diff --git a/.github/workflows/publish_pypi.yml b/.github/workflows/publish_pypi.yml
index 7c8b2fb..cddc952 100644
--- a/.github/workflows/publish_pypi.yml
+++ b/.github/workflows/publish_pypi.yml
@@ -11,90 +11,348 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  build_wheels:
-    name: Build wheels on ${{ matrix.os }}
-    runs-on: ${{ matrix.os }}
-    continue-on-error: true
-    env:
-      MACOSX_DEPLOYMENT_TARGET: "10.15"
-    strategy:
-      matrix:
-        os: [ubuntu-22.04, macos-15-intel, windows-latest]
-        cibw_archs: ["auto64"]
-        cibw_build: ["cp310-*", "cp311-*", "cp312-*", "cp313-*"]
+  # ──────────────────────────────────────────────────────────────────────────
+  # Stage 1 — Prebuild jobs (one per OS/arch, run in parallel)
+  # ──────────────────────────────────────────────────────────────────────────
 
+  prebuild_linux:
+    name: Prebuild (Linux)
+    runs-on: ubuntu-22.04
     steps:
       - uses: actions/checkout@v3
-        name: Check out
 
-      - uses: ilammy/msvc-dev-cmd@v1
-        name: Add MSVS Path
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-ubuntu-22.04-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-ubuntu-22.04-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Restore ccache
+        id: ccache
+        uses: actions/cache@v4
+        with:
+          path: ccache_dir
+          key: ccache-linux-auto64-${{ github.sha }}
+          restore-keys: |
+            ccache-linux-auto64-
+
+      - name: Run Linux prebuild in manylinux container
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true' || steps.ccache.outputs.cache-hit != 'true'
+        run: |
+          docker run --rm \
+            -v ${{ github.workspace }}:/project \
+            -w /project \
+            quay.io/pypa/manylinux_2_28_x86_64 \
+            bash ci-utils/prebuild_linux.sh
+
+  prebuild_macos_intel:
+    name: Prebuild (macOS intel)
+    runs-on: macos-15-intel
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-15-intel-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-15-intel-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Run macOS intel prebuild
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
+        run: |
+          brew install nasm
+          brew uninstall --ignore-dependencies openssl || true
+          sudo rm -rf /usr/local/include/openssl
+          bash ci-utils/install_prereq_linux.sh ./prereq_cache/local_install
+          cmake -S ci-utils/tensorstore_prebuild -B /tmp/ts_cmake_prebuild \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DTENSORSTORE_USE_SYSTEM_JPEG=ON \
+            -DTENSORSTORE_USE_SYSTEM_ZLIB=ON \
+            -DTENSORSTORE_USE_SYSTEM_PNG=ON \
+            -DCMAKE_CXX_FLAGS=-Wno-missing-template-arg-list-after-template-kw \
+            -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache
+          cmake --build /tmp/ts_cmake_prebuild -j$(sysctl -n hw.ncpu)
+
+  prebuild_windows:
+    name: Prebuild (Windows)
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-windows-latest-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-windows-latest-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Set up MSVC environment
+        uses: ilammy/msvc-dev-cmd@v1
 
       - name: Add NASM
-        if: matrix.os == 'windows-latest'
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
         uses: ilammy/setup-nasm@v1
 
       - name: Add Ninja
-        if: matrix.os == 'windows-latest'
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
         uses: seanmiddleditch/gha-setup-ninja@master
 
+      - name: Run Windows prebuild
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
+        shell: cmd
+        run: |
+          call ci-utils\install_prereq_win.bat
+          if %errorlevel% neq 0 exit /b %errorlevel%
+          xcopy /E /I /y local_install prereq_cache\local_install
+          cmake -S ci-utils\tensorstore_prebuild -B C:\ts_cmake_prebuild -DCMAKE_GENERATOR=Ninja -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache -DCMAKE_BUILD_TYPE=Release
+          if %errorlevel% neq 0 exit /b %errorlevel%
+          cmake --build C:\ts_cmake_prebuild --parallel
+          if %errorlevel% neq 0 exit /b %errorlevel%
+
+  prebuild_macos_arm64:
+    name: Prebuild (macOS arm64)
+    runs-on: macos-14
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-14-arm64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-14-arm64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Run macOS arm64 prebuild
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
+        run: |
+          brew install nasm
+          brew uninstall --ignore-dependencies jpeg-turbo || true
+          bash ci-utils/install_prereq_linux.sh ./prereq_cache/local_install
+          cmake -S ci-utils/tensorstore_prebuild -B /tmp/ts_cmake_prebuild \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DTENSORSTORE_USE_SYSTEM_JPEG=ON \
+            -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache
+          cmake --build /tmp/ts_cmake_prebuild -j$(sysctl -n hw.ncpu)
+
+  # ──────────────────────────────────────────────────────────────────────────
+  # Stage 2 — Wheel-build + publish jobs (4 Python versions per OS, in parallel)
+  #   Each job depends only on its own OS's prebuild.
+  # ──────────────────────────────────────────────────────────────────────────
+
+  build_wheels_linux:
+    name: Build wheels (Linux, ${{ matrix.cibw_build }})
+    runs-on: ubuntu-22.04
+    needs: [prebuild_linux]
+    continue-on-error: true
+    strategy:
+      matrix:
+        cibw_build: ["cp310-*", "cp311-*", "cp312-*", "cp313-*"]
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-ubuntu-22.04-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-ubuntu-22.04-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Restore ccache
+        uses: actions/cache@v4
+        with:
+          path: ccache_dir
+          key: ccache-linux-auto64-${{ github.sha }}-${{ matrix.cibw_build }}
+          restore-keys: |
+            ccache-linux-auto64-${{ github.sha }}
+            ccache-linux-auto64-
+
       - uses: actions/setup-python@v4
-        name: Install Python
         with:
           python-version: '3.11'
 
       - name: Install cibuildwheel
-        run: |
-          python -m pip install cibuildwheel delvewheel wheel
+        run: python -m pip install cibuildwheel wheel
 
-      - name: Building wheels 
-        run: |
-          python -m cibuildwheel --output-dir dist
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
         env:
           CIBW_BUILD: ${{ matrix.cibw_build }}
           CIBW_BUILD_VERBOSITY: 3
           CIBW_SKIP: "*musllinux*"
           CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
+          CIBW_BEFORE_ALL_LINUX: bash /project/ci-utils/before_all_linux.sh
+          CIBW_ENVIRONMENT_LINUX: >-
+            LD_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64:$LD_LIBRARY_PATH"
+            ON_GITHUB="TRUE"
+            ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
+            CCACHE_DIR=/project/ccache_dir
+            CCACHE_MAXSIZE=5G
+            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON -DFETCHCONTENT_BASE_DIR=/project/ts_fc_cache"
+          CIBW_ARCHS: auto64
+          CIBW_BEFORE_TEST_LINUX: dnf -y install maven java
+          CIBW_TEST_REQUIRES: bfio>=2.4.0 tensorstore numpy pytest
+          CIBW_TEST_COMMAND: python -W default -m pytest -vv -s {project}/tests/python
+
+      - name: Install Dependencies
+        run: python -m pip install --upgrade twine requests
+
+      - name: Publish to PyPi
+        run: python -m twine upload dist/*.whl
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PYPI_TOKEN }}
+          TWINE_REPOSITORY: pypi
+
+  build_wheels_macos_intel:
+    name: Build wheels (macOS intel, ${{ matrix.cibw_build }})
+    runs-on: macos-15-intel
+    needs: [prebuild_macos_intel]
+    continue-on-error: true
+    env:
+      MACOSX_DEPLOYMENT_TARGET: "10.15"
+    strategy:
+      matrix:
+        cibw_build: ["cp310-*", "cp311-*", "cp312-*", "cp313-*"]
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-15-intel-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-15-intel-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install cibuildwheel
+        run: python -m pip install cibuildwheel wheel
+
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
+        env:
+          CIBW_BUILD: ${{ matrix.cibw_build }}
+          CIBW_BUILD_VERBOSITY: 3
           CIBW_BEFORE_ALL_MACOS: brew install nasm &&
-                          bash ci-utils/install_prereq_linux.sh &&
-                          mkdir -p /tmp/argolid_bld &&
+                          bash ci-utils/before_all_macos.sh &&
                           brew uninstall --ignore-dependencies openssl || true &&
-                          sudo rm -rf /usr/local/include/openssl &&
-                          cp -r local_install /tmp/argolid_bld
-          CIBW_BEFORE_ALL_LINUX: curl -L https://www.nasm.us/pub/nasm/releasebuilds/2.15.05/nasm-2.15.05.tar.bz2 -o nasm-2.15.05.tar.bz2 &&
-                                   tar -xjf nasm-2.15.05.tar.bz2 &&
-                                   cd nasm-2.15.05 &&
-                                   ./configure && 
-                                   make && 
-                                   make install &&
-                                   cd .. &&
-                                   bash ci-utils/install_prereq_linux.sh &&
-                                   mkdir -p /tmp/argolid_bld &&
-                                   cp -r local_install /tmp/argolid_bld
-          CIBW_BEFORE_ALL_WINDOWS: ci-utils\install_prereq_win.bat &&
-                                   xcopy /E /I /y local_install C:\TEMP\argolid_bld\local_install 
+                          sudo rm -rf /usr/local/include/openssl
           CIBW_ENVIRONMENT_MACOS: >-
             MACOSX_DEPLOYMENT_TARGET=11.0
             REPAIR_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64"
             ON_GITHUB="TRUE"
             ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
-            CMAKE_ARGS="-DCMAKE_CXX_FLAGS=-Wno-missing-template-arg-list-after-template-kw -DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON"
-          CIBW_ENVIRONMENT_LINUX: >-
-            LD_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64:$LD_LIBRARY_PATH"
-            ON_GITHUB="TRUE" 
-            ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
-            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON"
+            CMAKE_ARGS="-DCMAKE_CXX_FLAGS=-Wno-missing-template-arg-list-after-template-kw -DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache"
           CIBW_REPAIR_WHEEL_COMMAND_MACOS: DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-listdeps {wheel} && DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-wheel --require-archs {delocate_archs} -w {dest_dir} {wheel}
-          CIBW_ENVIRONMENT_WINDOWS: PATH="$TEMP\\argolid\\bin;$PATH" ON_GITHUB="TRUE" ARGOLID_DEP_DIR="C:\\TEMP\\argolid_bld\\local_install" CMAKE_ARGS="-DCMAKE_GENERATOR=Ninja"
+          CIBW_ARCHS: auto64
+
+      - name: Install Dependencies
+        run: python -m pip install --upgrade twine requests
+
+      - name: Publish to PyPi
+        run: python -m twine upload dist/*.whl
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PYPI_TOKEN }}
+          TWINE_REPOSITORY: pypi
+
+  build_wheels_windows:
+    name: Build wheels (Windows, ${{ matrix.cibw_build }})
+    runs-on: windows-latest
+    needs: [prebuild_windows]
+    continue-on-error: true
+    strategy:
+      matrix:
+        cibw_build: ["cp310-*", "cp311-*", "cp312-*", "cp313-*"]
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-windows-latest-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-windows-latest-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - uses: ilammy/msvc-dev-cmd@v1
+        name: Add MSVS Path
+
+      - name: Add NASM
+        uses: ilammy/setup-nasm@v1
+
+      - name: Add Ninja
+        uses: seanmiddleditch/gha-setup-ninja@master
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install cibuildwheel
+        run: python -m pip install cibuildwheel delvewheel wheel
+
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
+        env:
+          CIBW_BUILD: ${{ matrix.cibw_build }}
+          CIBW_BUILD_VERBOSITY: 3
+          CIBW_BEFORE_ALL_WINDOWS: ci-utils\before_all_win.bat
+          CIBW_ENVIRONMENT_WINDOWS: PATH="$TEMP\\argolid\\bin;$PATH" ON_GITHUB="TRUE" ARGOLID_DEP_DIR="C:\\TEMP\\argolid_bld\\local_install" CMAKE_ARGS="-DCMAKE_GENERATOR=Ninja -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache"
           CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: "delvewheel repair -w {dest_dir} {wheel}"
-          CIBW_ARCHS: ${{ matrix.cibw_archs }}
-          CIBW_BEFORE_TEST_LINUX: dnf -y install maven java
-          CIBW_TEST_REQUIRES: bfio>=2.4.0 tensorstore numpy pytest
-          CIBW_TEST_COMMAND: python -W default -m pytest -vv -s {project}/tests/python
+          CIBW_ARCHS: auto64
 
       - name: Install Dependencies
         run: python -m pip install --upgrade twine requests
-        
+
       - name: Publish to PyPi
         run: python -m twine upload dist/*.whl
         env:
@@ -103,61 +361,66 @@ jobs:
           TWINE_REPOSITORY: pypi
 
   build_wheels_apple_arm64:
-    name: Build wheels on ${{ matrix.os }}
-    runs-on: ${{ matrix.os }}
+    name: Build wheels (macOS arm64, ${{ matrix.cibw_build }})
+    runs-on: macos-14
+    needs: [prebuild_macos_arm64]
     continue-on-error: true
     env:
       MACOSX_DEPLOYMENT_TARGET: "11.0"
     strategy:
       matrix:
-        os: [macos-14]
-        cibw_archs: ["arm64"]
         cibw_build: ["cp310-*", "cp311-*", "cp312-*", "cp313-*"]
 
     steps:
       - uses: actions/checkout@v3
-        name: Check out
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-14-arm64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-14-arm64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
 
       - uses: actions/setup-python@v4
-        name: Install Python
         with:
           python-version: '3.11'
-  
+
       - name: Install cibuildwheel
-        run: |
-          python -m pip install cibuildwheel delocate wheel
+        run: python -m pip install cibuildwheel delocate wheel
 
-      - name: Building wheels 
-        run: |
-          python -m cibuildwheel --output-dir dist
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
         env:
           CIBW_BUILD: ${{ matrix.cibw_build }}
           CIBW_BUILD_VERBOSITY: 3
           CIBW_ARCHS_MACOS: arm64
-          CIBW_BEFORE_ALL_MACOS:  brew install nasm &&
-                                  brew uninstall --ignore-dependencies jpeg-turbo && 
-                                  bash ci-utils/install_prereq_linux.sh &&
-                                  mkdir -p /tmp/argolid_bld &&
-                                  cp -r local_install /tmp/argolid_bld
+          CIBW_BEFORE_ALL_MACOS: brew install nasm &&
+                                  brew uninstall --ignore-dependencies jpeg-turbo &&
+                                  bash ci-utils/before_all_macos.sh
           CIBW_ENVIRONMENT_MACOS: >-
             MACOSX_DEPLOYMENT_TARGET=11.0
             REPAIR_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64"
             ON_GITHUB="TRUE"
             ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
-            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON"
+            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache"
           CIBW_REPAIR_WHEEL_COMMAND_MACOS: |
             DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-listdeps {wheel} &&
             MACOSX_DEPLOYMENT_TARGET=11.0 DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-wheel --require-archs {delocate_archs} -w {dest_dir} {wheel} -e libc++ -e libunwind
-          CIBW_ARCHS: ${{ matrix.cibw_archs }}
+          CIBW_ARCHS: arm64
           CIBW_TEST_REQUIRES: bfio>=2.4.0 tensorstore numpy pytest
           CIBW_TEST_COMMAND: python -W default -m pytest -vv -s {project}/tests/python
 
       - name: Install Dependencies
         run: python -m pip install --upgrade twine requests
-        
+
       - name: Publish to PyPi
         run: python -m twine upload dist/*.whl
         env:
           TWINE_USERNAME: __token__
           TWINE_PASSWORD: ${{ secrets.PYPI_TOKEN }}
-          TWINE_REPOSITORY: pypi
\ No newline at end of file
+          TWINE_REPOSITORY: pypi
diff --git a/.github/workflows/wheel_build.yml b/.github/workflows/wheel_build.yml
index 02c3138..8be4dd4 100644
--- a/.github/workflows/wheel_build.yml
+++ b/.github/workflows/wheel_build.yml
@@ -9,148 +9,410 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  build_wheels:
-    name: Build wheels on ${{ matrix.os }}
-    runs-on: ${{ matrix.os }}
-    continue-on-error: true
+  # ──────────────────────────────────────────────────────────────────────────
+  # Stage 1 — Prebuild jobs (one per OS/arch, run in parallel)
+  #   Each populates two caches:
+  #     prereq_cache/ — filepattern + pybind11
+  #     ts_fc_cache/  — TensorStore FetchContent source tree
+  #   Wheel-build jobs below depend only on their matching prebuild job,
+  #   so each OS starts building wheels as soon as its own prebuild is done.
+  # ──────────────────────────────────────────────────────────────────────────
 
-    env:
-      MACOSX_DEPLOYMENT_TARGET: "10.15"
-    strategy:
-      matrix:
-        os: [ubuntu-22.04, macos-15-intel, windows-latest]
-        cibw_archs: ["auto64"]
-        cibw_build: ["cp310", "cp311", "cp312", "cp313"]
+  prebuild_linux:
+    name: Prebuild (Linux)
+    runs-on: ubuntu-22.04
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-ubuntu-22.04-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
 
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-ubuntu-22.04-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Restore ccache
+        id: ccache
+        uses: actions/cache@v4
+        with:
+          path: ccache_dir
+          key: ccache-linux-auto64-${{ github.sha }}
+          restore-keys: |
+            ccache-linux-auto64-
+
+      - name: Run Linux prebuild in manylinux container
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true' || steps.ccache.outputs.cache-hit != 'true'
+        run: |
+          docker run --rm \
+            -v ${{ github.workspace }}:/project \
+            -w /project \
+            quay.io/pypa/manylinux_2_28_x86_64 \
+            bash ci-utils/prebuild_linux.sh
+
+  prebuild_macos_intel:
+    name: Prebuild (macOS intel)
+    runs-on: macos-15-intel
     steps:
       - uses: actions/checkout@v3
-        name: Check out
 
-      - uses: ilammy/msvc-dev-cmd@v1
-        name: Add MSVS Path
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-15-intel-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-15-intel-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Run macOS intel prebuild
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
+        run: |
+          brew install nasm
+          brew uninstall --ignore-dependencies openssl || true
+          sudo rm -rf /usr/local/include/openssl
+          bash ci-utils/install_prereq_linux.sh ./prereq_cache/local_install
+          cmake -S ci-utils/tensorstore_prebuild -B /tmp/ts_cmake_prebuild \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DTENSORSTORE_USE_SYSTEM_JPEG=ON \
+            -DTENSORSTORE_USE_SYSTEM_ZLIB=ON \
+            -DTENSORSTORE_USE_SYSTEM_PNG=ON \
+            -DCMAKE_CXX_FLAGS=-Wno-missing-template-arg-list-after-template-kw \
+            -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache
+          cmake --build /tmp/ts_cmake_prebuild -j$(sysctl -n hw.ncpu)
+
+  prebuild_windows:
+    name: Prebuild (Windows)
+    runs-on: windows-latest
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-windows-latest-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-windows-latest-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Set up MSVC environment
+        uses: ilammy/msvc-dev-cmd@v1
 
       - name: Add NASM
-        if: matrix.os == 'windows-latest'
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
         uses: ilammy/setup-nasm@v1
 
       - name: Add Ninja
-        if: matrix.os == 'windows-latest'
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
         uses: seanmiddleditch/gha-setup-ninja@master
 
+      - name: Run Windows prebuild
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
+        shell: cmd
+        run: |
+          call ci-utils\install_prereq_win.bat
+          if %errorlevel% neq 0 exit /b %errorlevel%
+          xcopy /E /I /y local_install prereq_cache\local_install
+          cmake -S ci-utils\tensorstore_prebuild -B C:\ts_cmake_prebuild -DCMAKE_GENERATOR=Ninja -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache -DCMAKE_BUILD_TYPE=Release
+          if %errorlevel% neq 0 exit /b %errorlevel%
+          cmake --build C:\ts_cmake_prebuild --parallel
+          if %errorlevel% neq 0 exit /b %errorlevel%
+
+  prebuild_macos_arm64:
+    name: Prebuild (macOS arm64)
+    runs-on: macos-14
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        id: prereq-cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-14-arm64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        id: ts-cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-14-arm64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Run macOS arm64 prebuild
+        if: steps.prereq-cache.outputs.cache-hit != 'true' || steps.ts-cache.outputs.cache-hit != 'true'
+        run: |
+          brew install nasm
+          brew uninstall --ignore-dependencies jpeg-turbo || true
+          bash ci-utils/install_prereq_linux.sh ./prereq_cache/local_install
+          cmake -S ci-utils/tensorstore_prebuild -B /tmp/ts_cmake_prebuild \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DTENSORSTORE_USE_SYSTEM_JPEG=ON \
+            -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache
+          cmake --build /tmp/ts_cmake_prebuild -j$(sysctl -n hw.ncpu)
+
+  # ──────────────────────────────────────────────────────────────────────────
+  # Stage 2 — Wheel-build jobs (4 Python versions in parallel per OS)
+  #   Each job needs only its own OS's prebuild, so it starts as soon as
+  #   that prebuild completes — independent of the other OSes.
+  # ──────────────────────────────────────────────────────────────────────────
+
+  build_wheels_linux:
+    name: Build wheels (Linux, ${{ matrix.cibw_build }})
+    runs-on: ubuntu-22.04
+    needs: [prebuild_linux]
+    continue-on-error: true
+    strategy:
+      matrix:
+        cibw_build: ["cp310", "cp311", "cp312", "cp313"]
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-ubuntu-22.04-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-ubuntu-22.04-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - name: Restore ccache
+        uses: actions/cache@v4
+        with:
+          path: ccache_dir
+          key: ccache-linux-auto64-${{ github.sha }}-${{ matrix.cibw_build }}
+          restore-keys: |
+            ccache-linux-auto64-${{ github.sha }}
+            ccache-linux-auto64-
+
       - uses: actions/setup-python@v4
-        name: Install Python
         with:
           python-version: '3.11'
 
       - name: Install cibuildwheel
-        run: |
-          python -m pip install cibuildwheel delvewheel wheel
+        run: python -m pip install cibuildwheel wheel
 
-      - name: Building wheels 
-        run: |
-          python -m cibuildwheel --output-dir dist
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
         env:
           CIBW_BUILD: ${{ matrix.cibw_build }}-*
           CIBW_BUILD_VERBOSITY: 3
           CIBW_SKIP: "*musllinux*"
           CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
+          CIBW_BEFORE_ALL_LINUX: bash /project/ci-utils/before_all_linux.sh
+          CIBW_ENVIRONMENT_LINUX: >-
+            LD_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64:$LD_LIBRARY_PATH"
+            ON_GITHUB="TRUE"
+            ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
+            CCACHE_DIR=/project/ccache_dir
+            CCACHE_MAXSIZE=5G
+            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON -DFETCHCONTENT_BASE_DIR=/project/ts_fc_cache"
+          CIBW_ARCHS: auto64
+          CIBW_BEFORE_TEST_LINUX: dnf -y install maven java
+          CIBW_TEST_REQUIRES: bfio>=2.4.0 tensorstore numpy pytest
+          CIBW_TEST_COMMAND: python -W default -m pytest -vv -s {project}/tests/python
+
+      - name: Upload Artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: argolid-wheels-linux-auto64-${{ matrix.cibw_build }}
+          path: dist/*.whl
+          retention-days: 1
+
+  build_wheels_macos_intel:
+    name: Build wheels (macOS intel, ${{ matrix.cibw_build }})
+    runs-on: macos-15-intel
+    needs: [prebuild_macos_intel]
+    continue-on-error: true
+    env:
+      MACOSX_DEPLOYMENT_TARGET: "10.15"
+    strategy:
+      matrix:
+        cibw_build: ["cp310", "cp311", "cp312", "cp313"]
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-15-intel-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-15-intel-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install cibuildwheel
+        run: python -m pip install cibuildwheel wheel
+
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
+        env:
+          CIBW_BUILD: ${{ matrix.cibw_build }}-*
+          CIBW_BUILD_VERBOSITY: 3
           CIBW_BEFORE_ALL_MACOS: brew install nasm &&
-                          bash ci-utils/install_prereq_linux.sh &&
-                          mkdir -p /tmp/argolid_bld &&
+                          bash ci-utils/before_all_macos.sh &&
                           brew uninstall --ignore-dependencies openssl || true &&
-                          sudo rm -rf /usr/local/include/openssl &&
-                          cp -r local_install /tmp/argolid_bld
-          CIBW_BEFORE_ALL_LINUX: curl -L https://www.nasm.us/pub/nasm/releasebuilds/2.15.05/nasm-2.15.05.tar.bz2 -o nasm-2.15.05.tar.bz2 &&
-                                   tar -xjf nasm-2.15.05.tar.bz2 &&
-                                   cd nasm-2.15.05 &&
-                                   ./configure && 
-                                   make && 
-                                   make install &&
-                                   cd .. &&
-                                   bash ci-utils/install_prereq_linux.sh &&
-                                   mkdir -p /tmp/argolid_bld &&
-                                   cp -r local_install /tmp/argolid_bld
-          CIBW_BEFORE_ALL_WINDOWS: ci-utils\install_prereq_win.bat &&
-                                   xcopy /E /I /y local_install C:\TEMP\argolid_bld\local_install 
+                          sudo rm -rf /usr/local/include/openssl
           CIBW_ENVIRONMENT_MACOS: >-
             MACOSX_DEPLOYMENT_TARGET=11.0
             REPAIR_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64"
             ON_GITHUB="TRUE"
             ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
-            CMAKE_ARGS="-DCMAKE_CXX_FLAGS=-Wno-missing-template-arg-list-after-template-kw -DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON"
-          CIBW_ENVIRONMENT_LINUX: >-
-            LD_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64:$LD_LIBRARY_PATH"
-            ON_GITHUB="TRUE" 
-            ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
-            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON"
+            CMAKE_ARGS="-DCMAKE_CXX_FLAGS=-Wno-missing-template-arg-list-after-template-kw -DTENSORSTORE_USE_SYSTEM_JPEG=ON -DTENSORSTORE_USE_SYSTEM_ZLIB=ON -DTENSORSTORE_USE_SYSTEM_PNG=ON -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache"
           CIBW_REPAIR_WHEEL_COMMAND_MACOS: DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-listdeps {wheel} && DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-wheel --require-archs {delocate_archs} -w {dest_dir} {wheel}
-          CIBW_ENVIRONMENT_WINDOWS: PATH="$TEMP\\argolid\\bin;$PATH" ON_GITHUB="TRUE" ARGOLID_DEP_DIR="C:\\TEMP\\argolid_bld\\local_install" CMAKE_ARGS="-DCMAKE_GENERATOR=Ninja"
+          CIBW_ARCHS: auto64
+
+      - name: Upload Artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: argolid-wheels-macos-15-intel-auto64-${{ matrix.cibw_build }}
+          path: dist/*.whl
+          retention-days: 1
+
+  build_wheels_windows:
+    name: Build wheels (Windows, ${{ matrix.cibw_build }})
+    runs-on: windows-latest
+    needs: [prebuild_windows]
+    continue-on-error: true
+    strategy:
+      matrix:
+        cibw_build: ["cp310", "cp311", "cp312", "cp313"]
+
+    steps:
+      - uses: actions/checkout@v3
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-windows-latest-auto64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-windows-latest-auto64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
+
+      - uses: ilammy/msvc-dev-cmd@v1
+        name: Add MSVS Path
+
+      - name: Add NASM
+        uses: ilammy/setup-nasm@v1
+
+      - name: Add Ninja
+        uses: seanmiddleditch/gha-setup-ninja@master
+
+      - uses: actions/setup-python@v4
+        with:
+          python-version: '3.11'
+
+      - name: Install cibuildwheel
+        run: python -m pip install cibuildwheel delvewheel wheel
+
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
+        env:
+          CIBW_BUILD: ${{ matrix.cibw_build }}-*
+          CIBW_BUILD_VERBOSITY: 3
+          CIBW_BEFORE_ALL_WINDOWS: ci-utils\before_all_win.bat
+          CIBW_ENVIRONMENT_WINDOWS: PATH="$TEMP\\argolid\\bin;$PATH" ON_GITHUB="TRUE" ARGOLID_DEP_DIR="C:\\TEMP\\argolid_bld\\local_install" CMAKE_ARGS="-DCMAKE_GENERATOR=Ninja -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache"
           CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: "delvewheel repair -w {dest_dir} {wheel}"
-          CIBW_ARCHS: ${{ matrix.cibw_archs }}
-          CIBW_BEFORE_TEST_LINUX: dnf -y install maven java
-          CIBW_TEST_REQUIRES: bfio>=2.4.0 tensorstore numpy pytest
-          CIBW_TEST_COMMAND: python -W default -m pytest -vv -s {project}/tests/python
+          CIBW_ARCHS: auto64
 
       - name: Upload Artifact
         uses: actions/upload-artifact@v4
         with:
-          name: argolid-wheels-${{ matrix.os }}-${{ matrix.cibw_archs }}-${{ matrix.cibw_build }}
+          name: argolid-wheels-windows-auto64-${{ matrix.cibw_build }}
           path: dist/*.whl
           retention-days: 1
 
   build_wheels_apple_arm64:
-    name: Build wheels on ${{ matrix.os }}
-    runs-on: ${{ matrix.os }}
+    name: Build wheels (macOS arm64, ${{ matrix.cibw_build }})
+    runs-on: macos-14
+    needs: [prebuild_macos_arm64]
     continue-on-error: true
     env:
       MACOSX_DEPLOYMENT_TARGET: "11.0"
     strategy:
       matrix:
-        os: [macos-14]
-        cibw_archs: ["arm64"]
         cibw_build: ["cp310", "cp311", "cp312", "cp313"]
 
     steps:
       - uses: actions/checkout@v3
-        name: Check out
+
+      - name: Restore prereq cache
+        uses: actions/cache@v4
+        with:
+          path: prereq_cache
+          key: prereq-macos-14-arm64-${{ hashFiles('ci-utils/install_prereq_linux.sh', 'ci-utils/install_prereq_win.bat') }}
+
+      - name: Restore TensorStore FetchContent cache
+        uses: actions/cache@v4
+        with:
+          path: ts_fc_cache
+          key: ts-fc-macos-14-arm64-${{ hashFiles('CMakeLists.txt', 'ci-utils/tensorstore_prebuild/CMakeLists.txt') }}
 
       - uses: actions/setup-python@v4
-        name: Install Python
         with:
           python-version: '3.11'
-  
+
       - name: Install cibuildwheel
-        run: |
-          python -m pip install cibuildwheel delocate wheel
+        run: python -m pip install cibuildwheel delocate wheel
 
-      - name: Building wheels 
-        run: |
-          python -m cibuildwheel --output-dir dist
+      - name: Build wheels
+        run: python -m cibuildwheel --output-dir dist
         env:
           CIBW_BUILD: ${{ matrix.cibw_build }}-*
           CIBW_BUILD_VERBOSITY: 3
           CIBW_ARCHS_MACOS: arm64
-          CIBW_BEFORE_ALL_MACOS:  brew install nasm &&
+          CIBW_BEFORE_ALL_MACOS: brew install nasm &&
                                   brew uninstall --ignore-dependencies jpeg-turbo &&
-                                  bash ci-utils/install_prereq_linux.sh &&
-                                  mkdir -p /tmp/argolid_bld &&
-                                  cp -r local_install /tmp/argolid_bld
+                                  bash ci-utils/before_all_macos.sh
           CIBW_ENVIRONMENT_MACOS: >-
             MACOSX_DEPLOYMENT_TARGET=11.0
             REPAIR_LIBRARY_PATH="/tmp/argolid_bld/local_install/lib:/tmp/argolid_bld/local_install/lib64"
             ON_GITHUB="TRUE"
             ARGOLID_DEP_DIR="/tmp/argolid_bld/local_install"
-            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON"
+            CMAKE_ARGS="-DTENSORSTORE_USE_SYSTEM_JPEG=ON -DFETCHCONTENT_BASE_DIR=${{ github.workspace }}/ts_fc_cache"
           CIBW_REPAIR_WHEEL_COMMAND_MACOS: |
             DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-listdeps {wheel} &&
             MACOSX_DEPLOYMENT_TARGET=11.0 DYLD_LIBRARY_PATH=$REPAIR_LIBRARY_PATH delocate-wheel --require-archs {delocate_archs} -w {dest_dir} {wheel} -e libc++ -e libunwind
-          CIBW_ARCHS: ${{ matrix.cibw_archs }}
+          CIBW_ARCHS: arm64
           CIBW_TEST_REQUIRES: bfio>=2.4.0 tensorstore numpy pytest
           CIBW_TEST_COMMAND: python -W default -m pytest -vv -s {project}/tests/python
 
       - name: Upload Artifact
         uses: actions/upload-artifact@v4
         with:
-          name: argolid-wheels-apple-arm64-${{ matrix.os }}-${{ matrix.cibw_archs }}-${{ matrix.cibw_build }}
+          name: argolid-wheels-apple-arm64-${{ matrix.cibw_build }}
           path: dist/*.whl
-          retention-days: 1
\ No newline at end of file
+          retention-days: 1
diff --git a/.gitignore b/.gitignore
index cce0baa..2b8b807 100644
--- a/.gitignore
+++ b/.gitignore
@@ -562,9 +562,14 @@ dmypy.json
 # Cython debug symbols
 cython_debug/
 
-# scratch 
+# scratch
 scratch/
 
+# CI build caches (populated by prebuild job, restored by wheel-build jobs)
+prereq_cache/
+ts_fc_cache/
+ccache_dir/
+
 build/
 build_man/
 venv/
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 34554cc..4f18f12 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -36,12 +36,21 @@ set(SOURCE
            src/cpp/utilities/utilities.cpp
 )
 
+# Use ccache if available to speed up repeated TensorStore compilation in CI
+find_program(CCACHE_EXECUTABLE ccache)
+if(CCACHE_EXECUTABLE)
+    message(STATUS "ccache found: ${CCACHE_EXECUTABLE}")
+    set(CMAKE_CXX_COMPILER_LAUNCHER "${CCACHE_EXECUTABLE}" CACHE STRING "" FORCE)
+    set(CMAKE_C_COMPILER_LAUNCHER   "${CCACHE_EXECUTABLE}" CACHE STRING "" FORCE)
+endif()
+
 include(FetchContent)
 
 FetchContent_Declare(
   tensorstore
   URL "https://github.com/google/tensorstore/archive/refs/tags/v0.1.80.tar.gz"
   URL_HASH SHA256=94866de34b6139d77d30e828a50f9e8df98e7dd68e848393470879aeb50ea7bf
+  FIND_PACKAGE_ARGS
 )
 
 # Additional FetchContent_Declare calls as needed...
diff --git a/ci-utils/before_all_linux.sh b/ci-utils/before_all_linux.sh
new file mode 100755
index 0000000..3dfef2d
--- /dev/null
+++ b/ci-utils/before_all_linux.sh
@@ -0,0 +1,33 @@
+#!/bin/bash
+# before_all_linux.sh
+# Replaces the inline CIBW_BEFORE_ALL_LINUX command.
+# Runs inside the manylinux container during the wheel-build job.
+# Expects:
+#   /project/prereq_cache/local_install  — populated by prebuild job (may be absent on cache miss)
+set -e
+
+cd /project
+
+# Always install NASM (required by TensorStore's aom/dav1d deps)
+curl -L https://www.nasm.us/pub/nasm/releasebuilds/2.15.05/nasm-2.15.05.tar.bz2 \
+     -o nasm.tar.bz2
+tar -xjf nasm.tar.bz2
+cd nasm-2.15.05 && ./configure && make -j"$(nproc)" && make install
+cd /project
+
+# Install ccache (CCACHE_DIR and CCACHE_MAXSIZE are set via CIBW_ENVIRONMENT_LINUX)
+dnf install -y ccache
+
+mkdir -p /tmp/argolid_bld
+
+if [ -d "/project/prereq_cache/local_install" ]; then
+    echo "==> Prereq cache hit: restoring filepattern + pybind11"
+    cp -r /project/prereq_cache/local_install /tmp/argolid_bld/local_install
+    # zlib/jpeg/png cannot be cached from /usr/local (container-internal path),
+    # so rebuild them quickly from source (~1 min total).
+    bash ci-utils/install_sys_deps_linux.sh
+else
+    echo "==> Prereq cache miss: building filepattern + pybind11 + sys deps from scratch"
+    bash ci-utils/install_prereq_linux.sh
+    cp -r /project/local_install /tmp/argolid_bld/local_install
+fi
diff --git a/ci-utils/before_all_macos.sh b/ci-utils/before_all_macos.sh
new file mode 100755
index 0000000..bef1520
--- /dev/null
+++ b/ci-utils/before_all_macos.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+# before_all_macos.sh
+# Replaces the inline CIBW_BEFORE_ALL_MACOS commands for wheel-build jobs.
+# brew uninstalls (openssl for intel, jpeg-turbo for arm64) should be done
+# by the caller (CIBW_BEFORE_ALL inline) BEFORE invoking this script.
+# Expects:
+#   prereq_cache/local_install  — populated by prebuild job (may be absent on cache miss)
+set -e
+
+mkdir -p /tmp/argolid_bld
+
+if [ -d "prereq_cache/local_install" ]; then
+    echo "==> Prereq cache hit: restoring filepattern + pybind11"
+    cp -r prereq_cache/local_install /tmp/argolid_bld/local_install
+    # Rebuild sys deps (zlib/jpeg/png) to /usr/local — these are fast and
+    # cannot be cached across jobs since they live outside the workspace.
+    bash ci-utils/install_sys_deps_linux.sh
+else
+    echo "==> Prereq cache miss: building filepattern + pybind11 + sys deps from scratch"
+    bash ci-utils/install_prereq_linux.sh
+    cp -r local_install /tmp/argolid_bld/local_install
+fi
diff --git a/ci-utils/before_all_win.bat b/ci-utils/before_all_win.bat
new file mode 100644
index 0000000..4c3deab
--- /dev/null
+++ b/ci-utils/before_all_win.bat
@@ -0,0 +1,24 @@
+@echo off
+REM before_all_win.bat
+REM Replaces the inline CIBW_BEFORE_ALL_WINDOWS command for wheel-build jobs.
+REM Expects:
+REM   prereq_cache\local_install — populated by prebuild job (may be absent on cache miss)
+
+if exist prereq_cache\local_install (
+    echo =^> Prereq cache hit: restoring filepattern + pybind11
+    xcopy /E /I /y prereq_cache\local_install C:\TEMP\argolid_bld\local_install
+    REM Replicate the DLL copy that install_prereq_win.bat does for PATH
+    if exist prereq_cache\local_install\bin (
+        xcopy /E /I /y prereq_cache\local_install\bin %TEMP%\argolid\bin
+    )
+) else (
+    echo =^> Prereq cache miss: building filepattern + pybind11 from scratch
+    call ci-utils\install_prereq_win.bat
+    if errorlevel 1 exit 1
+    xcopy /E /I /y local_install C:\TEMP\argolid_bld\local_install
+    REM Explicitly copy DLLs to PATH location (don't rely on install_prereq_win.bat's
+    REM conditional copy which requires ON_GITHUB to be already set)
+    if exist local_install\bin (
+        xcopy /E /I /y local_install\bin %TEMP%\argolid\bin
+    )
+)
diff --git a/ci-utils/install_prereq_linux.sh b/ci-utils/install_prereq_linux.sh
index 4b233bd..86ef95e 100755
--- a/ci-utils/install_prereq_linux.sh
+++ b/ci-utils/install_prereq_linux.sh
@@ -1,8 +1,16 @@
 #!/bin/bash
-# Usage: $bash install_prereq_linux.sh $INSTALL_DIR
+# Usage: bash install_prereq_linux.sh [$INSTALL_DIR] [--skip-sys-deps]
 # Default $INSTALL_DIR = ./local_install
+# --skip-sys-deps: skip building zlib, libjpeg-turbo, libpng (use when already installed)
 #
-if [ -z "$1" ]
+SKIP_SYS_DEPS=false
+for arg in "$@"; do
+    case "$arg" in
+        --skip-sys-deps) SKIP_SYS_DEPS=true ;;
+    esac
+done
+
+if [ -z "$1" ] || [ "$1" = "--skip-sys-deps" ]
 then
       echo "No path to the Argolid source location provided"
       echo "Creating local_install directory"
@@ -34,14 +42,15 @@ make install -j4
 cd ../../
 
 
+if [ "$SKIP_SYS_DEPS" != "true" ]; then
 curl -L https://github.com/madler/zlib/releases/download/v1.3.1/zlib131.zip -o zlib131.zip
 unzip zlib131.zip
 cd zlib-1.3.1
 mkdir build_man
 cd build_man
-cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_INSTALL_PREFIX=/usr/local ..  
-cmake --build . 
-cmake --build . --target install 
+cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_INSTALL_PREFIX=/usr/local ..
+cmake --build .
+cmake --build . --target install
 cd ../../
 
 curl -L https://github.com/libjpeg-turbo/libjpeg-turbo/archive/refs/tags/3.1.0.zip -o 3.1.0.zip
@@ -67,3 +76,4 @@ cd build_man
 cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_INSTALL_PREFIX=/usr/local ..
 make install -j4
 cd ../../
+fi
diff --git a/ci-utils/install_sys_deps_linux.sh b/ci-utils/install_sys_deps_linux.sh
new file mode 100755
index 0000000..24a74fb
--- /dev/null
+++ b/ci-utils/install_sys_deps_linux.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+# install_sys_deps_linux.sh
+# Builds and installs zlib, libjpeg-turbo, and libpng to /usr/local.
+# Works on Linux (manylinux) and macOS.
+# These are required by TensorStore when built with TENSORSTORE_USE_SYSTEM_* flags.
+set -e
+
+JOBS=$(nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null || echo 4)
+BUILD_TMPDIR="$(mktemp -d)"
+
+# zlib
+curl -L https://github.com/madler/zlib/releases/download/v1.3.1/zlib131.zip -o "${BUILD_TMPDIR}/zlib131.zip"
+unzip -q "${BUILD_TMPDIR}/zlib131.zip" -d "${BUILD_TMPDIR}"
+mkdir -p "${BUILD_TMPDIR}/zlib-1.3.1/build_man"
+cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_INSTALL_PREFIX=/usr/local \
+      -S "${BUILD_TMPDIR}/zlib-1.3.1" -B "${BUILD_TMPDIR}/zlib-1.3.1/build_man"
+cmake --build "${BUILD_TMPDIR}/zlib-1.3.1/build_man" -j"${JOBS}"
+cmake --build "${BUILD_TMPDIR}/zlib-1.3.1/build_man" --target install
+
+# libjpeg-turbo
+curl -L https://github.com/libjpeg-turbo/libjpeg-turbo/archive/refs/tags/3.1.0.zip \
+     -o "${BUILD_TMPDIR}/3.1.0.zip"
+unzip -q "${BUILD_TMPDIR}/3.1.0.zip" -d "${BUILD_TMPDIR}"
+mkdir -p "${BUILD_TMPDIR}/libjpeg-turbo-3.1.0/build_man"
+cmake -DCMAKE_INSTALL_PREFIX=/usr/local -DENABLE_STATIC=FALSE \
+      -DCMAKE_BUILD_TYPE=Release \
+      -S "${BUILD_TMPDIR}/libjpeg-turbo-3.1.0" \
+      -B "${BUILD_TMPDIR}/libjpeg-turbo-3.1.0/build_man"
+if [[ "$OSTYPE" == "darwin"* ]]; then
+  sudo cmake --build "${BUILD_TMPDIR}/libjpeg-turbo-3.1.0/build_man" \
+             --target install -j"${JOBS}"
+else
+  cmake --build "${BUILD_TMPDIR}/libjpeg-turbo-3.1.0/build_man" \
+        --target install -j"${JOBS}"
+fi
+
+# libpng
+curl -L https://github.com/glennrp/libpng/archive/refs/tags/v1.6.53.zip \
+     -o "${BUILD_TMPDIR}/v1.6.53.zip"
+unzip -q "${BUILD_TMPDIR}/v1.6.53.zip" -d "${BUILD_TMPDIR}"
+mkdir -p "${BUILD_TMPDIR}/libpng-1.6.53/build_man"
+cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_INSTALL_PREFIX=/usr/local \
+      -S "${BUILD_TMPDIR}/libpng-1.6.53" \
+      -B "${BUILD_TMPDIR}/libpng-1.6.53/build_man"
+cmake --build "${BUILD_TMPDIR}/libpng-1.6.53/build_man" \
+      --target install -j"${JOBS}"
+
+rm -rf "${BUILD_TMPDIR}"
diff --git a/ci-utils/prebuild_linux.sh b/ci-utils/prebuild_linux.sh
new file mode 100755
index 0000000..09ef2cb
--- /dev/null
+++ b/ci-utils/prebuild_linux.sh
@@ -0,0 +1,41 @@
+#!/bin/bash
+# prebuild_linux.sh
+# Runs inside the manylinux_2_28_x86_64 Docker container during the prebuild job.
+# Populates:
+#   /project/prereq_cache/local_install  — filepattern + pybind11 headers/libs
+#   /project/ts_fc_cache/               — TensorStore FetchContent source + build
+set -e
+
+cd /project
+
+# Install NASM (required by aom/dav1d inside TensorStore)
+curl -L https://www.nasm.us/pub/nasm/releasebuilds/2.15.05/nasm-2.15.05.tar.bz2 \
+     -o nasm.tar.bz2
+tar -xjf nasm.tar.bz2
+cd nasm-2.15.05 && ./configure && make -j"$(nproc)" && make install
+cd /project
+
+# Install ccache and point it at the workspace-mounted cache dir
+dnf install -y ccache
+mkdir -p /project/ccache_dir
+export CCACHE_DIR=/project/ccache_dir
+export CCACHE_MAXSIZE=5G
+
+# Build zlib, libjpeg-turbo, libpng to /usr/local (needed for TENSORSTORE_USE_SYSTEM_* flags)
+bash ci-utils/install_sys_deps_linux.sh
+
+# Build filepattern + pybind11 → prereq_cache/local_install
+# (--skip-sys-deps: sys deps already installed above)
+bash ci-utils/install_prereq_linux.sh ./prereq_cache/local_install --skip-sys-deps
+
+# Pre-build TensorStore using the minimal cmake project (no Python, no pybind11 needed).
+# FETCHCONTENT_BASE_DIR stores TensorStore source + build state in ts_fc_cache/
+# so wheel-build jobs can skip the expensive TensorStore download and reuse cmake state.
+cmake -S /project/ci-utils/tensorstore_prebuild \
+      -B /tmp/ts_cmake_prebuild \
+      -DFETCHCONTENT_BASE_DIR=/project/ts_fc_cache \
+      -DTENSORSTORE_USE_SYSTEM_JPEG=ON \
+      -DTENSORSTORE_USE_SYSTEM_ZLIB=ON \
+      -DTENSORSTORE_USE_SYSTEM_PNG=ON \
+      -DCMAKE_BUILD_TYPE=Release
+cmake --build /tmp/ts_cmake_prebuild -j"$(nproc)"
diff --git a/ci-utils/tensorstore_prebuild/CMakeLists.txt b/ci-utils/tensorstore_prebuild/CMakeLists.txt
new file mode 100644
index 0000000..5a30369
--- /dev/null
+++ b/ci-utils/tensorstore_prebuild/CMakeLists.txt
@@ -0,0 +1,33 @@
+cmake_minimum_required(VERSION 3.24)
+project(TensorStorePrebuild)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+# Use ccache if available to speed up repeated TensorStore compilation in CI
+find_program(CCACHE_EXECUTABLE ccache)
+if(CCACHE_EXECUTABLE)
+    message(STATUS "ccache found: ${CCACHE_EXECUTABLE}")
+    set(CMAKE_CXX_COMPILER_LAUNCHER "${CCACHE_EXECUTABLE}" CACHE STRING "" FORCE)
+    set(CMAKE_C_COMPILER_LAUNCHER   "${CCACHE_EXECUTABLE}" CACHE STRING "" FORCE)
+endif()
+
+include(FetchContent)
+
+# Keep URL + SHA256 in sync with root CMakeLists.txt
+# Cache key hashes both this file and root CMakeLists.txt
+FetchContent_Declare(
+  tensorstore
+  URL "https://github.com/google/tensorstore/archive/refs/tags/v0.1.80.tar.gz"
+  URL_HASH SHA256=94866de34b6139d77d30e828a50f9e8df98e7dd68e848393470879aeb50ea7bf
+)
+
+FetchContent_MakeAvailable(tensorstore)
+
+# Executable target forces TensorStore and all its transitive deps to be compiled.
+# No Python, no pybind11, no filepattern needed.
+add_executable(ts_prebuild_marker ts_prebuild_marker.cc)
+target_link_libraries(ts_prebuild_marker PRIVATE
+  tensorstore::tensorstore
+  tensorstore::all_drivers
+)
diff --git a/ci-utils/tensorstore_prebuild/ts_prebuild_marker.cc b/ci-utils/tensorstore_prebuild/ts_prebuild_marker.cc
new file mode 100644
index 0000000..9ac9b7d
--- /dev/null
+++ b/ci-utils/tensorstore_prebuild/ts_prebuild_marker.cc
@@ -0,0 +1,6 @@
+#include "tensorstore/open.h"
+
+// Minimal stub that forces the linker to pull in TensorStore symbols.
+// This file exists solely to make ts_prebuild_marker a non-trivial executable
+// so cmake compiles all TensorStore transitive dependencies during the prebuild.
+int main() { return 0; }