[BUG]: tuple_size_structured_bindings test fails on MSVC — late tuple_size specialization not picked up by structured binding

[alliepiper]

Type of Bug

Runtime Error

Component

libcu++

Describe the bug

std/utilities/tuple/tuple.tuple/tuple.helper/tuple_size_structured_bindings.pass.cpp fails at line 153 on MSVC:

Assertion failed: p == -1, ... tuple_size_structured_bindings.pass.cpp, line 153

test_after_tuple_size_specialization declares Test const t{99}, then auto& [p] = t after a std::tuple_size<Test> specialization and matching get<0> returning -1. clang/gcc route the binding through the tuple protocol (p == -1); MSVC keeps binding to the underlying member (p == 99).

The test already guards an analogous NVRTC quirk:

#if !_CCCL_COMPILER(NVRTC) // nvbug4053842
  assert(p == -1);
#endif

MSVC needs an equivalent guard, or the structured-binding lookup needs to honor late tuple_size specializations under MSVC.

Reproduces on every MSVC libcu++ Test job: CTK12.9/MSVC14.44 and CTK13.2/MSVC14.50, C++20, RTX4090 and H100. Discovered while enabling libcu++ MSVC GPU jobs in #5725, which marks the test // UNSUPPORTED: msvc to unblock; this issue tracks the underlying behavior.

How to Reproduce

1. Check out PR Enable MSVC GPU jobs for CCCL projects #5725.
2. Build and run libcudacxx.test.lit against MSVC (CTK12.9/MSVC14.44 or CTK13.2/MSVC14.50, C++20).

Expected behavior

After specializing std::tuple_size<Test> with value = 1 and a matching get<0> returning -1, auto& [p] = t should bind p to -1, matching clang/gcc.

Operating System

Windows Server (CI runner image)

NVCC version

CTK 12.9 (MSVC 14.44.35207) and CTK 13.2 (MSVC 14.50.35717).

[miscco]

This is a bug in MSVC2026: https://developercommunity.visualstudio.com/t/VS-2026-fails-to-consider-specialization/11034185?q=tuple_size