diff --git a/.github/workflows/samples-integration-test.yml b/.github/workflows/samples-integration-test.yml
index 4574af56f..0aedb0608 100644
--- a/.github/workflows/samples-integration-test.yml
+++ b/.github/workflows/samples-integration-test.yml
@@ -16,6 +16,9 @@ on:
 permissions:
   contents: read
 
+env:
+  ORT_TELEMETRY_DISABLED: '1'
+
 jobs:
   # ── Python Samples ──────────────────────────────────────────────────
   python-samples:
@@ -32,7 +35,7 @@ jobs:
           clean: true
 
       - name: Setup Python
-        uses: actions/setup-python@v7
+        uses: actions/setup-python@v6
         with:
           python-version: '3.12'
 
@@ -96,17 +99,17 @@ jobs:
           clean: true
 
       - name: Setup Node.js
-        uses: actions/setup-node@v7
+        uses: actions/setup-node@v6
         with:
           node-version: '20.x'
 
       - name: Setup Python (for node-gyp)
-        uses: actions/setup-python@v7
+        uses: actions/setup-python@v6
         with:
           python-version: '3.x'
 
       - name: Setup .NET SDK for NuGet authentication
-        uses: actions/setup-dotnet@v6
+        uses: actions/setup-dotnet@v5
         with:
           dotnet-version: '10.0.x'
 
@@ -182,7 +185,7 @@ jobs:
           clean: true
 
       - name: Setup .NET SDK
-        uses: actions/setup-dotnet@v6
+        uses: actions/setup-dotnet@v5
         with:
           dotnet-version: |
             8.0.x
diff --git a/.pipelines/v2/templates/steps-build-js.yml b/.pipelines/v2/templates/steps-build-js.yml
index 657704bd9..cf18c80c4 100644
--- a/.pipelines/v2/templates/steps-build-js.yml
+++ b/.pipelines/v2/templates/steps-build-js.yml
@@ -156,6 +156,14 @@ steps:
         throw "WinML runtime DLL not found in native artifact: $winmlDll"
       }
 
+      $directMlDll = Join-Path "${{ parameters.nativeArtifactDir }}" 'DirectML.dll'
+      if (Test-Path $directMlDll) {
+        Copy-Item $directMlDll -Destination $dst -Force
+        Write-Host "  staged DirectML.dll"
+      } else {
+        throw "DirectML.dll not found in native artifact: $directMlDll"
+      }
+
       Get-ChildItem $dst | ForEach-Object { Write-Host "    $($_.Name)  $($_.Length) bytes" }
     displayName: 'Stage prebuild directory'
 
diff --git a/.pipelines/v2/templates/steps-build-linux.yml b/.pipelines/v2/templates/steps-build-linux.yml
index b1328cba2..20a2a08a0 100644
--- a/.pipelines/v2/templates/steps-build-linux.yml
+++ b/.pipelines/v2/templates/steps-build-linux.yml
@@ -75,27 +75,44 @@ steps:
   displayName: 'Dump vcpkg error logs'
   condition: failed()
 
-# Stage the redistributable native artifact. Vcpkg uses static linkage on
-# Linux (see sdk_v2/cpp/triplets/x64-linux.cmake), so libfoundry_local.so
-# carries its transitive deps (azure-*, spdlog, fmt, libcurl, libssl/libcrypto,
-# zlib, brotli*) inside itself — the only file we need to forward downstream
-# is libfoundry_local.so. ORT/GenAI .so files are also present in bin/ but are
-# supplied to consumers separately (pip on the Python side, NuGet on the C#
-# side); test/example binaries are not part of the redistributable surface.
+# Stage the redistributable native artifacts for standalone C++ consumers.
 - bash: |
     set -euo pipefail
     src='$(Build.SourcesDirectory)/sdk_v2/cpp/build/Linux/${{ parameters.buildConfig }}/bin'
     dst='$(Build.ArtifactStagingDirectory)/native'
     mkdir -p "$dst"
 
-    primary="$src/libfoundry_local.so"
-    if [ ! -f "$primary" ]; then
-      echo "ERROR: libfoundry_local.so not found at $primary" >&2
-      exit 1
-    fi
-    cp -P "$primary" "$dst/"
-    echo "  staged libfoundry_local.so"
-
+    required=(
+      "$src/libfoundry_local.so"
+      "$src/libonnxruntime-genai.so"
+      "$src/libonnxruntime.so"
+    )
+    for f in "${required[@]}"; do
+      if [ ! -e "$f" ]; then
+        echo "ERROR: required native artifact not found: $f" >&2
+        exit 1
+      fi
+      cp -P "$f" "$dst/"
+      echo "  staged $(basename "$f")"
+    done
+    for f in "$src"/libonnxruntime.so.*; do
+      if [ -e "$f" ]; then
+        cp -P "$f" "$dst/"
+        echo "  staged $(basename "$f")"
+      fi
+    done
+    for f in "$src"/libonnxruntime-genai*.so; do
+      if [ -e "$f" ] && [ "$(basename "$f")" != "libonnxruntime-genai.so" ]; then
+        cp -P "$f" "$dst/"
+        echo "  staged $(basename "$f")"
+      fi
+    done
+    for f in "$src"/libonnxruntime_providers_*.so; do
+      if [ -e "$f" ]; then
+        cp -P "$f" "$dst/"
+        echo "  staged $(basename "$f")"
+      fi
+    done
   displayName: 'Stage native artifacts'
 
 # Stage debug symbols into a separate artifact so they can be published
diff --git a/.pipelines/v2/templates/steps-build-macos.yml b/.pipelines/v2/templates/steps-build-macos.yml
index a5da9fab9..eeb90ab53 100644
--- a/.pipelines/v2/templates/steps-build-macos.yml
+++ b/.pipelines/v2/templates/steps-build-macos.yml
@@ -81,26 +81,38 @@ steps:
   displayName: 'Dump vcpkg error logs'
   condition: failed()
 
-# Stage the redistributable native artifact. macOS has no triplet overlay so
-# vcpkg uses its default arm64-osx triplet, which is static — libfoundry_local.dylib
-# carries its transitive deps (azure-*, spdlog, fmt, libcurl, libssl/libcrypto,
-# zlib, brotli*) inside itself. ORT/GenAI .dylib files are supplied to consumers
-# separately (pip on the Python side, NuGet on the C# side); test/example binaries
-# are not part of the redistributable surface.
+# Stage the redistributable native artifacts for standalone C++ consumers.
 - bash: |
     set -euo pipefail
     src='$(Build.SourcesDirectory)/sdk_v2/cpp/build/macOS/${{ parameters.buildConfig }}/bin'
     dst='$(Build.ArtifactStagingDirectory)/native'
     mkdir -p "$dst"
 
-    primary="$src/libfoundry_local.dylib"
-    if [ ! -f "$primary" ]; then
-      echo "ERROR: libfoundry_local.dylib not found at $primary" >&2
-      exit 1
-    fi
-    cp -P "$primary" "$dst/"
-    echo "  staged libfoundry_local.dylib"
-
+    required=(
+      "$src/libfoundry_local.dylib"
+      "$src/libonnxruntime-genai.dylib"
+      "$src/libonnxruntime.dylib"
+    )
+    for f in "${required[@]}"; do
+      if [ ! -e "$f" ]; then
+        echo "ERROR: required native artifact not found: $f" >&2
+        exit 1
+      fi
+      cp -P "$f" "$dst/"
+      echo "  staged $(basename "$f")"
+    done
+    for f in "$src"/libonnxruntime.*.dylib; do
+      if [ -e "$f" ]; then
+        cp -P "$f" "$dst/"
+        echo "  staged $(basename "$f")"
+      fi
+    done
+    for f in "$src"/libonnxruntime_providers_*.dylib; do
+      if [ -e "$f" ]; then
+        cp -P "$f" "$dst/"
+        echo "  staged $(basename "$f")"
+      fi
+    done
   displayName: 'Stage native artifacts'
 
 # Stage debug symbols into a separate artifact so they can be published
diff --git a/.pipelines/v2/templates/steps-build-python.yml b/.pipelines/v2/templates/steps-build-python.yml
index 50f86c9da..84201a682 100644
--- a/.pipelines/v2/templates/steps-build-python.yml
+++ b/.pipelines/v2/templates/steps-build-python.yml
@@ -106,10 +106,32 @@ steps:
         # deps (onnxruntime-{core,gpu} / onnxruntime-genai-{core,cuda}). Test
         # and example binaries are also filtered out upstream. We just copy
         # the curated artifact contents straight in.
-        $files = Get-ChildItem -Path "${{ parameters.nativeArtifactDir }}" -Recurse -File
+        $rid = '${{ parameters.rid }}'
+        if ($rid -like 'win-*') {
+            $allowed = @('foundry_local.dll', 'foundry_local.lib', 'Microsoft.Windows.AI.MachineLearning.dll', 'DirectML.dll')
+        } elseif ($rid -like 'linux-*') {
+            $allowed = @('libfoundry_local.so')
+        } elseif ($rid -like 'osx-*') {
+            $allowed = @('libfoundry_local.dylib')
+        } else {
+            throw "Unsupported Python wheel RID: $rid"
+        }
+        $files = foreach ($name in $allowed) {
+            Get-ChildItem -Path "${{ parameters.nativeArtifactDir }}" -Filter $name -File -ErrorAction SilentlyContinue
+        }
         if ($files.Count -eq 0) {
             throw "No native artifacts found under ${{ parameters.nativeArtifactDir }}"
         }
+        if ($rid -like 'win-*') {
+            $found = @{}
+            foreach ($f in $files) {
+                $found[$f.Name] = $true
+            }
+            $missing = @($allowed | Where-Object { -not $found.ContainsKey($_) })
+            if ($missing.Count -gt 0) {
+                throw "Missing required Windows native artifact(s): $($missing -join ', ') under ${{ parameters.nativeArtifactDir }}"
+            }
+        }
         foreach ($f in $files) {
             Copy-Item $f.FullName -Destination $dest -Force
         }
diff --git a/.pipelines/v2/templates/steps-build-windows.yml b/.pipelines/v2/templates/steps-build-windows.yml
index 5f1d0f761..e72fde515 100644
--- a/.pipelines/v2/templates/steps-build-windows.yml
+++ b/.pipelines/v2/templates/steps-build-windows.yml
@@ -138,7 +138,13 @@ steps:
         (Join-Path $binDir  'foundry_local.dll'),
         (Join-Path $binDir  'foundry_local.pdb'),
         (Join-Path $linkDir 'foundry_local.lib'),
-        (Join-Path $binDir  'Microsoft.Windows.AI.MachineLearning.dll')
+        (Join-Path $binDir  'Microsoft.Windows.AI.MachineLearning.dll'),
+        (Join-Path $binDir  'DirectML.dll'),
+        (Join-Path $binDir  'onnxruntime.dll'),
+        (Join-Path $binDir  'onnxruntime-genai.dll')
+      )
+      $optionalPatterns = @(
+        'onnxruntime_providers_*.dll'
       )
 
       foreach ($s in $sources) {
@@ -149,6 +155,12 @@ steps:
         Copy-Item -Path $s -Destination $dst -Force
         Write-Host "  staged $(Split-Path -Leaf $s)"
       }
+      foreach ($pattern in $optionalPatterns) {
+        Get-ChildItem -Path $binDir -Filter $pattern -File -ErrorAction SilentlyContinue | ForEach-Object {
+          Copy-Item -Path $_.FullName -Destination $dst -Force
+          Write-Host "  staged $($_.Name)"
+        }
+      }
 
 - ${{ if eq(parameters.stageHeaders, true) }}:
   - task: CopyFiles@2
diff --git a/.pipelines/v2/templates/steps-pack-cpp-sdk.yml b/.pipelines/v2/templates/steps-pack-cpp-sdk.yml
index a878dbf3d..984bc6a33 100644
--- a/.pipelines/v2/templates/steps-pack-cpp-sdk.yml
+++ b/.pipelines/v2/templates/steps-pack-cpp-sdk.yml
@@ -61,9 +61,8 @@ steps:
 
         Copy-Item -Path (Join-Path $includeSrc '*') -Destination $includeDst -Recurse -Force
 
-        # The SDK tgz should include only public wrapper headers.
+        # The SDK tgz should include public wrapper headers and their header-only dependencies.
         $pathsToPrune = @(
-          (Join-Path $includeDst 'gsl'),
           (Join-Path $includeDst '_manifest')
         )
         foreach ($pathToPrune in $pathsToPrune) {
@@ -72,13 +71,17 @@ steps:
           }
         }
 
-        foreach ($f in $Files) {
-          $src = Join-Path $nativeSrc $f
-          if (-not (Test-Path $src)) {
-            throw "Required native file not found: $src"
+        foreach ($entry in $Files) {
+          $optional = $entry.StartsWith('?')
+          $pattern = if ($optional) { $entry.Substring(1) } else { $entry }
+          $matches = @(Get-ChildItem -Path $nativeSrc -Filter $pattern -File -ErrorAction SilentlyContinue)
+          if ($matches.Count -eq 0 -and -not $optional) {
+            throw "Required native file not found: $(Join-Path $nativeSrc $pattern)"
+          }
+          foreach ($match in $matches) {
+            $dst = Join-Path $libDst $match.Name
+            Copy-Item -Path $match.FullName -Destination $dst -Force
           }
-          $dst = Join-Path $libDst (Split-Path -Leaf $src)
-          Copy-Item -Path $src -Destination $dst -Force
         }
 
         if ($SymbolsArtifact) {
@@ -115,23 +118,40 @@ steps:
       New-SdkArchive -Rid 'win-x64' -NativeArtifact 'cpp-native-win-x64' -Files @(
         'foundry_local.dll',
         'foundry_local.pdb',
-        'foundry_local.lib'
+        'foundry_local.lib',
+        'Microsoft.Windows.AI.MachineLearning.dll',
+        'DirectML.dll',
+        'onnxruntime.dll',
+        'onnxruntime-genai.dll',
+        '?onnxruntime_providers_*.dll'
       )
 
       New-SdkArchive -Rid 'win-arm64' -NativeArtifact 'cpp-native-win-arm64' -Files @(
         'foundry_local.dll',
         'foundry_local.pdb',
-        'foundry_local.lib'
+        'foundry_local.lib',
+        'Microsoft.Windows.AI.MachineLearning.dll',
+        'DirectML.dll',
+        'onnxruntime.dll',
+        'onnxruntime-genai.dll',
+        '?onnxruntime_providers_*.dll'
       )
 
       New-SdkArchive -Rid 'linux-x64' -NativeArtifact 'cpp-native-linux-x64' -Files @(
-        'libfoundry_local.so'
+        'libfoundry_local.so',
+        'libonnxruntime-genai.so',
+        '?libonnxruntime-genai*.so',
+        '?libonnxruntime_providers_*.so',
+        'libonnxruntime.so*'
       ) -SymbolsArtifact 'cpp-native-symbols-linux-x64' -SymbolItems @(
         'libfoundry_local.so.dbg'
       )
 
       New-SdkArchive -Rid 'osx-arm64' -NativeArtifact 'cpp-native-osx-arm64' -Files @(
-        'libfoundry_local.dylib'
+        'libfoundry_local.dylib',
+        'libonnxruntime-genai.dylib',
+        'libonnxruntime*.dylib',
+        '?libonnxruntime_providers_*.dylib'
       ) -SymbolsArtifact 'cpp-native-symbols-osx-arm64' -SymbolItems @(
         'libfoundry_local.dylib.dSYM'
       )
diff --git a/.pipelines/v2/templates/steps-pack-js.yml b/.pipelines/v2/templates/steps-pack-js.yml
index 439ab41fd..9d7a4aff1 100644
--- a/.pipelines/v2/templates/steps-pack-js.yml
+++ b/.pipelines/v2/templates/steps-pack-js.yml
@@ -74,6 +74,7 @@ steps:
       $dst = '${{ parameters.outputDir }}'
       New-Item -ItemType Directory -Force -Path $dst | Out-Null
       Set-Location "$(Build.SourcesDirectory)/sdk_v2/js"
+      Copy-Item "$(Build.SourcesDirectory)/sdk_v2/deps_versions.json" -Destination "deps_versions.json" -Force
       npm pack --pack-destination $dst
       if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
       Get-ChildItem $dst | ForEach-Object { Write-Host "  $($_.Name)  $($_.Length) bytes" }
diff --git a/sdk_v2/cpp/CMakeLists.txt b/sdk_v2/cpp/CMakeLists.txt
index fd1c86b51..d7009c5e8 100644
--- a/sdk_v2/cpp/CMakeLists.txt
+++ b/sdk_v2/cpp/CMakeLists.txt
@@ -404,6 +404,13 @@ if(TARGET OnnxRuntime::OnnxRuntime)
                     $<TARGET_FILE_DIR:foundry_local>
             )
         endif()
+        if(WinMLEpCatalog_FOUND AND EXISTS "${WINML_EP_CATALOG_DLL_DIR}/DirectML.dll")
+            add_custom_command(TARGET foundry_local POST_BUILD
+                COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                    "${WINML_EP_CATALOG_DLL_DIR}/DirectML.dll"
+                    $<TARGET_FILE_DIR:foundry_local>
+            )
+        endif()
     elseif(APPLE)
         # macOS: copy dylibs so consumers that only link libfoundry_local.dylib (e.g. cache_only_tests) find the correct
         # ORT version instead of any system-installed ORT, which would cause an Ort::InitApi() version mismatch.
diff --git a/sdk_v2/cpp/cmake/FindOnnxRuntime.cmake b/sdk_v2/cpp/cmake/FindOnnxRuntime.cmake
index f11385873..5ed855cf1 100644
--- a/sdk_v2/cpp/cmake/FindOnnxRuntime.cmake
+++ b/sdk_v2/cpp/cmake/FindOnnxRuntime.cmake
@@ -97,7 +97,7 @@ else()
             message(STATUS "Downloading ${ORT_PACKAGE_NAME} ${ORT_VERSION} from nuget.org")
         endif()
     else()
-        message(STATUS "Using pre-configured ORT_FETCH_URL: ${ORT_FETCH_URL}")
+        message(STATUS "Using pre-configured ORT_FETCH_URL")
     endif()
 
     # Normalize backslashes (Windows paths) and handle .nupkg extension
@@ -106,11 +106,15 @@ else()
         set(_ORT_ZIP_PATH "${CMAKE_BINARY_DIR}/_deps/ortlib-download/ort.zip")
         get_filename_component(_ORT_ZIP_DIR "${_ORT_ZIP_PATH}" DIRECTORY)
         file(MAKE_DIRECTORY "${_ORT_ZIP_DIR}")
-        file(COPY_FILE "${ORT_FETCH_URL}" "${_ORT_ZIP_PATH}")
+        configure_file("${ORT_FETCH_URL}" "${_ORT_ZIP_PATH}" COPYONLY)
         set(ORT_FETCH_URL "${_ORT_ZIP_PATH}")
     endif()
 
-    FetchContent_Declare(ortlib URL ${ORT_FETCH_URL} DOWNLOAD_EXTRACT_TIMESTAMP TRUE DOWNLOAD_NAME ort.zip)
+    set(_ORT_FETCH_ARGS URL ${ORT_FETCH_URL} DOWNLOAD_NAME ort.zip)
+    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+        list(APPEND _ORT_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP TRUE)
+    endif()
+    FetchContent_Declare(ortlib ${_ORT_FETCH_ARGS})
     FetchContent_MakeAvailable(ortlib)
 
     set(_ORT_HEADER_DIR "${ortlib_SOURCE_DIR}/build/native/include")
@@ -149,7 +153,7 @@ else()
                 message(STATUS "Downloading ${_ORT_GPU_LINUX_PACKAGE} ${ORT_VERSION} from nuget.org")
             endif()
         else()
-            message(STATUS "Using pre-configured ORT_GPU_LINUX_FETCH_URL: ${ORT_GPU_LINUX_FETCH_URL}")
+            message(STATUS "Using pre-configured ORT_GPU_LINUX_FETCH_URL")
         endif()
 
         # Normalize backslashes and handle .nupkg extension
@@ -158,11 +162,15 @@ else()
             set(_ORT_GPU_ZIP_PATH "${CMAKE_BINARY_DIR}/_deps/ort_gpu_linux-download/ort_gpu_linux.zip")
             get_filename_component(_ORT_GPU_ZIP_DIR "${_ORT_GPU_ZIP_PATH}" DIRECTORY)
             file(MAKE_DIRECTORY "${_ORT_GPU_ZIP_DIR}")
-            file(COPY_FILE "${ORT_GPU_LINUX_FETCH_URL}" "${_ORT_GPU_ZIP_PATH}")
+            configure_file("${ORT_GPU_LINUX_FETCH_URL}" "${_ORT_GPU_ZIP_PATH}" COPYONLY)
             set(ORT_GPU_LINUX_FETCH_URL "${_ORT_GPU_ZIP_PATH}")
         endif()
 
-        FetchContent_Declare(ort_gpu_linux URL ${ORT_GPU_LINUX_FETCH_URL} DOWNLOAD_EXTRACT_TIMESTAMP TRUE DOWNLOAD_NAME ort_gpu_linux.zip)
+        set(_ORT_GPU_FETCH_ARGS URL ${ORT_GPU_LINUX_FETCH_URL} DOWNLOAD_NAME ort_gpu_linux.zip)
+        if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            list(APPEND _ORT_GPU_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP TRUE)
+        endif()
+        FetchContent_Declare(ort_gpu_linux ${_ORT_GPU_FETCH_ARGS})
         FetchContent_MakeAvailable(ort_gpu_linux)
 
         set(_ORT_LIB_DIR "${ort_gpu_linux_SOURCE_DIR}/runtimes/${_ORT_PLATFORM}/native")
diff --git a/sdk_v2/cpp/cmake/FindOnnxRuntimeGenAI.cmake b/sdk_v2/cpp/cmake/FindOnnxRuntimeGenAI.cmake
index f83002d64..f8bdc4b10 100644
--- a/sdk_v2/cpp/cmake/FindOnnxRuntimeGenAI.cmake
+++ b/sdk_v2/cpp/cmake/FindOnnxRuntimeGenAI.cmake
@@ -160,7 +160,7 @@ else()
             message(STATUS "Downloading ${_GENAI_PACKAGE_NAME} ${ORT_GENAI_VERSION} from nuget.org")
         endif()
     else()
-        message(STATUS "Using caller-provided GENAI_FETCH_URL: ${GENAI_FETCH_URL}")
+        message(STATUS "Using caller-provided GENAI_FETCH_URL")
     endif()
 
     # Normalize to forward slashes — backslashes from Windows paths cause CMake
@@ -173,15 +173,19 @@ else()
         set(_GENAI_ZIP_PATH "${CMAKE_BINARY_DIR}/_deps/genai-download/genai.zip")
         get_filename_component(_GENAI_ZIP_DIR "${_GENAI_ZIP_PATH}" DIRECTORY)
         file(MAKE_DIRECTORY "${_GENAI_ZIP_DIR}")
-        file(COPY_FILE "${GENAI_FETCH_URL}" "${_GENAI_ZIP_PATH}")
+        configure_file("${GENAI_FETCH_URL}" "${_GENAI_ZIP_PATH}" COPYONLY)
         set(GENAI_FETCH_URL "${_GENAI_ZIP_PATH}")
-        message(STATUS "Copied .nupkg to .zip for CMake extraction: ${GENAI_FETCH_URL}")
+        message(STATUS "Copied local GenAI .nupkg to .zip for CMake extraction")
     endif()
 
-    FetchContent_Declare(genailib
+    set(_GENAI_FETCH_ARGS
         URL ${GENAI_FETCH_URL}
-        DOWNLOAD_EXTRACT_TIMESTAMP TRUE
-        DOWNLOAD_NAME genai.zip  # .nupkg is a ZIP; force CMake to recognize the format
+        DOWNLOAD_NAME genai.zip)
+    if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+        list(APPEND _GENAI_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP TRUE)
+    endif()
+    FetchContent_Declare(genailib
+        ${_GENAI_FETCH_ARGS}
     )
     FetchContent_MakeAvailable(genailib)
 
diff --git a/sdk_v2/cpp/cmake/FindWinMLEpCatalog.cmake b/sdk_v2/cpp/cmake/FindWinMLEpCatalog.cmake
index dc732a4ff..4bf4eb0f7 100644
--- a/sdk_v2/cpp/cmake/FindWinMLEpCatalog.cmake
+++ b/sdk_v2/cpp/cmake/FindWinMLEpCatalog.cmake
@@ -60,30 +60,32 @@ if(NOT _WINML_PLATFORM_UPPER MATCHES "^(AMD64|X64|X86_64|ARM64|AARCH64)$")
     return()
 endif()
 
-include(cmake/nuget.cmake)
-
 # WINML_EP_CATALOG_FETCH_URL can be set externally (e.g. for CI where nuget.org is blocked).
 set(WINML_EP_CATALOG_FETCH_URL "" CACHE STRING "Override URL or local path for the WinML EP Catalog NuGet package")
 
-if(WINML_EP_CATALOG_FETCH_URL)
-    # Use FetchContent to download/extract the pre-downloaded package
-    include(FetchContent)
-    string(REPLACE "\\" "/" WINML_EP_CATALOG_FETCH_URL "${WINML_EP_CATALOG_FETCH_URL}")
-    if(WINML_EP_CATALOG_FETCH_URL MATCHES "\\.nupkg$" AND NOT WINML_EP_CATALOG_FETCH_URL MATCHES "^https?://")
-        set(_WINML_ZIP_PATH "${CMAKE_BINARY_DIR}/_deps/winml_ep_catalog-download/winml_ep_catalog.zip")
-        get_filename_component(_WINML_ZIP_DIR "${_WINML_ZIP_PATH}" DIRECTORY)
-        file(MAKE_DIRECTORY "${_WINML_ZIP_DIR}")
-        file(COPY_FILE "${WINML_EP_CATALOG_FETCH_URL}" "${_WINML_ZIP_PATH}")
-        set(WINML_EP_CATALOG_FETCH_URL "${_WINML_ZIP_PATH}")
-    endif()
-    FetchContent_Declare(winml_ep_catalog URL ${WINML_EP_CATALOG_FETCH_URL} DOWNLOAD_EXTRACT_TIMESTAMP TRUE DOWNLOAD_NAME winml_ep_catalog.zip)
-    FetchContent_MakeAvailable(winml_ep_catalog)
-    set(_WINML_EP_ROOT "${winml_ep_catalog_SOURCE_DIR}")
-    message(STATUS "WinML EP Catalog via FetchContent: ${_WINML_EP_ROOT}")
-else()
-    install_nuget_package(Microsoft.Windows.AI.MachineLearning ${WINML_EP_CATALOG_VERSION} _WINML_EP_ROOT
-        SOURCE https://api.nuget.org/v3/index.json)
+include(FetchContent)
+set(_WINML_EP_CATALOG_URL "${WINML_EP_CATALOG_FETCH_URL}")
+if(NOT _WINML_EP_CATALOG_URL)
+    set(_WINML_EP_CATALOG_URL
+        "https://api.nuget.org/v3-flatcontainer/microsoft.windows.ai.machinelearning/${WINML_EP_CATALOG_VERSION}/microsoft.windows.ai.machinelearning.${WINML_EP_CATALOG_VERSION}.nupkg")
+endif()
+
+string(REPLACE "\\" "/" _WINML_EP_CATALOG_URL "${_WINML_EP_CATALOG_URL}")
+if(_WINML_EP_CATALOG_URL MATCHES "\\.nupkg$" AND NOT _WINML_EP_CATALOG_URL MATCHES "^https?://")
+    set(_WINML_ZIP_PATH "${CMAKE_BINARY_DIR}/_deps/winml_ep_catalog-download/winml_ep_catalog.zip")
+    get_filename_component(_WINML_ZIP_DIR "${_WINML_ZIP_PATH}" DIRECTORY)
+    file(MAKE_DIRECTORY "${_WINML_ZIP_DIR}")
+    configure_file("${_WINML_EP_CATALOG_URL}" "${_WINML_ZIP_PATH}" COPYONLY)
+    set(_WINML_EP_CATALOG_URL "${_WINML_ZIP_PATH}")
+endif()
+set(_WINML_FETCH_ARGS URL ${_WINML_EP_CATALOG_URL} DOWNLOAD_NAME winml_ep_catalog.zip)
+if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+    list(APPEND _WINML_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP TRUE)
 endif()
+FetchContent_Declare(winml_ep_catalog ${_WINML_FETCH_ARGS})
+FetchContent_MakeAvailable(winml_ep_catalog)
+set(_WINML_EP_ROOT "${winml_ep_catalog_SOURCE_DIR}")
+message(STATUS "WinML EP Catalog via FetchContent: ${_WINML_EP_ROOT}")
 
 # Load the package's first-party CMake config for target discovery and layout
 # resolution. The config lives at build/cmake/<lowercased-package>-config.cmake
diff --git a/sdk_v2/cpp/nuget/pack.py b/sdk_v2/cpp/nuget/pack.py
index 902ba288a..ec27eaf3b 100644
--- a/sdk_v2/cpp/nuget/pack.py
+++ b/sdk_v2/cpp/nuget/pack.py
@@ -55,11 +55,16 @@
     "osx_arm64":  ("osx-arm64",  "libfoundry_local.dylib"),
 }
 
-# Sibling files copied into runtimes/<rid>/native/ when present in the upstream
-# artifact. Windows builds drop Microsoft.Windows.AI.MachineLearning.dll alongside
-# foundry_local.dll; other platforms don't, so presence alone drives inclusion.
+# Optional sibling files copied into runtimes/<rid>/native/ when present in the upstream
+# artifact. Non-Windows platforms do not have these files, so presence alone drives inclusion there.
 OPTIONAL_SIBLINGS: tuple[str, ...] = (
+)
+
+# Required Windows runtime siblings. The reg-free WinML runtime depends on DirectML, so a Windows RID
+# with only Microsoft.Windows.AI.MachineLearning.dll is incomplete and must fail packaging.
+WINDOWS_REQUIRED_SIBLINGS: tuple[str, ...] = (
     "Microsoft.Windows.AI.MachineLearning.dll",
+    "DirectML.dll",
 )
 
 log = logging.getLogger("pack")
@@ -159,6 +164,16 @@ def stage(args: argparse.Namespace, staging: Path) -> int:
         log.info("  %s → runtimes/%s/native/%s", lib_path, rid, lib_path.name)
         staged = 1
 
+        required_siblings = WINDOWS_REQUIRED_SIBLINGS if rid.startswith("win-") else ()
+        for sibling in required_siblings:
+            sibling_path = src_dir / sibling
+            if not sibling_path.is_file():
+                log.error("Expected Windows runtime dependency %s at %s but file not found.", sibling, src_dir)
+                sys.exit(1)
+            shutil.copy2(sibling_path, native_dir)
+            log.info("  %s → runtimes/%s/native/%s", sibling_path, rid, sibling)
+            staged += 1
+
         for sibling in OPTIONAL_SIBLINGS:
             sibling_path = src_dir / sibling
             if sibling_path.is_file():
diff --git a/sdk_v2/js/script/copy-native.mjs b/sdk_v2/js/script/copy-native.mjs
index 494d26671..34682f7dc 100644
--- a/sdk_v2/js/script/copy-native.mjs
+++ b/sdk_v2/js/script/copy-native.mjs
@@ -55,6 +55,7 @@ const wanted = (() => {
       "onnxruntime-genai.dll",
       "onnxruntime_providers_shared.dll",
       "Microsoft.Windows.AI.MachineLearning.dll",
+      "DirectML.dll",
     ];
   }
   if (process.platform === "darwin") {
diff --git a/sdk_v2/js/script/install-native.cjs b/sdk_v2/js/script/install-native.cjs
index bbb372091..ae32ce0b4 100644
--- a/sdk_v2/js/script/install-native.cjs
+++ b/sdk_v2/js/script/install-native.cjs
@@ -30,7 +30,6 @@ const PLATFORM_MAP = {
     'win32-x64': 'win-x64',
     'win32-arm64': 'win-arm64',
     'linux-x64': 'linux-x64',
-    'linux-arm64': 'linux-arm64',
     'darwin-arm64': 'osx-arm64',
 };
 const platformKey = `${os.platform()}-${os.arch()}`;
diff --git a/sdk_v2/js/script/pack-prebuilds.mjs b/sdk_v2/js/script/pack-prebuilds.mjs
index 36e1d0913..5dbf558aa 100644
--- a/sdk_v2/js/script/pack-prebuilds.mjs
+++ b/sdk_v2/js/script/pack-prebuilds.mjs
@@ -43,18 +43,14 @@ const destDir = resolve(pkgRoot, "prebuilds", `${process.platform}-${process.arc
 mkdirSync(destDir, { recursive: true });
 
 // foundry_local is required; ORT/GenAI siblings are excluded (fetched at install
-// time). The WinML EP catalog DLL is an optional sibling bundled on Windows.
+// time). The WinML EP catalog and DirectML DLLs are required siblings on Windows.
 const wanted = (() => {
-  if (process.platform === "win32") return ["foundry_local.dll"];
+  if (process.platform === "win32") return ["foundry_local.dll", "Microsoft.Windows.AI.MachineLearning.dll", "DirectML.dll"];
   if (process.platform === "darwin") return ["libfoundry_local.dylib"];
   return ["libfoundry_local.so"];
 })();
 
-// Optional native siblings — copied when present, skipped (with a warning) when
-// not. The reg-free WinML 2.x runtime ships next to foundry_local.dll on Windows
-// so WinML hardware EPs work out of the box without an install-time download.
-const optional =
-  process.platform === "win32" ? ["Microsoft.Windows.AI.MachineLearning.dll"] : [];
+const optional = [];
 
 let copied = 0;
 const available = new Set(readdirSync(sourceDir));
diff --git a/sdk_v2/python/src/foundry_local_sdk/_native/lib_loader.py b/sdk_v2/python/src/foundry_local_sdk/_native/lib_loader.py
index 1369759c8..0658c08d7 100644
--- a/sdk_v2/python/src/foundry_local_sdk/_native/lib_loader.py
+++ b/sdk_v2/python/src/foundry_local_sdk/_native/lib_loader.py
@@ -15,6 +15,7 @@
 import pathlib
 import platform
 import sys
+import sysconfig
 
 logger = logging.getLogger(__name__)
 
@@ -29,7 +30,7 @@ def _lib_name() -> str:
 
 def _platform_subdir() -> str:
     if sys.platform == "win32":
-        return "win-x64"
+        return "win-arm64" if sysconfig.get_platform().lower() == "win-arm64" else "win-x64"
     if sys.platform == "darwin":
         return "osx-arm64" if platform.machine() == "arm64" else "osx-x64"
     return "linux-x64"
@@ -248,7 +249,7 @@ def prepare_native_dependencies(foundry_local_dir: pathlib.Path) -> list:
         if add_dll_directory is not None:
             for d in {ort_path.parent, genai_path.parent, foundry_local_dir}:
                 try:
-                    add_dll_directory(str(d))
+                    handles.append(add_dll_directory(str(d)))
                 except OSError as exc:
                     logger.warning("os.add_dll_directory(%s) failed: %s", d, exc)
 
@@ -278,22 +279,10 @@ def prepare_native_dependencies(foundry_local_dir: pathlib.Path) -> list:
         logger.warning("Failed to preload ORT (%s): %s", ort_path, exc)
         return handles
 
-    # macOS only: GenAI's static initializer does its own dlopen("libonnxruntime.dylib"),
-    # which on Darwin only matches by leafname against dyld search paths or images
-    # whose install_name leaf is exactly "libonnxruntime.dylib". ORT's dylib has a
-    # versioned install_name instead (the soversion, e.g. "@rpath/libonnxruntime.1.dylib"),
-    # so neither match path succeeds and GenAI aborts in dyld init before any of our code
-    # runs. The second name GenAI tries is "<genai_dir>/libonnxruntime.dylib", so a symlink
-    # there fixes it. Linux dlopen consults the loaded-soname table by leafname and finds
-    # our already-RTLD_GLOBAL'd image without help.
-    if sys.platform == "darwin":
-        symlink_path = genai_path.parent / "libonnxruntime.dylib"
-        try:
-            if not symlink_path.exists():
-                symlink_path.symlink_to(ort_path)
-                logger.info("Created macOS GenAI->ORT symlink: %s -> %s", symlink_path, ort_path)
-        except OSError as exc:
-            logger.warning("Failed to create macOS ORT symlink at %s: %s", symlink_path, exc)
+    # GenAI's lazy InitApi() first checks ORT_LIB_PATH before falling back to a
+    # platform leafname. Point it at the exact ORT dylib/so we resolved instead
+    # of mutating the installed GenAI package to add an unversioned alias.
+    os.environ["ORT_LIB_PATH"] = str(ort_path)
 
     try:
         handles.append(ctypes.CDLL(str(genai_path), mode=ctypes.RTLD_GLOBAL))