main does not compile under MSVC since #641, and check-windows-portability.py runs in no CI job, hook, or preflight

[localai-bot]

What

main (cefacd2d0, #641) fails scripts/check-windows-portability.py:

ERROR: src/vllm/multimodal/video_engine.cpp:21: unguarded POSIX include/call reaches Windows
ERROR: src/vllm/multimodal/video_engine.cpp:59: unguarded POSIX include/call reaches Windows
ERROR: src/vllm/multimodal/video_engine.cpp:64: unguarded POSIX include/call reaches Windows
EXIT=1

:21 is #include <sys/stat.h>; :59 and :64 are struct stat + ::stat inside
IsDir() / Exists(). The load-bearing part is S_ISDIR, which MSVC does not
define at all — MSVC does ship <sys/stat.h> and _stat, so the include alone
would survive; S_ISDIR does not.

CMakeLists.txt:919 puts the file in the main source list with no WIN32
guard, so it compiles into everything that links the library — including
server, which is on both Windows lanes' target lists.

This is a live regression, and the timings show it

Branch	base	windows-msvc-cpu	duration
#624, #636	pre-#641	fail	~22 min — reached the test run, died in test_openai_api_server.exe (#584)
#642, #643	cefacd2d0	fail	~1 min

A one-minute failure is a compile failure. Before #641 the lanes built for 22
minutes and got as far as executing tests. They no longer build at all.

Why nothing caught it

check-windows-portability.py runs nowhere. It is not in
.github/workflows/, not in .githooks/, and not in scripts/agent-preflight.sh:

$ grep -n 'check-windows-portability' .github/workflows/*.yml .githooks/* scripts/agent-preflight.sh
(no output)

So the checker is correct, current, and orphaned. A contributor running the full
local gate sees green; CI never runs it; and the MSVC lanes that would have caught
it were already red from #584, so a new red there reads as "the known Windows
failure" rather than "this PR stopped Windows compiling". #584 masked #603 the
same way — three defects deep on one lane now.

Scope

Two separable pieces:

1. Fix the break — guard the POSIX calls in video_engine.cpp, or provide the
   _WIN32 path. The repo already has both idioms (_putenv_s in
   tests/vllm/test_gguf.cpp:85), so this is a small change.
2. Wire the checker in — put check-windows-portability.py into
   agent-preflight.sh and/or a CI job, so an orphaned gate stops being a gate
   nobody runs. This changes what the gate enforces in practice, so per AGENTS.md
   it wants a red-before/green-after demonstration: cefacd2d0 is the red.

(2) is the one that matters. A checker that no automation runs is indistinguishable
from a deleted checker, and this one had been correct the whole time.

Found while reviewing #642/#643, whose Windows reds turned out not to be the
known #584/#603 pair at all.

[localai-bot]

Overlaps #648, filed three minutes after this one for the same break. Splitting them so neither is a duplicate.

#648 — "VideoEngine uses unguarded POSIX stat in native Windows build" — owns the
fix: the _WIN32 path for video_engine.cpp:21,59,64. That is the right home
for it; it is a src/ change on the row that added the file.

This issue should be read as the SECOND half only: the checker runs nowhere.

$ grep -n 'check-windows-portability' .github/workflows/*.yml .githooks/* scripts/agent-preflight.sh
(no output)

That is the part that outlives the fix. check-windows-portability.py is correct
and current, and it caught this break the moment it was pointed at main — but no
automation points it anywhere, so:

• feat(ltx-2.5): LTX-2.5 joint video+audio DiT, and a video seam that is no longer MiniMax-only (#435) #641 merged the break with both windows-msvc-* already red from test_openai_api_server crashes on Windows with STATUS_STACK_BUFFER_OVERRUN (0xC0000409), unmasked by the #512 fix #584, where a
  new red is indistinguishable from the known one;
• a contributor running the full local gate sees green, because preflight does not
  include it;
• an agent that ran it before a rebase and reported exit 0 was telling the truth
  and was still stale afterwards — which happened here, on record(TOOLS-PARSER-BREADTH): backfill the 41 parsers already shipped, and correct three counts (#608) #643, and was
  self-caught.

A checker no automation runs is indistinguishable from a deleted checker. Wiring it
into agent-preflight.sh and/or a CI job is the change that stops this recurring;
cefacd2d0 is the red-before evidence, and #648's fix will be the green-after.

Also worth recording for anyone triaging a Windows red: the lane now has three
distinct causes stacked — #584 (test_openai_api_server.exe → 0xC0000409, a
runtime crash), #603 (tests/vt/test_backend_cross_device.cpp bare POSIX setenv,
masked behind it), and now this compile break. "It's the known Windows failure" has
stopped being a safe read: pre-#641 the lane failed after ~22 minutes having built
successfully; it now fails in ~1.