diff --git a/.github/workflows/build-llama-cpp-windows.yml b/.github/workflows/build-llama-cpp-windows.yml
deleted file mode 100644
index cb6e6c19f..000000000
--- a/.github/workflows/build-llama-cpp-windows.yml
+++ /dev/null
@@ -1,182 +0,0 @@
-name: Build llama.cpp (Windows ARM64)
-
-# Manual-only full build of upstream llama.cpp on the Snapdragon Windows
-# ARM64 runner, with the same GGML backends geniex ships (Hexagon HTP +
-# Adreno OpenCL). Builds straight from the official ggml-org/llama.cpp
-# repo rather than the geniex submodule, so an arbitrary upstream commit
-# can be exercised without touching the submodule pin.
-
-on:
-  workflow_dispatch:
-    inputs:
-      llama_ref:
-        description: "llama.cpp git ref to build (branch, tag, or commit SHA). Blank = latest master."
-        required: false
-        type: string
-        default: ""
-
-permissions:
-  contents: read
-  packages: read
-
-jobs:
-  build:
-    name: build-llama-cpp
-    runs-on: windows-arm64
-    timeout-minutes: 120
-
-    steps:
-      - name: Configure Windows Git
-        shell: cmd
-        run: |
-          echo C:\Program Files\Git\bin>>%GITHUB_PATH%
-          git config --global core.longpaths true
-          git config --global core.protectNTFS false
-          git config --global core.fscache true
-          git config --global core.preloadindex true
-
-      # Needed for the composite actions; llama.cpp is cloned separately.
-      - name: Checkout geniex (for composite actions)
-        uses: actions/checkout@v7
-        with:
-          fetch-depth: 1
-
-      - name: Clone llama.cpp
-        id: llama
-        shell: powershell
-        env:
-          LLAMA_REF: ${{ inputs.llama_ref }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          $dir = "$env:GITHUB_WORKSPACE\llama.cpp"
-          if (Test-Path $dir) { Remove-Item -Recurse -Force $dir }
-          git clone https://github.com/ggml-org/llama.cpp.git $dir
-          if ($LASTEXITCODE -ne 0) { throw "git clone failed: $LASTEXITCODE" }
-          Push-Location $dir
-          if ($env:LLAMA_REF) {
-            git fetch --depth 1 origin $env:LLAMA_REF
-            if ($LASTEXITCODE -ne 0) { throw "git fetch '$env:LLAMA_REF' failed: $LASTEXITCODE" }
-            git checkout FETCH_HEAD
-            if ($LASTEXITCODE -ne 0) { throw "git checkout '$env:LLAMA_REF' failed: $LASTEXITCODE" }
-          }
-          $sha = (git rev-parse HEAD).Trim()
-          Pop-Location
-          Add-Content -Path $env:GITHUB_OUTPUT -Value "sha=$sha"
-          Add-Content -Path $env:GITHUB_OUTPUT -Value "dir=$dir"
-          $refLabel = if ($env:LLAMA_REF) { $env:LLAMA_REF } else { "master" }
-          Write-Host "llama.cpp ref=$refLabel sha=$sha"
-
-      - name: Install and configure ccache
-        shell: powershell
-        run: |
-          if (-not (Get-Command ccache -ErrorAction SilentlyContinue)) {
-            choco install ccache --no-progress -y
-          }
-          $env:CCACHE_DIR = "${{ github.workspace }}/.ccache"
-          Add-Content $env:GITHUB_ENV "CCACHE_DIR=$env:CCACHE_DIR"
-          # Tolerate __DATE__/__TIME__ and mtime jitter in ggml/llama sources.
-          Add-Content $env:GITHUB_ENV "CCACHE_SLOPPINESS=time_macros,include_file_mtime,include_file_ctime"
-          ccache --max-size=5G
-
-      - name: Restore ccache
-        uses: actions/cache@v6
-        with:
-          path: ${{ github.workspace }}/.ccache
-          key: ccache-llama-windows-arm64-${{ steps.llama.outputs.sha }}
-          restore-keys: ccache-llama-windows-arm64-
-
-      - name: Setup Snapdragon SDKs
-        uses: ./.github/actions/setup-snapdragon-sdks
-
-      - name: Configure HTP signing cert
-        shell: powershell
-        env:
-          HEXAGON_HTP_CERT_PFX: ${{ secrets.HEXAGON_HTP_CERT_PFX }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          if (-not $env:HEXAGON_HTP_CERT_PFX) {
-            throw "HEXAGON_HTP_CERT_PFX secret is empty; set it to the base64 of the HTP .pfx."
-          }
-          $dst = "$env:RUNNER_TEMP\ggml-htp-v1.pfx"
-          [IO.File]::WriteAllBytes($dst, [Convert]::FromBase64String($env:HEXAGON_HTP_CERT_PFX))
-          Add-Content -Path $env:GITHUB_ENV -Value "HEXAGON_HTP_CERT=$dst"
-          Write-Host "HEXAGON_HTP_CERT=$dst"
-
-      - name: Setup VS toolchain
-        uses: ./.github/actions/setup-vcvars
-
-      - name: Build llama.cpp
-        shell: powershell
-        env:
-          LLAMA_DIR: ${{ steps.llama.outputs.dir }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          $Preset = "arm64-windows-llvm-release"
-          $BuildDir = "$env:LLAMA_DIR/build-$Preset"
-
-          # Snapshot SDK vars before vcvars replay: the cmd.exe child does not
-          # inherit $GITHUB_ENV values, so the replay loop would clobber them.
-          $HexagonSdkRoot   = $env:HEXAGON_SDK_ROOT
-          $HexagonToolsRoot = $env:HEXAGON_TOOLS_ROOT
-          $HexagonHtpCert   = $env:HEXAGON_HTP_CERT
-          $OpenclSdkRoot    = $env:OPENCL_SDK_ROOT
-          $WindowsSdkBin    = $env:WINDOWS_SDK_BIN
-
-          # Source vcvars so cmake children (clang, ninja, inf2cat, signtool)
-          # inherit the Windows SDK PATH/INCLUDE/LIB.
-          $envDump = cmd /c "`"$env:VCVARS_BAT`" $env:VCVARS_ARGS && set"
-          foreach ($line in $envDump) {
-            if ($line -match "^(.+?)=(.*)$") {
-              [Environment]::SetEnvironmentVariable($matches[1], $matches[2], "Process")
-            }
-          }
-          $env:PATH = "$env:LLVM_BIN;$env:PATH"
-
-          # Restore SDK vars clobbered by the replay.
-          $env:HEXAGON_SDK_ROOT   = $HexagonSdkRoot
-          $env:HEXAGON_TOOLS_ROOT = $HexagonToolsRoot
-          $env:HEXAGON_HTP_CERT   = $HexagonHtpCert
-          $env:OPENCL_SDK_ROOT    = $OpenclSdkRoot
-          $env:WINDOWS_SDK_BIN    = $WindowsSdkBin
-
-          # CMAKE_INSTALL_LIBDIR=bin puts skels + .cat alongside the DLLs and
-          # executables in one runnable folder.
-          cmake -S $env:LLAMA_DIR --preset $Preset -B $BuildDir --log-level=VERBOSE `
-            -DGGML_HEXAGON=ON `
-            -DGGML_OPENCL=ON `
-            -DHEXAGON_SDK_ROOT="$HexagonSdkRoot" `
-            -DHEXAGON_TOOLS_ROOT="$HexagonToolsRoot" `
-            -DPREBUILT_LIB_DIR="windows_aarch64" `
-            -DCMAKE_PREFIX_PATH="$OpenclSdkRoot" `
-            -DCMAKE_INSTALL_LIBDIR=bin `
-            -DCMAKE_C_COMPILER_LAUNCHER=ccache `
-            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
-          if ($LASTEXITCODE -ne 0) { throw "cmake configure failed: $LASTEXITCODE" }
-
-          cmake --build $BuildDir -j $([Environment]::ProcessorCount)
-          if ($LASTEXITCODE -ne 0) { throw "cmake --build failed: $LASTEXITCODE" }
-
-          cmake --install $BuildDir --prefix "$env:LLAMA_DIR/dist"
-          if ($LASTEXITCODE -ne 0) { throw "cmake --install failed: $LASTEXITCODE" }
-
-          # Assemble the WIDO signing bundle: 6 skels + .inf + .cat in one folder.
-          # .inf has no upstream install() rule so it is copied from the build tree.
-          $signingDir = "$env:LLAMA_DIR/dist/libggml-htp"
-          New-Item -ItemType Directory -Force -Path $signingDir | Out-Null
-          $hexBuildDir = "$BuildDir/ggml/src/ggml-hexagon"
-          Copy-Item "$hexBuildDir/libggml-htp-v*.so" $signingDir
-          Copy-Item "$hexBuildDir/libggml-htp.inf"   $signingDir
-          Copy-Item "$hexBuildDir/libggml-htp.cat"   $signingDir
-
-      - name: ccache stats
-        shell: powershell
-        run: ccache --show-stats --verbose; $true
-
-      - name: Upload llama.cpp build artifact
-        uses: actions/upload-artifact@v7
-        with:
-          name: llama-cpp-windows-arm64-${{ steps.llama.outputs.sha }}
-          path: |
-            ${{ steps.llama.outputs.dir }}/dist/bin/
-            ${{ steps.llama.outputs.dir }}/dist/libggml-htp/
-          retention-days: 7