diff --git a/.github/workflows/engines-cross-ci.yml b/.github/workflows/engines-cross-ci.yml
new file mode 100644
index 00000000000..1500d130bb5
--- /dev/null
+++ b/.github/workflows/engines-cross-ci.yml
@@ -0,0 +1,230 @@
+# Cross-platform lanes for the engines (parakeet-cpp/, tts-cpp/ — engines/*
+# after the repo reorg QIP; update the path filters + SRC map when it lands).
+#
+# Complements parakeet-ci.yml / tts-ci.yml (linux+mac build+test):
+#   - windows: full build + non-GPU ctest (MSVC, static ggml)
+#   - android: compile smoke, arm64-v8a via NDK (no device to run tests on)
+#   - ios:     compile smoke, arm64 device slice, Metal embedded (flags
+#              mirror build-xcframework.sh)
+# On-device e2e stays downstream in tetherto/qvac; these lanes exist to
+# catch cross-compile/link breaks *before* the registry pin chain.
+name: engines cross CI
+
+on:
+  push:
+    branches: [master]
+    paths:
+      - 'parakeet-cpp/**'
+      - 'tts-cpp/**'
+      - '.github/workflows/engines-cross-ci.yml'
+  pull_request:
+    paths:
+      - 'parakeet-cpp/**'
+      - 'tts-cpp/**'
+      - '.github/workflows/engines-cross-ci.yml'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: engines-cross-ci-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+env:
+  GGML_REPO: https://github.com/tetherto/qvac-ext-ggml.git
+  GGML_BRANCH: speech # single ggml pin shared with parakeet-ci.yml / tts-ci.yml
+
+jobs:
+  windows:
+    strategy:
+      fail-fast: false
+      matrix:
+        engine: [parakeet, tts]
+        include:
+          - engine: parakeet
+            src: parakeet-cpp
+            cmake_flags: -DPARAKEET_USE_SYSTEM_GGML=ON -DPARAKEET_BUILD_TESTS=ON -DPARAKEET_BUILD_EXAMPLES=OFF
+          - engine: tts
+            src: tts-cpp
+            cmake_flags: -DTTS_CPP_BUILD_TESTS=ON
+    runs-on: windows-2022
+    timeout-minutes: 90
+    defaults:
+      run:
+        shell: bash
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Resolve ggml pin (tip of ${{ env.GGML_BRANCH }})
+        id: ggml
+        run: echo "sha=$(git ls-remote "$GGML_REPO" "refs/heads/$GGML_BRANCH" | cut -f1)" >> "$GITHUB_OUTPUT"
+
+      - name: Cache ggml install
+        id: ggml-cache
+        uses: actions/cache@v4
+        with:
+          path: ggml-install
+          key: ggml-install-windows-2022-msvc-static-${{ steps.ggml.outputs.sha }}
+
+      # Static ggml on Windows sidesteps DLL discovery for the test exes.
+      - name: Build ggml (MSVC, static, CPU-only)
+        if: steps.ggml-cache.outputs.cache-hit != 'true'
+        run: |
+          # Fetch the exact SHA the cache key was derived from (the branch
+          # tip may advance between ls-remote and this step).
+          git init -q ggml-src
+          git -C ggml-src remote add origin "$GGML_REPO"
+          git -C ggml-src fetch --depth 1 origin ${{ steps.ggml.outputs.sha }}
+          git -C ggml-src checkout -q FETCH_HEAD
+          cmake -S ggml-src -B ggml-src/build -A x64 \
+            -DBUILD_SHARED_LIBS=OFF \
+            -DGGML_BUILD_TESTS=OFF \
+            -DGGML_BUILD_EXAMPLES=OFF \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build --config Release -j4
+          cmake --install ggml-src/build --config Release
+
+      - name: Configure
+        run: |
+          cmake -S ${{ matrix.src }} -B build -A x64 \
+            ${{ matrix.cmake_flags }} \
+            -DCMAKE_PREFIX_PATH="$PWD/ggml-install"
+
+      - name: Build
+        run: cmake --build build --config Release -j4
+
+      - name: Test (non-GPU)
+        run: ctest --test-dir build -C Release -LE 'gpu|perf' --output-on-failure --timeout 600
+
+  android:
+    strategy:
+      fail-fast: false
+      matrix:
+        engine: [parakeet, tts]
+        include:
+          - engine: parakeet
+            src: parakeet-cpp
+            cmake_flags: -DPARAKEET_USE_SYSTEM_GGML=ON -DPARAKEET_BUILD_TESTS=OFF -DPARAKEET_BUILD_EXAMPLES=OFF
+          - engine: tts
+            src: tts-cpp
+            cmake_flags: -DTTS_CPP_BUILD_TESTS=OFF
+    runs-on: ubuntu-24.04
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Resolve ggml pin (tip of ${{ env.GGML_BRANCH }})
+        id: ggml
+        run: echo "sha=$(git ls-remote "$GGML_REPO" "refs/heads/$GGML_BRANCH" | cut -f1)" >> "$GITHUB_OUTPUT"
+
+      - name: Cache ggml install
+        id: ggml-cache
+        uses: actions/cache@v4
+        with:
+          path: ggml-install
+          key: ggml-install-android-arm64-${{ steps.ggml.outputs.sha }}
+
+      - name: Build ggml (NDK arm64-v8a, static, CPU-only)
+        if: steps.ggml-cache.outputs.cache-hit != 'true'
+        run: |
+          # Fetch the exact SHA the cache key was derived from (the branch
+          # tip may advance between ls-remote and this step).
+          git init -q ggml-src
+          git -C ggml-src remote add origin "$GGML_REPO"
+          git -C ggml-src fetch --depth 1 origin ${{ steps.ggml.outputs.sha }}
+          git -C ggml-src checkout -q FETCH_HEAD
+          cmake -S ggml-src -B ggml-src/build \
+            -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK_LATEST_HOME/build/cmake/android.toolchain.cmake" \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DBUILD_SHARED_LIBS=OFF \
+            -DGGML_OPENMP=OFF \
+            -DGGML_BUILD_TESTS=OFF \
+            -DGGML_BUILD_EXAMPLES=OFF \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build -j4
+          cmake --install ggml-src/build
+
+      # Compile smoke only: hosted runners can't execute arm64 Android
+      # binaries. CMAKE_FIND_ROOT_PATH (not PREFIX_PATH) because the NDK
+      # toolchain restricts find_package() to the find-root.
+      - name: Configure + build engine (compile smoke)
+        run: |
+          cmake -S ${{ matrix.src }} -B build \
+            -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK_LATEST_HOME/build/cmake/android.toolchain.cmake" \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DCMAKE_BUILD_TYPE=Release \
+            ${{ matrix.cmake_flags }} \
+            -DCMAKE_FIND_ROOT_PATH="$PWD/ggml-install"
+          cmake --build build -j4
+
+  ios:
+    strategy:
+      fail-fast: false
+      matrix:
+        engine: [parakeet, tts]
+        include:
+          - engine: parakeet
+            src: parakeet-cpp
+            cmake_flags: -DPARAKEET_USE_SYSTEM_GGML=ON -DPARAKEET_BUILD_TESTS=OFF -DPARAKEET_BUILD_EXAMPLES=OFF
+          - engine: tts
+            src: tts-cpp
+            cmake_flags: -DTTS_CPP_BUILD_TESTS=OFF
+    runs-on: macos-14
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Resolve ggml pin (tip of ${{ env.GGML_BRANCH }})
+        id: ggml
+        run: echo "sha=$(git ls-remote "$GGML_REPO" "refs/heads/$GGML_BRANCH" | cut -f1)" >> "$GITHUB_OUTPUT"
+
+      - name: Cache ggml install
+        id: ggml-cache
+        uses: actions/cache@v4
+        with:
+          path: ggml-install
+          key: ggml-install-ios-arm64-${{ steps.ggml.outputs.sha }}
+
+      # Flags mirror build-xcframework.sh's iOS-device slice (Metal embedded).
+      - name: Build ggml (iOS device arm64, static, Metal embedded)
+        if: steps.ggml-cache.outputs.cache-hit != 'true'
+        run: |
+          # Fetch the exact SHA the cache key was derived from (the branch
+          # tip may advance between ls-remote and this step).
+          git init -q ggml-src
+          git -C ggml-src remote add origin "$GGML_REPO"
+          git -C ggml-src fetch --depth 1 origin ${{ steps.ggml.outputs.sha }}
+          git -C ggml-src checkout -q FETCH_HEAD
+          cmake -S ggml-src -B ggml-src/build \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_SYSROOT=iphoneos \
+            -DCMAKE_OSX_ARCHITECTURES=arm64 \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=16.4 \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DBUILD_SHARED_LIBS=OFF \
+            -DGGML_METAL=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_OPENMP=OFF \
+            -DGGML_BUILD_TESTS=OFF \
+            -DGGML_BUILD_EXAMPLES=OFF \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build -j4
+          cmake --install ggml-src/build
+
+      # Compile smoke only (no iOS device on hosted runners).
+      - name: Configure + build engine (compile smoke)
+        run: |
+          cmake -S ${{ matrix.src }} -B build \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_SYSROOT=iphoneos \
+            -DCMAKE_OSX_ARCHITECTURES=arm64 \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=16.4 \
+            -DCMAKE_BUILD_TYPE=Release \
+            ${{ matrix.cmake_flags }} \
+            -DCMAKE_FIND_ROOT_PATH="$PWD/ggml-install"
+          cmake --build build -j4
diff --git a/.github/workflows/parakeet-ci.yml b/.github/workflows/parakeet-ci.yml
new file mode 100644
index 00000000000..02ce8ee9d49
--- /dev/null
+++ b/.github/workflows/parakeet-ci.yml
@@ -0,0 +1,146 @@
+# CI for the parakeet engine (parakeet-cpp/ — engines/parakeet/ after the
+# repo reorg QIP lands; update the path filters below when it does).
+#
+# Scope (QIP "PR 0" lane): build the engine + every test harness against
+# system ggml (qvac-ext-ggml @ speech, the same fork the ggml-speech vcpkg
+# port ships), then run the non-GPU ctest suites. Tests whose model/reference
+# fixtures aren't present are registered as DISABLED by the CMakeLists and
+# show up as "Not Run" — the lane stays green with the committed fixtures
+# only. GPU suites (`ctest -L gpu`) need real GPUs → self-hosted lane, see
+# the gpu job stub at the bottom.
+name: parakeet CI
+
+on:
+  push:
+    branches: [master]
+    paths:
+      - 'parakeet-cpp/**'
+      - '.github/workflows/parakeet-ci.yml'
+  pull_request:
+    paths:
+      - 'parakeet-cpp/**'
+      - '.github/workflows/parakeet-ci.yml'
+  workflow_dispatch:
+    inputs:
+      run_gpu:
+        description: 'Also run the GPU ctest lane (needs a [self-hosted, gpu] runner)'
+        type: boolean
+        default: false
+
+permissions:
+  contents: read
+
+concurrency:
+  group: parakeet-ci-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+env:
+  GGML_REPO: https://github.com/tetherto/qvac-ext-ggml.git
+  GGML_BRANCH: speech # keep in sync with parakeet-cpp/scripts/setup-ggml.sh
+
+jobs:
+  build-test:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-24.04, macos-14]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install ccache
+        run: |
+          if [ "$RUNNER_OS" = "macOS" ]; then brew install ccache; else sudo apt-get update -q && sudo apt-get install -y ccache; fi
+          echo "CCACHE_DIR=${{ github.workspace }}/.ccache" >> "$GITHUB_ENV"
+
+      - name: Resolve ggml pin (tip of ${{ env.GGML_BRANCH }})
+        id: ggml
+        run: echo "sha=$(git ls-remote "$GGML_REPO" "refs/heads/$GGML_BRANCH" | cut -f1)" >> "$GITHUB_OUTPUT"
+
+      - name: Cache ggml install
+        id: ggml-cache
+        uses: actions/cache@v4
+        with:
+          path: ggml-install
+          key: ggml-install-${{ matrix.os }}-${{ steps.ggml.outputs.sha }}
+
+      - name: Cache ccache
+        uses: actions/cache@v4
+        with:
+          path: .ccache
+          key: ccache-parakeet-${{ matrix.os }}-${{ github.sha }}
+          restore-keys: ccache-parakeet-${{ matrix.os }}-
+
+      # CPU-only ggml: this lane runs `-LE gpu` tests, and hosted runners
+      # have no Vulkan/OpenCL; Metal is exercised by the gpu lane instead.
+      - name: Build ggml (system ggml, qvac-ext-ggml@${{ env.GGML_BRANCH }})
+        if: steps.ggml-cache.outputs.cache-hit != 'true'
+        run: |
+          # Fetch the exact SHA the cache key was derived from — the branch
+          # tip may have advanced since the ls-remote resolution, and a
+          # tip-only shallow clone would then silently build (and cache)
+          # a different commit than the key claims.
+          git init -q ggml-src
+          git -C ggml-src remote add origin "$GGML_REPO"
+          git -C ggml-src fetch --depth 1 origin ${{ steps.ggml.outputs.sha }}
+          git -C ggml-src checkout -q FETCH_HEAD
+          cmake -S ggml-src -B ggml-src/build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DBUILD_SHARED_LIBS=ON \
+            -DGGML_METAL=OFF \
+            -DGGML_BUILD_TESTS=OFF \
+            -DGGML_BUILD_EXAMPLES=OFF \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build -j4
+          cmake --install ggml-src/build
+
+      - name: Configure
+        run: |
+          cmake -S parakeet-cpp -B build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DPARAKEET_USE_SYSTEM_GGML=ON \
+            -DPARAKEET_BUILD_TESTS=ON \
+            -DCMAKE_PREFIX_PATH="$PWD/ggml-install"
+
+      - name: Build
+        run: cmake --build build -j4
+
+      # Excludes gpu (no GPU on hosted runners) and perf (timing gates are
+      # meaningless on shared runners). Model-dependent tests are DISABLED
+      # at configure time when fixtures are absent and report as "Not Run".
+      # LD_LIBRARY_PATH: the ggml install ships libqvac-speech-ggml.so whose
+      # own dependency on the ggml-cpu/base siblings is resolved transitively
+      # — Linux RUNPATH doesn't apply to transitive loads, so the loader
+      # needs the directory explicitly (macOS @rpath chains handle it).
+      - name: Test (non-GPU)
+        env:
+          LD_LIBRARY_PATH: ${{ github.workspace }}/ggml-install/lib
+        run: ctest --test-dir build -LE 'gpu|perf' --output-on-failure --timeout 600
+
+  # GPU lane stub: opt-in until self-hosted GPU runners exist. Trial note:
+  # hosted macos-14 exposes Metal, so moving `-L gpu` into the macOS
+  # build-test job is a cheap first experiment.
+  gpu:
+    if: github.event_name == 'workflow_dispatch' && inputs.run_gpu
+    runs-on: [self-hosted, gpu]
+    timeout-minutes: 90
+    steps:
+      - uses: actions/checkout@v4
+      # GGML_VULKAN=ON is the backend the -L gpu suites primarily target
+      # (test_vk_vs_cpu et al); Metal is on by default on Apple hosts.
+      # Revisit the flag set (OpenCL? CUDA?) when the GPU runner fleet is
+      # provisioned — a CPU-only ggml here would skip or falsely pass the
+      # GPU suites.
+      - name: Build ggml + engine (native backends) and run GPU suites
+        run: |
+          git clone --depth 1 --branch "$GGML_BRANCH" "$GGML_REPO" ggml-src
+          cmake -S ggml-src -B ggml-src/build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON \
+            -DGGML_VULKAN=ON \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build -j && cmake --install ggml-src/build
+          cmake -S parakeet-cpp -B build -DCMAKE_BUILD_TYPE=Release \
+            -DPARAKEET_USE_SYSTEM_GGML=ON -DPARAKEET_BUILD_TESTS=ON \
+            -DCMAKE_PREFIX_PATH="$PWD/ggml-install"
+          cmake --build build -j
+          ctest --test-dir build -L gpu --output-on-failure --timeout 1200
diff --git a/.github/workflows/tts-ci.yml b/.github/workflows/tts-ci.yml
new file mode 100644
index 00000000000..00234903010
--- /dev/null
+++ b/.github/workflows/tts-ci.yml
@@ -0,0 +1,146 @@
+# CI for the TTS engines (tts-cpp/ — engines/tts/ after the repo reorg QIP
+# lands; update the path filters below when it does).
+#
+# Scope (QIP "PR 0" lane): build chatterbox + supertonic + lavasr and every
+# test harness against system ggml (qvac-ext-ggml @ speech — mandatory here:
+# TTS_CPP_USE_SYSTEM_GGML defaults ON and the bundled path is rejected in
+# this tree), then run the non-GPU ctest suites (~68 pass with committed
+# fixtures; model-dependent ones auto-DISABLE to "Not Run"). GPU suites need
+# real GPUs → self-hosted lane, see the gpu job stub at the bottom.
+name: tts CI
+
+on:
+  push:
+    branches: [master]
+    paths:
+      - 'tts-cpp/**'
+      - '.github/workflows/tts-ci.yml'
+  pull_request:
+    paths:
+      - 'tts-cpp/**'
+      - '.github/workflows/tts-ci.yml'
+  workflow_dispatch:
+    inputs:
+      run_gpu:
+        description: 'Also run the GPU ctest lane (needs a [self-hosted, gpu] runner)'
+        type: boolean
+        default: false
+
+permissions:
+  contents: read
+
+concurrency:
+  group: tts-ci-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+env:
+  GGML_REPO: https://github.com/tetherto/qvac-ext-ggml.git
+  GGML_BRANCH: speech # single ggml pin shared with parakeet-ci.yml
+
+jobs:
+  build-test:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-24.04, macos-14]
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install ccache
+        run: |
+          if [ "$RUNNER_OS" = "macOS" ]; then brew install ccache; else sudo apt-get update -q && sudo apt-get install -y ccache; fi
+          echo "CCACHE_DIR=${{ github.workspace }}/.ccache" >> "$GITHUB_ENV"
+
+      - name: Resolve ggml pin (tip of ${{ env.GGML_BRANCH }})
+        id: ggml
+        run: echo "sha=$(git ls-remote "$GGML_REPO" "refs/heads/$GGML_BRANCH" | cut -f1)" >> "$GITHUB_OUTPUT"
+
+      - name: Cache ggml install
+        id: ggml-cache
+        uses: actions/cache@v4
+        with:
+          path: ggml-install
+          key: ggml-install-${{ matrix.os }}-${{ steps.ggml.outputs.sha }}
+
+      - name: Cache ccache
+        uses: actions/cache@v4
+        with:
+          path: .ccache
+          key: ccache-tts-${{ matrix.os }}-${{ github.sha }}
+          restore-keys: ccache-tts-${{ matrix.os }}-
+
+      # CPU-only ggml: this lane runs `-LE gpu` tests, and hosted runners
+      # have no Vulkan/OpenCL; Metal is exercised by the gpu lane instead.
+      # Cache key is shared with parakeet-ci.yml so whichever lane builds
+      # first serves both.
+      - name: Build ggml (system ggml, qvac-ext-ggml@${{ env.GGML_BRANCH }})
+        if: steps.ggml-cache.outputs.cache-hit != 'true'
+        run: |
+          # Fetch the exact SHA the cache key was derived from — the branch
+          # tip may have advanced since the ls-remote resolution, and a
+          # tip-only shallow clone would then silently build (and cache)
+          # a different commit than the key claims.
+          git init -q ggml-src
+          git -C ggml-src remote add origin "$GGML_REPO"
+          git -C ggml-src fetch --depth 1 origin ${{ steps.ggml.outputs.sha }}
+          git -C ggml-src checkout -q FETCH_HEAD
+          cmake -S ggml-src -B ggml-src/build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DBUILD_SHARED_LIBS=ON \
+            -DGGML_METAL=OFF \
+            -DGGML_BUILD_TESTS=OFF \
+            -DGGML_BUILD_EXAMPLES=OFF \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build -j4
+          cmake --install ggml-src/build
+
+      - name: Configure
+        run: |
+          cmake -S tts-cpp -B build \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DTTS_CPP_BUILD_TESTS=ON \
+            -DCMAKE_PREFIX_PATH="$PWD/ggml-install"
+
+      - name: Build
+        run: cmake --build build -j4
+
+      # Excludes gpu (no GPU on hosted runners) and perf (timing gates are
+      # meaningless on shared runners). Model-dependent tests are DISABLED
+      # at configure time when fixtures are absent and report as "Not Run".
+      # LD_LIBRARY_PATH: the ggml install ships libqvac-speech-ggml.so whose
+      # own dependency on the ggml-cpu/base siblings is resolved transitively
+      # — Linux RUNPATH doesn't apply to transitive loads, so the loader
+      # needs the directory explicitly (macOS @rpath chains handle it).
+      - name: Test (non-GPU)
+        env:
+          LD_LIBRARY_PATH: ${{ github.workspace }}/ggml-install/lib
+        run: ctest --test-dir build -LE 'gpu|perf' --output-on-failure --timeout 600
+
+  # GPU lane stub: opt-in until self-hosted GPU runners exist. Trial note:
+  # hosted macos-14 exposes Metal, so moving `-L gpu` / the mtl-* suites
+  # into the macOS build-test job is a cheap first experiment.
+  gpu:
+    if: github.event_name == 'workflow_dispatch' && inputs.run_gpu
+    runs-on: [self-hosted, gpu]
+    timeout-minutes: 90
+    steps:
+      - uses: actions/checkout@v4
+      # GGML_VULKAN=ON is the backend the -L gpu suites primarily target
+      # (test_vk_vs_cpu et al); Metal is on by default on Apple hosts.
+      # Revisit the flag set (OpenCL? CUDA?) when the GPU runner fleet is
+      # provisioned — a CPU-only ggml here would skip or falsely pass the
+      # GPU suites.
+      - name: Build ggml + engine (native backends) and run GPU suites
+        run: |
+          git clone --depth 1 --branch "$GGML_BRANCH" "$GGML_REPO" ggml-src
+          cmake -S ggml-src -B ggml-src/build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON \
+            -DGGML_VULKAN=ON \
+            -DCMAKE_INSTALL_PREFIX="$PWD/ggml-install"
+          cmake --build ggml-src/build -j && cmake --install ggml-src/build
+          cmake -S tts-cpp -B build -DCMAKE_BUILD_TYPE=Release \
+            -DTTS_CPP_BUILD_TESTS=ON \
+            -DCMAKE_PREFIX_PATH="$PWD/ggml-install"
+          cmake --build build -j
+          ctest --test-dir build -L gpu --output-on-failure --timeout 1200
diff --git a/tts-cpp/CMakeLists.txt b/tts-cpp/CMakeLists.txt
index ae2a3a11e4f..86ddb0e6a6e 100644
--- a/tts-cpp/CMakeLists.txt
+++ b/tts-cpp/CMakeLists.txt
@@ -755,7 +755,8 @@ if (TTS_CPP_BUILD_TESTS)
 
     add_executable(test-campplus
         test/test_campplus.cpp
-        src/campplus.cpp)
+        src/campplus.cpp
+        src/backend_selection.cpp)
     target_link_libraries(test-campplus PRIVATE ggml)
     target_include_directories(test-campplus PRIVATE ggml/include src)
     if (OpenMP_CXX_FOUND)
@@ -1331,7 +1332,8 @@ if (TTS_CPP_BUILD_TESTS)
     add_executable(test-campplus-backward-parity
         test/test_campplus_backward_parity.cpp
         src/campplus_backward.cpp
-        src/campplus.cpp)
+        src/campplus.cpp
+        src/backend_selection.cpp)
     target_link_libraries(test-campplus-backward-parity PRIVATE ggml)
     target_include_directories(test-campplus-backward-parity PRIVATE ggml/include src)
     if (OpenMP_CXX_FOUND)
diff --git a/tts-cpp/test/test_env_portable.h b/tts-cpp/test/test_env_portable.h
new file mode 100644
index 00000000000..1eee6f3dfa6
--- /dev/null
+++ b/tts-cpp/test/test_env_portable.h
@@ -0,0 +1,30 @@
+#pragma once
+// Portable setenv/unsetenv for the test harnesses. MSVC's CRT has no
+// POSIX setenv/unsetenv (only _putenv_s), so provide same-named shims
+// there; every other platform gets the real functions from <cstdlib>.
+#include <cstdlib>
+#include <initializer_list>
+#include <string>
+#ifdef _MSC_VER
+inline int setenv(const char * name, const char * value, int overwrite) {
+    // POSIX: with overwrite=0 an existing variable must be left untouched
+    // (and 0 returned) — don't let the shim diverge from that silently.
+    if (!overwrite && std::getenv(name) != nullptr) {
+        return 0;
+    }
+    return _putenv_s(name, value);
+}
+inline int unsetenv(const char * name) {
+    return _putenv_s(name, ""); // empty value removes the variable
+}
+#endif
+
+// Writable scratch directory for tests that emit temp fixtures. POSIX
+// runners set TMPDIR; Windows sets TEMP/TMP and has no /tmp, so fall
+// through the whole chain before defaulting.
+inline std::string test_tmpdir() {
+    for (const char * var : { "TMPDIR", "TEMP", "TMP" }) {
+        if (const char * v = std::getenv(var); v && *v) return v;
+    }
+    return "/tmp";
+}
diff --git a/tts-cpp/test/test_gguf_stream.cpp b/tts-cpp/test/test_gguf_stream.cpp
index 1dd5f90f6db..113fea1a276 100644
--- a/tts-cpp/test/test_gguf_stream.cpp
+++ b/tts-cpp/test/test_gguf_stream.cpp
@@ -8,6 +8,7 @@
 // chunk boundary) into a temp file, loads it through both paths, and
 // compares.
 
+#include "test_env_portable.h"
 #include "ggml.h"
 #include "ggml-backend.h"
 #include "ggml-cpu.h"
@@ -47,8 +48,7 @@ static const spec SPECS[] = {
 static const size_t N_SPECS = sizeof(SPECS) / sizeof(SPECS[0]);
 
 static std::string write_fixture() {
-    const char * tmpdir = getenv("TMPDIR");
-    std::string path = std::string(tmpdir ? tmpdir : "/tmp") + "/test-gguf-stream-fixture.gguf";
+    std::string path = test_tmpdir() + "/test-gguf-stream-fixture.gguf";
 
     size_t total = 0;
     for (size_t i = 0; i < N_SPECS; ++i) {
diff --git a/tts-cpp/test/test_lavasr_enhancer_ggml.cpp b/tts-cpp/test/test_lavasr_enhancer_ggml.cpp
index a25cd443441..8eb962a586b 100644
--- a/tts-cpp/test/test_lavasr_enhancer_ggml.cpp
+++ b/tts-cpp/test/test_lavasr_enhancer_ggml.cpp
@@ -17,6 +17,7 @@
 // is present it additionally compares the ggml forward against the onnxruntime
 // golden real/imag with the real model weights.
 
+#include "test_env_portable.h"
 #include "backend_selection.h"
 #include "lavasr/enhancer.h"      // scalar enhance() + enhance_with()
 #include "lavasr/enhancer_core.h"
@@ -505,9 +506,7 @@ static std::string write_enhancer_gguf(const EnhancerWeights & w) {
         add_gguf_tensor(ctx, g, w, e.name, e.ne);
     }
 
-    const char * tmpdir = std::getenv("TMPDIR");
-    std::string  path =
-        std::string(tmpdir ? tmpdir : "/tmp") + "/test-lavasr-enhancer-load.gguf";
+    std::string path = test_tmpdir() + "/test-lavasr-enhancer-load.gguf";
     const bool ok = gguf_write_to_file(g, path.c_str(), /*only_meta=*/false);
     gguf_free(g);
     ggml_free(ctx);
diff --git a/tts-cpp/test/test_s3gen_sched_equivalence.cpp b/tts-cpp/test/test_s3gen_sched_equivalence.cpp
index 31b7dfb7c9a..90cf0a7102e 100644
--- a/tts-cpp/test/test_s3gen_sched_equivalence.cpp
+++ b/tts-cpp/test/test_s3gen_sched_equivalence.cpp
@@ -23,6 +23,7 @@
 //
 // usage: test-s3gen-sched-equivalence MODEL.gguf [n_gpu_layers]
 
+#include "test_env_portable.h"
 #include "tts-cpp/chatterbox/s3gen_pipeline.h"
 
 #include <cmath>
diff --git a/tts-cpp/test/test_supertonic_profile_csv.cpp b/tts-cpp/test/test_supertonic_profile_csv.cpp
index 780fc376e76..d07565d8f18 100644
--- a/tts-cpp/test/test_supertonic_profile_csv.cpp
+++ b/tts-cpp/test/test_supertonic_profile_csv.cpp
@@ -37,6 +37,7 @@
 // Registered with `LABEL "unit"` in CMakeLists.txt — no GGUF
 // required.
 
+#include "test_env_portable.h"
 #include "supertonic_internal.h"
 
 #include <cstdio>
diff --git a/tts-cpp/test/test_supertonic_sched_equivalence.cpp b/tts-cpp/test/test_supertonic_sched_equivalence.cpp
index 169aa081c6e..918b490164e 100644
--- a/tts-cpp/test/test_supertonic_sched_equivalence.cpp
+++ b/tts-cpp/test/test_supertonic_sched_equivalence.cpp
@@ -25,6 +25,7 @@
 //
 // usage: test-supertonic-sched-equivalence MODEL.gguf [n_gpu_layers]
 
+#include "test_env_portable.h"
 #include "tts-cpp/supertonic/engine.h"
 
 #include <cmath>
diff --git a/tts-cpp/test/test_supertonic_vulkan_env_overrides.cpp b/tts-cpp/test/test_supertonic_vulkan_env_overrides.cpp
index c578c096eab..516c94293ff 100644
--- a/tts-cpp/test/test_supertonic_vulkan_env_overrides.cpp
+++ b/tts-cpp/test/test_supertonic_vulkan_env_overrides.cpp
@@ -51,6 +51,7 @@
 // Whole TU MUST fail to compile before the symbols are added,
 // then pass after.
 
+#include "test_env_portable.h"
 #include "tts-cpp/supertonic/engine.h"
 #include "supertonic_internal.h"
 
diff --git a/tts-cpp/test/test_t3_alignment_analyzer.cpp b/tts-cpp/test/test_t3_alignment_analyzer.cpp
index af35ef36f76..65c18372e71 100644
--- a/tts-cpp/test/test_t3_alignment_analyzer.cpp
+++ b/tts-cpp/test/test_t3_alignment_analyzer.cpp
@@ -11,6 +11,7 @@
 //   - token repetition only forces EOS once complete (avoids #519/#587 early cut)
 //   - short text / empty row / disabled => no force
 
+#include "test_env_portable.h"
 #include "t3_alignment_analyzer.h"
 
 #include <cstdio>
diff --git a/tts-cpp/test/test_t3_caches.cpp b/tts-cpp/test/test_t3_caches.cpp
index ee079d4eb0b..85cb5f7ee02 100644
--- a/tts-cpp/test/test_t3_caches.cpp
+++ b/tts-cpp/test/test_t3_caches.cpp
@@ -28,6 +28,7 @@
 // Without arguments, runs only the lightweight default-state
 // invariants (no model load required).
 
+#include "test_env_portable.h"
 #include "chatterbox_t3_internal.h"
 #include "chatterbox_tts_test_hooks.h"
 
diff --git a/tts-cpp/test/test_t3_sched_dispatch.cpp b/tts-cpp/test/test_t3_sched_dispatch.cpp
index 5bc3119a0ea..ba242c19b5b 100644
--- a/tts-cpp/test/test_t3_sched_dispatch.cpp
+++ b/tts-cpp/test/test_t3_sched_dispatch.cpp
@@ -12,6 +12,7 @@
 // Registered with `LABEL "unit"` in CMakeLists.txt so a fresh
 // checkout's `ctest` exercises this without needing any fixture.
 
+#include "test_env_portable.h"
 #include "backend_selection.h"
 #include "sched_dispatch.h"
 
diff --git a/tts-cpp/test/test_t3_sched_equivalence.cpp b/tts-cpp/test/test_t3_sched_equivalence.cpp
index 9cdcdc98305..c19cd043f4a 100644
--- a/tts-cpp/test/test_t3_sched_equivalence.cpp
+++ b/tts-cpp/test/test_t3_sched_equivalence.cpp
@@ -23,6 +23,7 @@
 // it, phase B must BYPASS the cached graphs, and phase A' proves they
 // were left intact.
 
+#include "test_env_portable.h"
 #include "chatterbox_t3_internal.h"
 
 #include "ggml.h"
diff --git a/tts-cpp/test/test_t3_stop_controller.cpp b/tts-cpp/test/test_t3_stop_controller.cpp
index 55a28c44844..c0fdff90c45 100644
--- a/tts-cpp/test/test_t3_stop_controller.cpp
+++ b/tts-cpp/test/test_t3_stop_controller.cpp
@@ -16,6 +16,7 @@
 //     EOS, no post-check stop.
 //   - make_mtl_stop_params budget scaling + env overrides.
 
+#include "test_env_portable.h"
 #include "t3_stop_controller.h"
 
 #include <cstdint>