fix: address 25 PR review comments — SiLU zero-alloc, arena safety, cleanup

SiLU derivative optimization (18 files):
- Replaced ones_t tensor allocation + Fill(One) + TensorAdd pattern with
  Engine.TensorAddScalar(x, NumOps.One) — zero allocation for 1+x computation
- ABCLayer, GatedDeltaNet, GatedDeltaProduct, GatedSlotAttention, HGRN2,
  Hedgehog, Hyena, Kimi, LogLinear, MesaNet, MixtureOfMamba/Memories,
  MultiLatentAttention, Rebased, RetNet, Rodimus, TransNormerLLM, TTT

ForwardArena fixes:
- Reset() uses pre-cached keys array instead of allocating List each call
- ShapeKey defensive-copies the shape array to prevent mutation
- Made ForwardArena internal (implementation detail, not public API)

Other fixes:
- RWKV7Block: removed unused TsYt constant
- DiffusionConvLayer: removed duplicate using directive
- S6Scan: eliminated wasted Rent in initialState branch (clone instead)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: ooples

src/Memory/ForwardArena.cs

@@ -11,10 +11,11 @@ namespace AiDotNet.Memory;
 /// and GC pressure during the forward pass. Tensors are pre-created during warmup
 /// and recycled across calls.
 /// </summary>
-public sealed class ForwardArena<T>
+internal sealed class ForwardArena<T>
 {
     private readonly Dictionary<ShapeKey, Tensor<T>[]> _slabs = new();
     private readonly Dictionary<ShapeKey, int> _cursors = new();
+    private ShapeKey[]? _cursorKeysCache;
     private const int DefaultSlabSize = 4;
     private const int GrowthFactor = 2;
 
@@ -71,9 +72,12 @@ public Tensor<T> RentUninitialized(int[] shape)
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public void Reset()
     {
-        var keys = new List<ShapeKey>(_cursors.Keys);
-        foreach (var key in keys)
-            _cursors[key] = 0;
+        // Use cached keys array to avoid allocation during reset
+        if (_cursorKeysCache is null || _cursorKeysCache.Length != _cursors.Count)
+            _cursorKeysCache = new ShapeKey[_cursors.Count];
+        _cursors.Keys.CopyTo(_cursorKeysCache, 0);
+        for (int i = 0; i < _cursorKeysCache.Length; i++)
+            _cursors[_cursorKeysCache[i]] = 0;
     }
 
     /// <summary>
@@ -136,7 +140,7 @@ private Tensor<T> GrowAndRent(ShapeKey key, int[] shape, bool clear = true)
 
     public ShapeKey(int[] shape)
     {
-        _dims = shape;
+        _dims = (int[])shape.Clone();
         unchecked
         {
             int hash = (int)2166136261;

src/NeuralNetworks/Layers/DiffusionConvLayer.cs

@@ -5,7 +5,6 @@
 using AiDotNet.Enums.AlgorithmTypes;
 using AiDotNet.Interfaces;
 using AiDotNet.Memory;
-using AiDotNet.Interfaces;
 using AiDotNet.Tensors.Engines;
 using AiDotNet.Tensors.Engines.DirectGpu;
 using AiDotNet.Tensors.Engines.Gpu;

src/NeuralNetworks/Layers/SSM/ABCLayer.cs

@@ -478,7 +478,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var onePlusXSig = Engine.TensorAdd(CreateOnesLike(xTimesOneMinusSig), xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/GatedDeltaNetLayer.cs

@@ -470,7 +470,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var onePlusXSig = Engine.TensorAdd(CreateOnesLike(xTimesOneMinusSig), xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/GatedDeltaProductLayer.cs

@@ -609,7 +609,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var onePlusXSig = Engine.TensorAdd(CreateOnesLike(xTimesOneMinusSig), xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/GatedSlotAttentionLayer.cs

@@ -446,7 +446,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var onePlusXSig = Engine.TensorAdd(CreateOnesLike(xTimesOneMinusSig), xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/HGRN2Layer.cs

@@ -392,7 +392,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var onePlusXSig = Engine.TensorAdd(CreateOnesLike(xTimesOneMinusSig), xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/HedgehogLayer.cs

@@ -576,8 +576,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var ones_t = new Tensor<T>(xTimesOneMinusSig.Shape.ToArray()); ones_t.Fill(NumOps.One);
-        var onePlusXSig = Engine.TensorAdd(ones_t, xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/HyenaLayer.cs

@@ -479,8 +479,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var onesT = new Tensor<T>(xTimesOneMinusSig.Shape.ToArray()); onesT.Fill(NumOps.One);
-        var onePlusXSig = Engine.TensorAdd(onesT, xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }
 

src/NeuralNetworks/Layers/SSM/KimiLinearAttentionLayer.cs

@@ -373,8 +373,7 @@ private Tensor<T> ComputeSiLUDerivative(Tensor<T> x)
         var sig = Engine.Sigmoid(x);
         var oneMinusSig = Engine.ScalarMinusTensor(NumOps.One, sig);
         var xTimesOneMinusSig = Engine.TensorMultiply(x, oneMinusSig);
-        var ones_t = new Tensor<T>(xTimesOneMinusSig.Shape.ToArray()); ones_t.Fill(NumOps.One);
-        var onePlusXSig = Engine.TensorAdd(ones_t, xTimesOneMinusSig);
+        var onePlusXSig = Engine.TensorAddScalar(xTimesOneMinusSig, NumOps.One);
         return Engine.TensorMultiply(sig, onePlusXSig);
     }