fix(metal): scope MLX AppleClang warning suppression

docs/BENCHMARKS.md

@@ -6883,11 +6883,12 @@ us by about 1.5 points.
 The default (non-MLX) build is **95.9%** against this corrected baseline, not
 96.4%.
 
-The 2026-08-03 MLX system-header dependency correction is **NOT APPLICABLE** to
-benchmark results: it replaces translation-unit warning suppression with an
-imported CMake dependency whose public headers are `SYSTEM`. Darwin CI remains
-the build-verification gate; no runtime path, measurement, or binding number
-changed.
+The 2026-08-04 MLX system-header warning correction is **NOT APPLICABLE** to
+benchmark results: the imported CMake dependency keeps its public headers
+`SYSTEM`, and the provider now also scopes AppleClang's
+`-Wgnu-folding-constant` suppression to those MLX public includes. Darwin CI
+remains the build-verification gate; no runtime path, measurement, or binding
+number changed.
 
 **Everything qualitative in the entry below still holds** — MLX wins prefill, the
 shape gate is the right disposition, the fallback hoist was worth 27.2 vs 17.8 —

docs/STATUS.md

@@ -1447,8 +1447,10 @@ per-lever chronology: [docs/BENCHMARKS.md](BENCHMARKS.md) and
 [.agents/specs/metal-dispatch-attribution.md](../.agents/specs/metal-dispatch-attribution.md).
 The MLX-enabled Darwin build remains **build-verification pending**: MLX is now
 an imported CMake dependency whose public include directory is explicitly
-`SYSTEM`; third-party header diagnostics stay outside project `-Werror`, while
-vllm.cpp warnings remain fatal. Darwin CI is the binding AppleClang gate.
+`SYSTEM`, and the provider additionally scopes AppleClang's
+`-Wgnu-folding-constant` suppression to the MLX public includes. Third-party
+header diagnostics stay outside project `-Werror`, while vllm.cpp warnings
+remain fatal. Darwin CI is the binding AppleClang gate.
 
 **CUDA architectures.** The production target is GB10/`sm_121a` (runtime-gated,
 both gate models token-exact + at/above vLLM speed). The arch-additivity

src/vt/metal/metal_mlx_provider.mm

@@ -52,12 +52,19 @@
 
 // MLX public headers. Deliberately NOT mlx/backend/metal/*: those pull in
 // metal-cpp and, as noted above, their entry points are not exported anyway.
+#if defined(__clang__)
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wgnu-folding-constant"
+#endif
 #include "mlx/allocator.h"
 #include "mlx/array.h"
 #include "mlx/dtype.h"
 #include "mlx/ops.h"
 #include "mlx/stream.h"
 #include "mlx/transforms.h"
+#if defined(__clang__)
+#pragma clang diagnostic pop
+#endif
 
 #include "metal_buffers.h"
 #include "metal_context.h"

tests/scripts/test_mlx_system_headers.py

@@ -15,8 +15,50 @@
 
 
 class MlxSystemHeadersTest(unittest.TestCase):
-    def test_dependency_module_is_shipped(self) -> None:
-        self.assertTrue((ROOT / "cmake" / "MLXDependency.cmake").is_file())
+    def test_dependency_target_marks_mlx_headers_as_system(self) -> None:
+        dependency = (ROOT / "cmake" / "MLXDependency.cmake").read_text(
+            encoding="utf-8"
+        )
+        self.assertIn("INTERFACE_SYSTEM_INCLUDE_DIRECTORIES", dependency)
+
+    def test_mlx_includes_have_source_scoped_clang_warning_suppression(self) -> None:
+        provider = (ROOT / "src" / "vt" / "metal" / "metal_mlx_provider.mm").read_text(
+            encoding="utf-8"
+        )
+        required_pragmas = (
+            "#pragma clang diagnostic push",
+            '#pragma clang diagnostic ignored "-Wgnu-folding-constant"',
+            "#pragma clang diagnostic pop",
+        )
+        for pragma in required_pragmas:
+            self.assertIn(pragma, provider)
+        self.assertEqual(provider.count("#if defined(__clang__)"), 2)
+        self.assertIn(
+            "#if defined(__clang__)\n"
+            "#pragma clang diagnostic push\n"
+            '#pragma clang diagnostic ignored "-Wgnu-folding-constant"\n'
+            "#endif\n",
+            provider,
+        )
+        self.assertIn(
+            '#include "mlx/transforms.h"\n'
+            "#if defined(__clang__)\n"
+            "#pragma clang diagnostic pop\n"
+            "#endif\n",
+            provider,
+        )
+        push = provider.index("#pragma clang diagnostic push")
+        ignored = provider.index(
+            '#pragma clang diagnostic ignored "-Wgnu-folding-constant"'
+        )
+        first_mlx_include = provider.index('#include "mlx/allocator.h"')
+        last_mlx_include = provider.index('#include "mlx/transforms.h"')
+        pop = provider.index("#pragma clang diagnostic pop")
+        first_project_include = provider.index('#include "metal_buffers.h"')
+        self.assertLess(push, ignored)
+        self.assertLess(ignored, first_mlx_include)
+        self.assertLess(last_mlx_include, pop)
+        self.assertLess(pop, first_project_include)
 
     @unittest.skipUnless(
         shutil.which("cmake") and shutil.which("clang++") and shutil.which("ar"),