fix(GATE-WINDOWS-WARNING-POLICY): /WX- is not /WX, and OBJCXX is not the C++ compile (#774)

[localai-bot]

Closes #774.

Spec: .agents/specs/windows-msvc-warning-policy-tokens.md
Roadmap issue table: .agents/roadmap_v1.md, #774 / ENG-RELEASE-WINDOWS.

Root cause

scripts/check-windows-portability.py:1710 asserted the MSVC warning policy
with a substring test:

if not all(token in warnings for token in ("/W4", "/WX")):

"/WX" in "/WX-" is True, and /WX- is MSVC's spelling for disable
warnings-as-errors. Two more blindnesses fall out of the same in, both
confirmed against this tree:

• "/W4" in "/W44996" is True — /W44996 sets warning C4996 to level 4, it
  does not raise the warning level.
• CMakeLists.txt:30 contains the literal /W4 /WX inside a # comment, so
  the prose alone satisfied the whole policy. Deleting every real flag in
  cmake/CompilerWarnings.cmake would still have passed.

And the assertion never asked which language the flag reaches. The measured
demonstration is PR #640 commit 74ba3823f (reverted in 15aa963a6 after
review), where the only bare /WX left was on $<COMPILE_LANGUAGE:OBJCXX> —
Objective-C++, the Metal backend, which never compiles under MSVC — while the
arm that matters got /WX-.

RED before

Same harness both sides: the HEAD test file run against the BASE checker
(af026e524) and against the HEAD checker, differing in nothing else.

===== BASE checker af026e524 + HEAD tests =====
FAILED ...::test_real_tree_msvc_warning_policy_reaches_the_cxx_compile
FAILED ...::test_rejects_policy_declared_only_in_a_comment
FAILED ...::test_rejects_policy_satisfied_only_on_objcxx
FAILED ...::test_rejects_prefix_lookalike_warning_flags
FAILED ...::test_rejects_warning_level_disabled_alongside_the_policy
FAILED ...::test_rejects_warnings_as_errors_disabled_by_wx_minus
6 failed, 1 passed, 71 deselected, 4 subtests passed

E   AssertionError: 0 == 0 : Windows portability contract OK

===== HEAD checker + HEAD tests =====
7 passed, 71 deselected, 9 subtests passed

The base failure message is the point: the checker reports Windows portability contract OK on a tree whose C++ arm says /WX-.

The one base test that passes is the inverse pin — an honest tree still passes
both before and after, so this is not "stricter about everything".

The fix

msvc_cxx_flag_text() reduces the CMake text to what can reach an MSVC C/C++
translation unit: # comments removed, and every generator expression whose
COMPILE_LANGUAGE names only non-C/C++ languages blanked (spans from a
$< / > stack scan, blanked in place so offsets survive, matching
without_set_source_properties). A genex naming no language is kept, because
it does reach C/C++, so the failure direction is never "silently excluded".

has_msvc_flag() matches with (?<![A-Za-z0-9_-]) / (?![A-Za-z0-9_-])
boundaries, case-sensitively — cl treats /w and /W4 as different flags.

The assertion then requires /W4 and /WX as tokens on that text, and refuses
the disable spellings of the same two flags — /WX-, /W0, /w — as a
separate error, because "absent" and "present and cancelled" are different
repairs.

cmake/CompilerWarnings.cmake is deliberately not edited: the tree already
satisfies the repaired policy, and a checker repair that also moved the thing it
checks could not show that.

The sibling evasions #640 also used — decided, not silently widened

• COMPILE_WARNING_AS_ERROR OFF: out of scope. CMake uses that property
  only to decide whether it adds a warnings-as-errors flag; nothing about it
  removes a literal flag from target_compile_options. Under the VS generator
  it lowers to MSBuild TreatWarningAsError=false, whose /WX- precedes
  AdditionalOptions, and MSVC takes the last of the pair — our explicit /WX
  still wins. The converse already fails closed: a tree that dropped the literal
  /WX in favour of the property would go RED on the missing token. Asserting a
  property value is a new policy about spelling, which check-windows-portability's /WX assertion is satisfied by /WX-, so the gate is blind to its own inversion #774 does not describe.
• Blanket /wd####: out of scope, and pinned as a deliberate non-goal by
  test_accepts_the_policy_on_every_shape_that_reaches_cxx. A per-warning
  suppression narrows what /W4 reports; it does not invert /W4 or /WX.
  Refusing it wholesale would forbid the legitimate targeted suppression (the
  tree has none today, so nothing is grandfathered in) and "how many is too
  many" is an undecided threshold. Any future /wd is a review judgement on its
  own PR.
• /W0 and /w: in scope, unlike the two above, because they are the
  disable spellings of /W4 itself. Closing /WX- while leaving its exact twin
  one character away would be the same defect check-windows-portability's /WX assertion is satisfied by /WX-, so the gate is blind to its own inversion #774 reports.

One pre-existing red had to be cleared to land this

check-pr-size.py's executable-evidence contract runs the whole recognized
module and requires it green at HEAD. tests/scripts/test_check_windows_portability.py
has been RED on main since e8a9e74e1 — verified on a pristine base worktree
at af026e524: 71 tests, 1 failure. That is the "one fake-runner case" of #680,
so no change to scripts/check-windows-portability.py could satisfy the gate.

It is a stale mutation anchor, not a checker gap. The checker reads
$calls.Add( in exactly one place — good_runner, the $good scriptblock of
Invoke-UnsupportedTierContractTests (check-windows-portability.py:1372) —
and script.replace("$calls.Add(", ..., 1) takes the first occurrence in the
file, which since #512/#583 added Invoke-CheckedContractTests above it lives
in a function the checker never looks at. The mutation left the governed text
intact, the checker correctly stayed green, and the assertion read that as a
defect.

The third commit anchors the mutation to the governed occurrence and asserts
that occurrence is unique. No assertion is deleted or widened, and the mutation
is now rejected by the base checker and this one — a fixture repair, not a
semantics change. #680 stays open for its other half: no workflow invokes
this suite at all.

Gates

• scripts/agent-preflight.sh — green (70 ok, 0 failed).
• python3 scripts/check-pr-size.py --base af026e524 --head HEAD — OK. That
  is the gate's own re-execution of the red-before/green-after evidence in an
  isolated worktree, so base_failed and head_passed are proven by the gate,
  not asserted by me.
• python3 scripts/check-commit-trailers.py --range af026e524..HEAD — OK.
• python3 -m unittest tests.scripts.test_check_windows_portability under
  check-pr-size's exact sanitized environment — 78 tests, OK.
• python3 -m pytest tests/scripts/ --ignore=tests/scripts/test_cpu_kernel_bench.py
  — 8 failed, 1368 passed, 3 skipped, 1541 subtests passed (18:51). All 8
  are pre-existing; the same 8 reproduce on a pristine origin/main worktree
  under the same names.

Baseline to subtract

Run on a /dev/shm worktree detached at origin/main, no diff applied:

SUBFAILED(shader='vt_attn_qk_norm_rope_gate.comp') test_gen_vulkan_spirv.py::CommittedArtifact::test_other_shaders_declare_none_yet
SUBFAILED(shader='vt_gdn_decode.comp')             (same test)
SUBFAILED(shader='vt_gdn_post_conv.comp')          (same test)
SUBFAILED(shader='vt_gdn_prefill.comp')            (same test)
SUBFAILED(shader='vt_matmul_vec.comp')             (same test)
SUBFAILED(shader='vt_sigmoid_gate_bf16.comp')      (same test)
FAILED test_mlx_system_headers.py::...::test_mlx_includes_have_source_scoped_clang_warning_suppression
FAILED test_now_render.py::DerivedFromPerRowRecords::test_a_spec_now_line_is_picked_up
8 failed, 20 passed, 2 skipped

test_now_render reads .agents/specs/now-derived.md, not the spec added here.

Separately, tests/scripts/test_cpu_x86_llamacpp_floor.py failed on an earlier
run of this branch and on the pristine base tree (af026e524: 2 failures,
including test_a_contended_leg_is_discarded_and_never_summarised, at loadavg
~29 with many agents on the box) and passed on the run above. It is
contention-sensitive, not attributable to this change. windows-msvc-* are
PR-only and red on every PR (#584).

CPU-only; no GPU, no benchmark.