diff --git a/.github/workflows/testing-linux.yml b/.github/workflows/testing-linux.yml
new file mode 100644
index 000000000000..3548fdf2c2a0
--- /dev/null
+++ b/.github/workflows/testing-linux.yml
@@ -0,0 +1,105 @@
+name: Linux
+
+on:
+  pull_request:
+    types: [ opened, synchronize, reopened ]
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  linux:
+    name: linux-${{ matrix.bits }} / ${{ matrix.uv_group }}
+    runs-on: ubuntu-24.04
+    strategy:
+      fail-fast: false
+      matrix:
+        bits: [ "32" ]  # Intentionally not 64, as we haven't configured self-hosted runners for it yet.
+        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21", "ci-llvm-20" ]
+        include:
+          - bits: 32
+            arch: i686
+            python: cpython-3.11.5-linux-x86-gnu  # latest available on uv
+          # - bits: 64  # Intentionally not 64, as we haven't configured self-hosted runners for it yet.
+          #   arch: x86_64
+          #   python: cpython-3.10-linux-x86_64-gnu
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v5
+
+      - name: Install system dependencies
+        run: |
+          if [[ "${{ matrix.bits }}" == "32" ]]; then
+            sudo dpkg --add-architecture i386
+          fi
+
+          apt_update() {
+            for i in 1 2 3; do
+              if sudo apt-get update; then return 0; fi
+              echo "apt-get update failed (attempt $i/3), retrying in 10s..."
+              sleep 10
+            done
+            return 1
+          }
+          apt_update
+
+          if [[ "${{ matrix.bits }}" == "32" ]]; then
+            sudo apt-get install -y \
+              gcc-multilib \
+              g++-multilib \
+              libpng-dev:i386 \
+              libjpeg-dev:i386
+          else
+            sudo apt-get install -y \
+              libpng-dev \
+              libjpeg-dev
+          fi
+
+      - name: Sync CI environment
+        run: |
+          setarch ${{ matrix.arch }} bash -ec "
+            CC='gcc -m${{ matrix.bits }}' CXX='g++ -m${{ matrix.bits }}' \
+            uv sync --python '${{ matrix.python }}' --group '${{ matrix.uv_group }}' --no-install-project
+            echo '${GITHUB_WORKSPACE}/.venv/bin' >> '$GITHUB_PATH'
+            echo 'VIRTUAL_ENV=${GITHUB_WORKSPACE}/.venv' >> '$GITHUB_ENV'
+          "
+
+      - name: Configure LLVM
+        run: echo "Halide_LLVM_ROOT=$(halide-llvm --prefix)" >> "$GITHUB_ENV"
+
+      - name: Configure CMake
+        run: |
+          TOOLCHAIN_ARGS=()
+          if [[ "${{ matrix.bits }}" == "32" ]]; then
+            TOOLCHAIN_ARGS+=(
+              "-DCMAKE_TOOLCHAIN_FILE=${GITHUB_WORKSPACE}/cmake/toolchain.linux-i386.cmake"
+              "-DWITH_PYTHON_BINDINGS=OFF"
+            )
+          fi
+
+          cmake -G Ninja -S . -B build \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DHalide_LLVM_ROOT="${Halide_LLVM_ROOT}" \
+            "${TOOLCHAIN_ARGS[@]}"
+
+      - name: Initial build
+        run: cmake --build build
+
+      - name: Test (host)
+        run: |
+          cmake -S . -B build -DHalide_TARGET=host
+          cmake --build build
+          ctest --test-dir build --build-config RelWithDebInfo --output-on-failure -j "$(nproc)"
+
+      - name: Test (no extensions)
+        run: |
+          cmake -S . -B build -DHalide_TARGET=cmake
+          cmake --build build
+          ctest --test-dir build --build-config RelWithDebInfo --output-on-failure -j "$(nproc)"
diff --git a/.github/workflows/testing-windows.yml b/.github/workflows/testing-windows.yml
new file mode 100644
index 000000000000..27e566c16aea
--- /dev/null
+++ b/.github/workflows/testing-windows.yml
@@ -0,0 +1,95 @@
+name: Windows
+
+on:
+  pull_request:
+    types: [ opened, synchronize, reopened ]
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+defaults:
+  run:
+    shell: bash
+
+jobs:
+  windows:
+    name: windows-${{ matrix.bits }} / ${{ matrix.uv_group }}
+    runs-on: windows-2022
+    strategy:
+      fail-fast: false
+      matrix:
+        bits: [ "32" ]  # Intentionally not 64, as we haven't configured self-hosted runners for it yet.
+        uv_group: [ "ci-llvm-main", "ci-llvm-22", "ci-llvm-21", "ci-llvm-20" ]
+        include:
+          - bits: 32
+            arch: x86
+            python: cpython-3.10.20-windows-x86-none
+          # - bits: 64  # Intentionally not 64, as we haven't configured self-hosted runners for it yet.
+          #   arch: x64
+          #   python: cpython-3.10.20-windows-x86_64-none
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v5
+
+      - uses: ilammy/msvc-dev-cmd@v1
+        with:
+          arch: ${{ matrix.arch }}
+
+      - name: Sync CI environment
+        run: |
+          # uv uses the host platform for dependency resolution regardless of the target
+          # Python, so on Windows it would download x86_64 halide-llvm (~250MB) even for
+          # a 32-bit build. Instead: read the pinned version from the lockfile (no download),
+          # sync only ci-base (no LLVM), then fetch the correct win32 wheel directly.
+          LLVM_VER=$(uv export --group '${{ matrix.uv_group }}' --no-emit-project \
+            | awk -F'==' '/^halide-llvm==/ { gsub(/ .*/, "", $2); print $2 }')
+
+          uv sync --python '${{ matrix.python }}' --group ci-base --no-install-project
+          uv pip install --python-platform i686-pc-windows-msvc --only-binary :all: \
+            --extra-index-url https://pypi.halide-lang.org/simple/ \
+            "halide-llvm==${LLVM_VER}"
+
+          echo "${GITHUB_WORKSPACE}/.venv/Scripts" >> "$GITHUB_PATH"
+          echo "VIRTUAL_ENV=${GITHUB_WORKSPACE}/.venv" >> "$GITHUB_ENV"
+
+      - name: Configure LLVM
+        run: echo "Halide_LLVM_ROOT=$(halide-llvm --prefix)" >> "$GITHUB_ENV"
+
+      - name: Configure CMake
+        run: |
+          # NB: vcpkg docs use VCPKG_ROOT, but GHA runners use VCPKG_INSTALLATION_ROOT
+          TOOLCHAIN_ARGS=(
+            "-DCMAKE_TOOLCHAIN_FILE=${VCPKG_INSTALLATION_ROOT}/scripts/buildsystems/vcpkg.cmake"
+            "-DVCPKG_TARGET_TRIPLET=${{ matrix.arch }}-windows"
+            "-DVCPKG_MANIFEST_FEATURES=developer"
+          )
+          if [[ "${{ matrix.bits }}" == "32" ]]; then
+            TOOLCHAIN_ARGS+=("-DWITH_PYTHON_BINDINGS=OFF")
+          fi
+
+          cmake -G Ninja -S . -B build \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DHalide_LLVM_ROOT="${Halide_LLVM_ROOT}" \
+            "${TOOLCHAIN_ARGS[@]}"
+
+      - name: Initial build
+        run: cmake --build build
+
+      - name: Test (host)
+        run: |
+          cmake -S . -B build -DHalide_TARGET=host
+          cmake --build build
+          ctest --test-dir build --build-config RelWithDebInfo --output-on-failure -j "$(nproc)"
+
+      - name: Test (no extensions)
+        run: |
+          cmake -S . -B build -DHalide_TARGET=cmake
+          cmake --build build
+          ctest --test-dir build --build-config RelWithDebInfo --output-on-failure -j "$(nproc)"
diff --git a/cmake/toolchain.linux-i386.cmake b/cmake/toolchain.linux-i386.cmake
index bca44475e752..f7b4dae5df70 100644
--- a/cmake/toolchain.linux-i386.cmake
+++ b/cmake/toolchain.linux-i386.cmake
@@ -15,6 +15,4 @@ set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
 set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
 
-# Set to empty string to indicate the resulting binaries can be natively executed
-set(CMAKE_CROSSCOMPILING_EMULATOR)
-
+set(CMAKE_CROSSCOMPILING_EMULATOR /usr/bin/env)
diff --git a/test/performance/packed_planar_fusion.cpp b/test/performance/packed_planar_fusion.cpp
index 3c23f9f5f62d..9819db4c88ee 100644
--- a/test/performance/packed_planar_fusion.cpp
+++ b/test/performance/packed_planar_fusion.cpp
@@ -79,9 +79,9 @@ int main(int argc, char **argv) {
     delete[] storage_1;
     delete[] storage_2;
 
-    if (t_planar_planar > t_packed_packed * 2 ||
-        t_packed_packed > t_packed_planar * 2 ||
-        t_planar_packed > t_packed_planar * 2) {
+    if (t_planar_planar > t_packed_packed * 4 ||
+        t_packed_packed > t_packed_planar * 4 ||
+        t_planar_packed > t_packed_planar * 4) {
         printf("Times were not in expected order:\n"
                "planar -> planar: %f \n"
                "packed -> packed: %f \n"