diff --git a/.github/workflows/build-and-publish.yml b/.github/workflows/build-and-publish.yml
index 5d69d81..6b4a85f 100644
--- a/.github/workflows/build-and-publish.yml
+++ b/.github/workflows/build-and-publish.yml
@@ -69,7 +69,9 @@ jobs:
           CIBW_ARCHS_LINUX: "auto aarch64"
           CIBW_BEFORE_BUILD_LINUX: "yum install -y openblas-devel || apt-get install -y libopenblas-dev"
           CIBW_BEFORE_BUILD_WINDOWS: "pip install delvewheel && vcpkg install openblas:x64-windows"
-          CIBW_ENVIRONMENT_WINDOWS: "CMAKE_TOOLCHAIN_FILE=C:/vcpkg/scripts/buildsystems/vcpkg.cmake"
+          # The engine turns MSVC /WX off under scikit-build for downstream
+          # builds; keep it on for our own wheels.
+          CIBW_ENVIRONMENT_WINDOWS: "CMAKE_TOOLCHAIN_FILE=C:/vcpkg/scripts/buildsystems/vcpkg.cmake SKBUILD_CMAKE_ARGS=-DSDE_WARNINGS_AS_ERRORS=ON"
           # Bundle openblas.dll into the wheel (auditwheel/delocate handle this on Linux/macOS)
           CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: "delvewheel repair -w {dest_dir} {wheel} --add-path C:/vcpkg/installed/x64-windows/bin"
           # On free-threaded builds, assert the import did not re-enable the GIL.
diff --git a/.github/workflows/conda-recipe.yml b/.github/workflows/conda-recipe.yml
new file mode 100644
index 0000000..6493098
--- /dev/null
+++ b/.github/workflows/conda-recipe.yml
@@ -0,0 +1,37 @@
+name: conda-forge recipe
+
+# Builds conda.recipe/ with rattler-build against conda-forge, using the subset
+# of conda-forge's pinning in conda.recipe/variants.yaml (e.g. netlib
+# libblas/libcblas on Linux), as the feedstock would.
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+  workflow_dispatch:
+
+concurrency:
+  group: conda-recipe-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ ubuntu-latest, macos-14, windows-latest ]
+
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          submodules: recursive
+
+      - uses: prefix-dev/rattler-build-action@v0.2.39
+        with:
+          recipe-path: conda.recipe/recipe.yaml
+          rattler-build-version: v0.76.1
+          # --experimental enables load_from_file() for the version
+          build-args: --experimental --channel conda-forge
+          artifact-name: conda-package-${{ matrix.os }}
diff --git a/SparseDiffEngine b/SparseDiffEngine
index d8cb9b8..9ec1164 160000
--- a/SparseDiffEngine
+++ b/SparseDiffEngine
@@ -1 +1 @@
-Subproject commit d8cb9b88453dae339c74755f7c3082dafaccbecc
+Subproject commit 9ec1164079db7abecc1051513d84ca4da65e87be
diff --git a/conda.recipe/recipe.yaml b/conda.recipe/recipe.yaml
new file mode 100644
index 0000000..70394b3
--- /dev/null
+++ b/conda.recipe/recipe.yaml
@@ -0,0 +1,83 @@
+# Mirror of the conda-forge recipe (conda-forge/staged-recipes#34978), built
+# from this checkout instead of the PyPI sdist and without the BLAS patch, so
+# CI catches changes that would break the conda-forge feedstock.
+context:
+  name: sparsediffpy
+  version: ${{ load_from_file("pyproject.toml").project.version }}
+
+package:
+  name: ${{ name }}
+  version: ${{ version }}
+
+source:
+  path: ../
+
+build:
+  number: 0
+  skip:
+    - match(python, "<3.11")
+  script:
+    - if: linux
+      then:
+        - export SKBUILD_CMAKE_ARGS="-DBLA_VENDOR=Generic"
+        # scikit-build-core otherwise falls back to "Unix Makefiles", and make
+        # isn't in the build env
+        - export CMAKE_GENERATOR=Ninja
+    - ${{ PYTHON }} -m pip install . -vv --no-deps --no-build-isolation
+
+requirements:
+  build:
+    - if: build_platform != target_platform
+      then:
+        - python
+        - cross-python_${{ target_platform }}
+        - numpy
+    - ${{ compiler("c") }}
+    - ${{ stdlib("c") }}
+    - cmake
+    - ninja
+  host:
+    - python
+    - pip
+    - scikit-build-core >=0.10
+    - numpy
+    # macOS links the Accelerate framework
+    - if: linux
+      then:
+        - libblas
+        - libcblas
+        - blas-devel
+    - if: win
+      then:
+        - openblas
+        # openblas has no run_exports; libopenblas (which ships openblas.dll)
+        # does, so listing it here adds the runtime dependency
+        - libopenblas
+  run:
+    - python
+    - numpy >=2
+
+tests:
+  - python:
+      imports:
+        - sparsediffpy
+        - sparsediffpy._sparsediffengine
+      pip_check: true
+  - files:
+      recipe:
+        - test_sparsediffpy.py
+    requirements:
+      run:
+        - python
+        - numpy
+    script:
+      - python test_sparsediffpy.py
+
+about:
+  homepage: https://github.com/SparseDifferentiation/SparseDiffPy
+  repository: https://github.com/SparseDifferentiation/SparseDiffPy
+  summary: Python bindings for SparseDiffEngine algorithmic differentiation
+  license: Apache-2.0
+  license_file:
+    - LICENSE
+    - SparseDiffEngine/LICENSE
diff --git a/conda.recipe/test_sparsediffpy.py b/conda.recipe/test_sparsediffpy.py
new file mode 100644
index 0000000..49dbd09
--- /dev/null
+++ b/conda.recipe/test_sparsediffpy.py
@@ -0,0 +1,31 @@
+# Smoke test: value, gradient and Hessian of small expressions through the
+# C bindings. The dense quad_form path goes through CBLAS, so this also
+# checks that BLAS is linked and callable.
+import numpy as np
+
+from sparsediffpy import _sparsediffengine as de
+
+
+def check(expr, u, value, grad, hess):
+    prob = de.make_problem(expr, [], False)
+    de.problem_init_jacobian_coo(prob)
+    de.problem_init_hessian_coo_lower_triangular(prob)
+    np.testing.assert_allclose(de.problem_objective_forward(prob, u), value)
+    np.testing.assert_allclose(de.problem_gradient(prob), grad)
+    rows, cols, _ = de.get_problem_hessian_sparsity_coo(prob)
+    vals = de.problem_eval_hessian_vals_coo(prob, 1.0, np.zeros(0))
+    H = np.zeros((u.size, u.size))
+    H[rows, cols] = vals
+    np.testing.assert_allclose(np.tril(H), np.tril(hess))
+
+
+u = np.array([1.0, -2.0, 0.5])
+
+x = de.make_variable(3, 1, 0, 3)
+check(de.make_sum(de.make_exp(x), -1), u, np.exp(u).sum(), np.exp(u), np.diag(np.exp(u)))
+
+P = np.array([[2.0, 1.0, 0.0], [1.0, 3.0, 0.0], [0.0, 0.0, 4.0]])
+x = de.make_variable(3, 1, 0, 3)
+check(de.make_quad_form(None, x, "dense", P.flatten(order="F"), 3), u, u @ P @ u, 2 * P @ u, 2 * P)
+
+print("sparsediffpy smoke test passed")
diff --git a/conda.recipe/variants.yaml b/conda.recipe/variants.yaml
new file mode 100644
index 0000000..2c59b55
--- /dev/null
+++ b/conda.recipe/variants.yaml
@@ -0,0 +1,36 @@
+# Subset of conda-forge's global pinning (conda-forge-pinning-feedstock
+# recipe/conda_build_config.yaml), which rattler-build can't load directly.
+# Only the oldest and newest supported Pythons are built.
+c_compiler:
+  - if: linux
+    then: gcc
+  - if: osx
+    then: clang
+  - if: win
+    then: vs2022
+c_compiler_version:
+  - if: linux
+    then: "15"
+  - if: osx
+    then: "21"
+c_stdlib:
+  - if: linux
+    then: sysroot
+  - if: osx
+    then: macosx_deployment_target
+  - if: win
+    then: vs
+c_stdlib_version:
+  - if: linux
+    then: "2.17"
+  - if: osx
+    then: "11.0"
+python:
+  - 3.11.* *_cpython
+  - 3.14.* *_cp314
+# conda-forge builds against netlib so packages link the switchable
+# libblas.so.3/libcblas.so.3
+libblas:
+  - 3.9.* *netlib
+libcblas:
+  - 3.9.* *netlib