windows-msvc is RED on every PR: glm5_next_weights.cpp shadows 'v' and /WX makes it an error

[localai-org-maint-bot]

What

Both Windows jobs fail to COMPILE on every pull request:

glm5_next_weights.cpp(223,31): warning C4456: declaration of 'v' hides previous local declaration
glm5_next_weights.cpp(226,31): warning C4456: declaration of 'v' hides previous local declaration
glm5_next_weights.cpp(223,31): error C2220: the following warning is treated as an error

windows-msvc-cpu and windows-msvc-vulkan both die this way; the cause is identical in each.

KdaHeadCount declares const GgufValue* v in an if, then re-declares v in each else if rung. GCC and Clang do not warn here, so it built green on Linux and landed.

Introduced by b326ea003 (#2242 / #2292).

Why it reached main green, which is the part worth fixing

windows-msvc is a PR-only job with no main baseline. main never runs it, so a Windows-only compile break lands green and is then charged to whichever unrelated pull request runs CI next — every one of them, until someone fixes it.

This is structurally the same failure as #2389 (check-env-doc runs ahead of a push but never on main) and as the two env-doc reds before it: a gate that does not run on main converts a landed defect into a red on innocent branches. Three instances now, in three different gates.

Fix

Rename the shadowed declarations (v_group, v_inner). Behaviour-preserving, and it restores the Windows build. Fixed in-flow because it blocks an unrelated PR from reaching green, which is exactly the cost the mechanism above imposes.

Row: ENG-RELEASE-WINDOWS

[localai-org-maint-bot]

Fixed and verified on PR #2378: the C4456 shadowing is gone and both Windows jobs now COMPILE and run their suites.

A second, separate failure is behind it, and it is not this one. With the build fixed, windows-msvc-cpu reaches the tests and three api_server cases crash with rc=-1073740791 (STATUS_STACK_BUFFER_OVERRUN):

FAILING CASE 56/79 : api_server: an explicit-cpu device-selected engine serves /v1/completions
FAILING CASE 57/79 : api_server: embeddings dispatch — OpenAI shape over the engine path
FAILING CASE 59/79 : api_server: embeddings socket smoke; generate routes 404 on the embedding server

That is pre-existing and unrelated to the shadowing fix. Unrelated open PR #2399 fails on the same three case numbers with the same crash code (job 99495331472), on a branch sharing nothing with #2378. It is worth stating that the compile break was masking it: until this fix, no Windows run got far enough to execute these tests at all, so the crash was invisible behind a build error.

The shadowing fix therefore does not turn windows-msvc green; it removes the first of two failures and exposes the second. Whoever picks up the crash now has a suite that builds.

[localai-org-maint-bot]

Fixed in 5f230020f (#2378): KdaHeadCount no longer shadows v (renamed v_group / v_inner), and both Windows jobs now COMPILE.

Did not auto-close for the same reason as #2383 — the squash message comes from the PR body, which did not carry this keyword.

windows-msvc is still RED, on the separate pre-existing crash this fix exposed: three api_server cases exit -1073740791 (STATUS_STACK_BUFFER_OVERRUN). Verified pre-existing — unrelated PR #2399 fails on the same three case numbers with the same code. The compile break was masking it; the suite now builds far enough to reach it.