[GH 28654] Address keidiai security issue

onnxruntime/core/mlas/lib/kleidiai/convolve_kleidiai.cpp

@@ -447,6 +447,9 @@ static std::unique_ptr<std::byte[]> LhsPackImageDataSme(const size_t ci, const s
 
     const auto lhs_size = kai_get_lhs_packed_size_lhs_imatmul_pack_x32p2vlx1_x32p_sme(m,kh*kw,ci);
     auto lhs = std::make_unique<std::byte[]>(lhs_size);
+    // Some ukernel packing paths may not overwrite every byte in the packed buffer for partial tiles.
+    // Start from a deterministic zero state so stale heap contents cannot influence later math.
+    std::fill_n(lhs.get(), lhs_size, std::byte{0});
 
     std::unique_ptr<float[]> nhwc_holder;
     const float* activation_src = nullptr;

onnxruntime/test/mlas/unittest/test_conv2d.h

@@ -716,12 +716,16 @@ class MlasConv2DTest : public MlasTestBase {
     //
     // This sequence forces pad-buffer growth by running a smaller-CI convolution followed by a larger-CI
     // convolution (with padding to ensure pad pointers are used), then runs the smaller-CI convolution again.
+    // Execute both orders to validate that results are stable regardless of invocation order.
     // Repeat a few times to increase the likelihood of triggering a reallocation and verify the path.
     //
     for (int i = 0; i < 4; ++i) {
       Test(1, 1, 64, 11, 11, 32, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1);   // smaller CI
       Test(1, 1, 320, 11, 11, 32, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1);  // larger CI forces pad buffer growth
       Test(1, 1, 64, 11, 11, 32, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1);   // sanity: back to smaller CI after growth
+
+      Test(1, 1, 320, 11, 11, 32, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1);  // reversed order start
+      Test(1, 1, 64, 11, 11, 32, 3, 3, 1, 1, 1, 1, 1, 1, 1, 1);   // reversed order end
     }
   }