[BUG] Ambiguous union_member::get overloads with GCC 16 break ORC/Parquet writer builds

Describe the bug

With GCC 16 as the host compiler, the constrained overloads of union_member::get in cpp/src/io/statistics/typed_statistics_chunk.cuh become ambiguous, and libcudf fails to build.

The overload set (starting at typed_statistics_chunk.cuh:35) is:

template <typename T, typename U>
__device__ static type<T, U> get(U& val)
  requires(std::is_integral_v<T> and std::is_unsigned_v<T>)   // line 56

template <typename T, typename U>
__device__ static type<T, U> get(U& val)
  requires(std::is_integral_v<T> and std::is_signed_v<T>)     // line 63

template <typename T, typename U>
__device__ static type<T, U> get(U& val)
  requires(std::is_same_v<T, __int128_t>)                     // line 70

Under GCC 16 the line 56 and line 63 overloads both become viable for the same T, so the call is ambiguous:

typed_statistics_chunk.cuh(268): error: more than one instance of overloaded function
  "cudf::io::union_member::get" matches the argument list:
    function template "cudf::io::union_member::type<T, U> cudf::io::union_member::get<T,U>(U &)" (declared at line 56)
    function template "cudf::io::union_member::type<T, U> cudf::io::union_member::get<T,U>(U &)" (declared at line 63)
    argument types are: (cudf::io::statistics_val)
      union_member::get<E>(stat.min_value) = chunk.minimum_value;
      ^
typed_statistics_chunk.cuh(84): note #3473-D: substituting
  "<__int128_t, cudf::io::statistics_val>" for function template
  "cudf::io::union_member::get<T,U>(U &)" failed constraints

It reproduces on the fixed-point statistics path, with T=numeric::decimal32 and T=numeric::decimal64:

instantiation of "cudf::io::statistics_chunk cudf::io::get_untyped_chunk(
  const cudf::io::typed_statistics_chunk<T, include_aggregate> &)
  [with T=numeric::decimal32, include_aggregate=true]"
  at line 123 of cpp/src/io/statistics/column_statistics.cuh
instantiation of "void cudf::io::calculate_group_statistics_functor<...>::operator()<T>(...)
  [with block_size=256, IO=cudf::io::detail::io_file_format::ORC, ..., T=numeric::decimal32]"
  at line 318 of cpp/src/io/statistics/column_statistics.cuh

Affected lines in typed_statistics_chunk.cuh are 139-142, 180-181 and 263-269.

Steps/Code to reproduce bug

Build libcudf with GCC 16 as host compiler and CUDA 13.4:

nvcc -ccbin=/opt/rh/gcc-toolset-16/root/usr/bin/g++ -std=c++20 ...

These two objects fail to build (both are required for libcudf.so):

src/io/orc/writer_impl.cu.o
src/io/parquet/writer_impl.cu.o

Expected behavior

The requires constraints remain mutually exclusive, and union_member::get resolves unambiguously for the decimal types.

Isolation: this is the host compiler, not nvcc

The diagnostic is emitted by the device frontend, so it reads like an nvcc problem, but it is driven by the host compiler's <type_traits>. I recompiled the identical failing TU changing only -ccbin, with CUDA 13.4 held constant:

So GCC 16's libstdc++ changes the relevant type-trait results such that the two constraints are no longer mutually exclusive for these types.

I did not attempt a fix, since disambiguating this dispatch correctly depends on the intended semantics for the fixed-point/decimal statistics path, and guessing risks silently wrong statistics.

Environment details

• cuDF: commit 456580fcdd726380dcb3de6b9686e7d47d6dd0a2 (release/26.10)
• Compiler: GCC 16.2.1 20260807 (Red Hat 16.2.1-1, gcc-toolset-16)
• CUDA: 13.4 (nvcc V13.4.92)
• CCCL: 3.5.0
• OS: CentOS Stream 9
• Standard: -std=gnu++20
• Consumer: Velox (VELOX_ENABLE_CUDF=ON), which builds libcudf from source via CPM

Note: GCC 16 requires CUDA >= 13, since CUDA 12.x nvcc hard-rejects GCC > 14.

Additional context

Found while evaluating GCC 16 support for Velox's cuDF backend. This is currently the blocking issue for that configuration: with it unresolved, libcudf.so cannot be produced and the cuDF test suite cannot run.

A separate, unrelated GCC 16 issue in cuDF (missing #include <variant> in cudf/ast/expressions.hpp) is filed separately. Two further GCC 16 issues live in other repos: rtcx (missing <variant> in include/rtcx/embed.hpp) and cuco (includes the deprecated <cuda/stream_ref>, which CCCL 3.5 warns on and -Werror promotes to an error).