Problem

ICX 2026 on Windows strictly enforces that C99 _Complex double cannot be passed to UCRT's creal()/cimag() functions, which expect _Dcomplex (an MSVC struct type). This causes hard compilation errors in all three build targets (mkl_umath_loops, _ufuncs, _patch_numpy).

Root cause chain:

1. CMakeLists.txt sets C_STANDARD 99 → CMake injects -Qstd:c99 for ICX
2. With -Qstd:c99, ICX enables strict C99 complex type semantics
3. numpy's npy_common.h defines npy_cdouble as double _Complex for __INTEL_LLVM_COMPILER (regardless of platform)
4. numpy's npy_math.h calls creal(z) / cimag(z) where z is npy_cdouble
5. UCRT's creal() expects _Dcomplex — ICX 2025 emitted a warning, ICX 2026 makes it a hard error

This is why Intel's own CI did not catch this: Intel's channel provides dpcpp_win-64 2025.x which was lenient. conda-forge uses dpcpp_win-64 2026.0 which enforces strict types.

Fix

Unset C_STANDARD for all three targets when building on Windows with IntelLLVM. Without -Qstd:c99, ICX operates in its default clang-cl compatibility mode and handles the _Complex/_Dcomplex coercion without error — matching the behavior of ICX 2025 and earlier.

if(WIN32 AND CMAKE_C_COMPILER_ID STREQUAL "IntelLLVM")
  set_target_properties(<target> PROPERTIES C_STANDARD "")
endif()

Note

The underlying incompatibility is in numpy's npy_math.h, which should guard creal/cimag calls for __INTEL_LLVM_COMPILER on Windows — similar to how npy_common.h already guards the npy_cdouble typedef. This fix works around it at the mkl_umath build level until numpy addresses it upstream.