windows-msvc CI is RED on main: test_backend_cross_device.cpp uses POSIX setenv/unsetenv, which MSVC does not provide

[localai-bot]

The windows-msvc-vulkan (and windows-msvc-cpu) CI legs fail on main and therefore on every PR branched from it.

tests\vt\test_backend_cross_device.cpp(1004,5): error C3861: setenv: identifier not found
tests\vt\test_backend_cross_device.cpp(1048,5): error C3861: setenv: identifier not found
tests\vt\test_backend_cross_device.cpp(1050,5): error C3861: unsetenv: identifier not found

setenv/unsetenv are POSIX; MSVC provides _putenv_s instead. The three call sites manipulate VT_FUSED_TIER around a tier sweep.

Introduced by 13bb7241f ("feat(backend): Metal W0 — enabling fixes + vt::Backend/Platform skeleton"). Confirmed present in git show origin/main:tests/vt/test_backend_cross_device.cpp.

Not introduced by any row currently in flight. Found while reviewing CI on #601, whose diff does not contain this file (git diff --name-only origin/main <branch> | grep test_backend_cross_device → no match). Any PR opened against current main will show this as a red Windows leg; the main baseline must be subtracted before attributing it.

Fix is a small portability shim — the repo already needs one wherever env vars are set from tests on Windows. Options: a helper that maps to _putenv_s under _MSC_VER, or guard the tier sweep out on MSVC (worse: it silently drops coverage on that platform).

Worth checking whether other tests set environment variables the same way, so the shim lands once rather than per-site.

[localai-bot]

Still live, and confirmed present on both Windows lanes' target lists.

Evidence gathered while attributing a red Windows lane on an unrelated PR:

• tests/vt/test_backend_cross_device.cpp calls bare POSIX setenv/unsetenv at
  :1004, :1048, :1050 — no _WIN32 guard anywhere in the file.
• tests/CMakeLists.txt:979 registers it unconditionally
  (vllm_cpp_add_test(test_backend_cross_device vt/test_backend_cross_device.cpp)),
  with a comment explaining the unconditional registration on capability grounds —
  which is right for backends and silent about portability.
• scripts/build-windows-release.ps1:263-284 builds a fixed target list that
  includes test_backend_cross_device on both the CPU and Vulkan lanes.

The portability pattern already exists in this repo — tests/vllm/v1/kv_offload/lmcache/test_lmcache_client.cpp:70
and tests/vllm/test_gguf.cpp:85 both use the MSVC _putenv_s form. This file
simply never adopted it.

Why it has been invisible

The lane REPORTS #584 (test_openai_api_server.exe exits -1073740791 /
0xC0000409), which is hit first. So every attribution of a red Windows lane
lands on #584 and stops there, and this one sits behind it unreported. Both are
live on the same lane; fixing #584 alone would likely just uncover this.

Worth noting for whoever fixes it: I initially searched tests/vllm/ for this
file, found nothing, and concluded it had been moved or fixed. It was in
tests/vt/ all along — a null grep proved my path wrong, not the file absent.

Not fixing it here: adding a _WIN32 shim to a test is a small change but it is
not mine to slip into an unrelated row, and the lane cannot be verified green
until #584 is also resolved.