Address windowed pow2 review feedback

PerfectNumbers.Core/Gpu/Kernels/DivisorKernels.cs

@@ -21,7 +21,7 @@ internal static class DivisorKernels
         ReadOnlyGpuUInt128 readOnlyMod = divisor;
         if (x == 0)
         {
-            baseVal = GpuUInt128.Pow2Mod(exponent, in readOnlyMod); // Uses the eight-bit windowed ladder measured fastest in GPU benchmarks.
+            baseVal = GpuUInt128.Pow2Mod(exponent, in readOnlyMod);
             result[0] = baseVal.High == 0UL && baseVal.Low == 1UL ? (byte)1 : (byte)0;
             return;
         }
@@ -38,7 +38,7 @@ internal static class DivisorKernels
 
         GpuUInt128 pow = baseVal;
         pow.ModPow(exponent / ux, in readOnlyMod);
-        var part2 = GpuUInt128.Pow2Mod(exponent % ux, in readOnlyMod); // Windowed pow2 keeps the mixed-radix tail fast.
+        var part2 = GpuUInt128.Pow2Mod(exponent % ux, in readOnlyMod);
         GpuUInt128 product = pow;
         ReadOnlyGpuUInt128 part2ReadOnly = part2.AsReadOnly();
         product.MulMod(in part2ReadOnly, in readOnlyMod);

PerfectNumbers.Core/Gpu/Kernels/IncrementalKernels.cs

@@ -96,7 +96,7 @@ internal static class IncrementalKernels
         }
 
         ulong phi64 = phi.Low;
-        if (GpuUInt128.Pow2Mod(phi64, in readOnlyQ) != GpuUInt128.One) // Windowed pow2 keeps the totient check fast.
+        if (GpuUInt128.Pow2Mod(phi64, in readOnlyQ) != GpuUInt128.One)
         {
             orders[index] = 0UL;
             return;
@@ -192,7 +192,7 @@ internal static class IncrementalKernels
         }
 
         ulong phi64 = phi.Low;
-        if (GpuUInt128.Pow2Mod(phi64, in readOnlyQ) != GpuUInt128.One) // Windowed pow2 keeps the order probe aligned with benchmarks.
+        if (GpuUInt128.Pow2Mod(phi64, in readOnlyQ) != GpuUInt128.One)
         {
             return;
         }

PerfectNumbers.Core/Gpu/Kernels/OrderKernels.cs

@@ -22,7 +22,7 @@ internal static class OrderKernels
         }
 
         ulong phi64 = phi.Low;
-        GpuUInt128 pow = GpuUInt128.Pow2Mod(phi64, in readOnlyQ); // Windowed pow2 keeps the cycle check aligned with benchmark wins.
+        GpuUInt128 pow = GpuUInt128.Pow2Mod(phi64, in readOnlyQ);
         if (pow != GpuUInt128.One)
         {
             orders[index] = 0UL;

PerfectNumbers.Core/Gpu/Kernels/Pow2ModKernels.cs

@@ -77,7 +77,6 @@ internal static class Pow2ModKernels
                 return;
             }
         }
-        // Pow2Minus1Mod already routes through the windowed pow2 helper benchmarked as fastest for large divisors.
         if (GpuUInt128.Pow2Minus1Mod(exponent, in readOnlyQ) != GpuUInt128.Zero)
         {
             orders[index] = 0UL;
@@ -169,7 +168,7 @@ internal static class Pow2ModKernels
                 return;
             }
         }
-        if (GpuUInt128.Pow2Mod(exponent, in readOnlyQ) != GpuUInt128.One) // Windowed pow2 keeps the order guard aligned with GPU benchmarks.
+        if (GpuUInt128.Pow2Mod(exponent, in readOnlyQ) != GpuUInt128.One)
         {
             return;
         }

PerfectNumbers.Core/Gpu/MersenneNumberLucasLehmerGpuTester.cs

@@ -112,7 +112,7 @@ public class MersenneNumberLucasLehmerGpuTester
 
         var gpu = GpuContextPool.RentPreferred(preferCpu: false);
         var accelerator = gpu.Accelerator;
-        var kernel = GetBatchKernel(accelerator); // Batch kernel already relies on the Mersenne squaring reducer; windowed pow2 helpers do not apply here.
+        var kernel = GetBatchKernel(accelerator);
 
         var expBuffer = accelerator.Allocate1D<ulong>(count);
         ulong[] expArray = ArrayPool<ulong>.Shared.Rent(count);
@@ -479,7 +479,6 @@ public class MersenneNumberLucasLehmerGpuTester
         {
             if ((exponent & 1UL) != 0UL)
             {
-                // The base varies per primitive-root candidate, so pow2-specific ladders do not apply.
                 result = MulMod(result, value, modulus);
             }
 

PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs

@@ -22,9 +22,7 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
                 {
                         GpuKernelType.Pow2Mod => gpuLease.Pow2ModOrderKernel,
                         _ => gpuLease.IncrementalOrderKernel,
-                }; // Pow2ModOrderKernel already routes through the windowed pow2 helper measured fastest for large divisors.
-                // Both GPU order kernels already rely on the windowed pow2 helper, so this wrapper only selects the delegate
-                // and manages the scoped GPU resources.
+                };
 
                 var foundBuffer = accelerator.Allocate1D<int>(1); // TODO: Replace this allocation with the pooled device buffer from GpuOrderKernelBenchmarks so repeated scans reuse the pinned staging memory instead of allocating per run.
                 exponent.Mod10_8_5_3Steps(out ulong step10, out ulong step8, out ulong step5, out ulong step3);

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -985,8 +985,6 @@ public class MersenneDivisorCycles
     private static ulong CalculateCycleLengthFallback(ulong divisor)
     {
         // Otherwise, find order of 2 mod divisor using the shared unrolled-hex stepping sequence.
-        // Divisors <= 3 hit this fallback, so the lightweight CPU path already matches the windowed pow2 helper
-        // without paying for a GPU round-trip.
         return divisor.CalculateMersenneDivisorCycleLengthUnrolledHexCpu();
     }
 

PerfectNumbers.Core/MersenneNumberTester.cs

@@ -319,8 +319,6 @@ public sealed class MersenneNumberTester(
         private static void LegacyOrderKernel(Index1D index, ulong exponent, ArrayView<UInt128> qs, ArrayView<ulong> orders)
         {
                 UInt128 q = qs[index];
-                // The legacy kernel already relies on the adaptive PowMod helper, which switches to the windowed pow2
-                // ladder for large divisors.
                 UInt128 pow = exponent.PowMod(q);
                 orders[index] = pow == 1UL ? exponent : 0UL;
         }

PerfectNumbers.Core/MersenneResidueStepper.cs

@@ -17,10 +17,15 @@ public class MersenneResidueStepper
         m = modulus;
         p = p0;
 
-        // Callers hand this stepper pre-reduced residues from the windowed pow2 pipeline; keep the modulo guard for tests
-        // while production inputs already arrive inside [0, m).
-        M_mod = M_mod_m_at_p0 % m;
-        pow2_mod = (M_mod + 1) % m;
+        // Inputs are already reduced in production; keep the old modulo guard commented for debugging reference.
+        // M_mod = M_mod_m_at_p0 % m;
+        // pow2_mod = (M_mod + 1) % m;
+        M_mod = M_mod_m_at_p0;
+        pow2_mod = M_mod + 1;
+        if (pow2_mod >= m)
+        {
+            pow2_mod -= m;
+        }
     }
 
     // TODO: Inline this accessor into callers so the hot residue path reads the backing fields directly