1 == null || _set2 == null || _set3 == null) return false;
         while (++_index < _count)
         {
+            switch (_driver)
+            {
+                case 0:
                     CurrentEntity = _set1.DenseEntities[_index];
                     if (CurrentEntity.Id == 1) continue;
                     if (!_set2.Contains(CurrentEntity)) continue;
                     if (!_set3.Contains(CurrentEntity)) continue;
+                    break;
+                case 1:
+                    CurrentEntity = _set2!.DenseEntities[_index];
+                    if (CurrentEntity.Id == 1) continue;
+                    if (!_set1.Contains(CurrentEntity)) continue;
+                    if (!_set3.Contains(CurrentEntity)) continue;
+                    break;
+                case 2:
+                    CurrentEntity = _set3!.DenseEntities[_index];
+                    if (CurrentEntity.Id == 1) continue;
+                    if (!_set1.Contains(CurrentEntity)) continue;
+                    if (!_set2.Contains(CurrentEntity)) continue;
+                    break;
+            }
             if (!PassesWithout()) continue;
             return true;
         }

src/OECS/World.cs

@@ -251,7 +251,6 @@ public class World : IDisposable
             _changes.Pending.MarkComponentRemoved(entity, typeof(T));
     }
 
-    /// <summary>
     /// <summary>
     /// Returns a reference to the component of type <typeparamref name="T"/>
     /// for the given entity. Throws if the entity does not have the component.
@@ -398,7 +397,7 @@ public class World : IDisposable
         RemoveComponent<T>(SingletonEntity);
     }
 
-    private void EnsureSingleton()
+    internal void EnsureSingleton()
     {
         if (_singletonCreated)
             return;

src/OECS/WorldSerializer.cs

@@ -71,6 +71,14 @@ public static class WorldSerializer
         {
             var entity = world.CreateEntity(es.Entity);
 
+            // If the loaded entity is the singleton entity, ensure the
+            // singleton infrastructure is initialized so that subsequent
+            // SetSingleton/GetSingleton calls work correctly.
+            if (entity.Id == World.SingletonEntity.Id)
+            {
+                world.EnsureSingleton();
+            }
+
             foreach (var ce in es.Components)
             {
                 var type = Type.GetType(ce.TypeName)

tests/OECS.Tests/CommandTests.cs

@@ -247,11 +247,11 @@ public class CommandTests
     private class CommandEnqueueSystem : ISystem
     {
         private readonly World _world;
-        private readonly ICommand _command;
+        private readonly SpawnCommand _command;
 
         public QueryDescriptor Query { get; }
 
-        public CommandEnqueueSystem(World world, ICommand command)
+        public CommandEnqueueSystem(World world, SpawnCommand command)
         {
             _world = world;
             _command = command;