build: make the tree compile on gcc 16, and add a CI lane so it stays that way

[jimmykarily]

::getpid() without <unistd.h> compiles on the distro g++ every Linux lane
installs (gcc 13 on ubuntu-latest) because a transitive include supplies the
declaration. On gcc 16 it does not, and five files in this tree rely on it —
one of them src/vllm/entrypoints/openai/server_main.cpp, a shipped binary.
Each gains the include it actually needs.

build-newest-gcc is a compile-only lane on container: gcc:16, so the class
cannot come back silently. It deliberately does not run ctest — build-test-cpu
owns execution — which keeps it to roughly one extra compile per pull request.
The image was confirmed to exist (library/gcc:16, pushed 2026-08-12), so the
lane will not fail on a missing tag.

cmake/CompilerWarnings.cmake stops treating -Warray-bounds as fatal on
GNU >= 16 only. That diagnostic fires inside libstdc++ and the vendored nlohmann
json for code that is correct; making the vendored headers SYSTEM does not
suppress it, because the reported location lands in our own translation unit
after inlining. The warning stays VISIBLE, and every compiler at 15 or below is
unchanged and still fails the build on a genuine out-of-bounds.

Maintainer change on top

The server_main.cpp include was added unguarded, among the plain includes.
That file is one of five in REQUIRED_CPP in
scripts/check-windows-portability.py, whose POSIX_PATTERNS[0] matches a bare
#include <unistd.h> on any line reachable from a Windows build — so as written
it would newly red both windows-msvc-* lanes, the opposite of this PR's intent.

Dropping it would also have been wrong: the ::getpid() that needs it (:158)
is guarded by #if defined(VT_BENCH_PROFILE_CONTROL) && !defined(_WIN32), and
that macro is set under VLLM_CPP_CUDA plus the opt-in
VLLM_CPP_BENCH_PROFILE_CONTROL. A CUDA build with the profiler enabled, on
gcc 16, genuinely needs the declaration — and this PR's own lane is CPU-only, so
it would not have caught that.

The include therefore moves into the VT_BENCH_PROFILE_CONTROL && !_WIN32 block
that already carries <cerrno>, <chrono>, <fcntl.h> and <sys/stat.h> for
the same feature. It is now guarded by exactly the condition that guards its only
user. check-windows-portability.py returns Windows portability contract OK.
The four test-file includes are untouched: none is in REQUIRED_CPP, none is
built on Windows, and each has an unguarded ::getpid() that needs the header.

The lane earned its keep immediately

build-newest-gcc failed on this very pull request, on a SIXTH file the five
include repairs above do not cover:

src/vllm/model_executor/models/qwen3_5.cpp:5296:27: error: '::sysconf' has not
  been declared; did you mean 'swscanf'?
src/vllm/model_executor/models/qwen3_5.cpp:5296:35: error: '_SC_PAGESIZE' was
  not declared in this scope

g++ (GCC) 16.2.0. A different symbol pair from the same header, so a ::getpid
sweep would never have found it; <sys/mman.h> was already included for the same
block and carried the file this far on gcc 13 through a transitive declaration.
Fixed in this branch, inside the #if defined(__unix__) guard the file already
has, which is the same condition guarding the call site.

Owed, not done here

build-newest-gcc is not wired into baseline-summary or
scripts/main-baseline.py's EXPECTED_JOBS, so the published baseline verdict
does not yet cover gcc 16. That is deliberate: #1155 is editing the same list to
address the structural half of #503, and two branches rewriting it would conflict. Wire it in after
whichever lands second.

Issue: #503

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: ClaudeCode:claude-opus-5 [ClaudeCode]
Signed-off-by: Dimitris Karakasilis dimitris@karakasilis.me

[localai-org-maint-bot]

The GCC 16 lane is useful, but the current head introduces a Windows portability regression while fixing the Linux one.

All five source changes add <unistd.h> unconditionally. MSVC does not provide that header, and these targets are part of the Windows build/test surface. Current main already fixed server_main.cpp with the portable split (<process.h> under _WIN32, POSIX handling otherwise), so rebasing the branch will also collide with or supersede that part of this diff. The four loader tests still call ::getpid() and need the same portable treatment (or a shared cross-platform process-ID helper), not an unconditional POSIX include.

The branch is currently 18 commits behind main and has no checks reported. Please rebase, preserve only the still-needed GCC 16 fixes, make the process-ID use portable on Windows, and let the new GCC lane plus the Windows lanes run on the updated head. I can re-review that result.

[localai-org-maint-bot]

Re-reviewing the updated head: the server include is now correctly guarded, but the original Windows blocker remains in all four loader tests. Each still includes <unistd.h> unconditionally and calls ::getpid(), so both Windows lanes fail. The new GCC 16 lane also fails, along with the documentation and commit-protocol gates.

Please apply the same portable process-ID treatment to the four tests (or use a shared helper), then resolve the actual GCC 16 compiler error from the failed log and update the required repository records/protocol metadata. The branch is mergeable now, but it is not ready while these real gates remain red.