test: validate malformed Git paths without filesystem assumptions

architecture/notes/scripts/ci/2026-09-23-prohibited-name-range-check.md

@@ -14,6 +14,10 @@ The repository already used an ignored `scripts/check_prohibited_names.py` in st
 - The local checker already owns the prohibited-name list and hook behavior; duplicating the list in workflow YAML would create a second policy source.
 - Inline copyright/third-party attribution can contain a project name, while Cargo git/source references and protocol/external-format identifiers include key-reporting constants, multiplexer escape sequences, and theme-format variants.
 
+- macOS run 35950152526 rejected a malformed UTF-8 filename before the checker
+  could run. The fixture now writes a Git tree directly, preserving the same
+  invalid-path regression without depending on filesystem filename support.
+
 ## Decision
 
 - Publish the existing checker as the single policy implementation and add a fail-closed range mode.
@@ -34,7 +38,7 @@ New PR/merge-group/main-push text and commit subjects are checked by the existin
 
 ## Validation
 
-`scripts/tests/test_prohibited_names.py` covers ordinary comparison failure, inline attribution positive/negative paths, protocol identifiers, Cargo dependency references, GitHub CLI text, range text, range commit messages, diverged related bases, unrelated/missing bases, old-history exclusion, added-then-deleted text, and merge-resolution additions. The checker also covers malformed UTF-8 paths/text and normal Unicode/control-character paths; the raw-byte path fixture runs on POSIX, where that filename can exist.
+`scripts/tests/test_prohibited_names.py` covers ordinary comparison failure, inline attribution positive/negative paths, protocol identifiers, Cargo dependency references, GitHub CLI text, range text, range commit messages, diverged related bases, unrelated/missing bases, old-history exclusion, added-then-deleted text, and merge-resolution additions. The checker also covers malformed UTF-8 paths/text and normal Unicode/control-character paths; the malformed-path fixture constructs Git tree objects without checking them out, so it also runs on filesystems that reject those bytes.
 
 ## Supersedes
 

scripts/tests/test_prohibited_names.py

@@ -27,15 +27,35 @@ class ProhibitedNamesTests(unittest.TestCase):
             env={**os.environ, "PYTHONUTF8": "1"},
         )
 
-    @unittest.skipIf(os.name == "nt", "Windows filenames cannot contain raw invalid UTF-8 bytes")
     def test_invalid_utf8_path_cannot_turn_into_an_empty_diff(self) -> None:
         with tempfile.TemporaryDirectory() as directory:
             repository = Path(directory)
             run_git(repository, "init", "-q")
-            with open(os.fsencode(repository) + b"/bad-\xff.rs", "wb") as output:
-                output.write(b"Ghostty\n")
-            run_git(repository, "add", "--all")
-            result = self.check_staged(repository)
+            run_git(repository, "config", "user.name", "CI test")
+            run_git(repository, "config", "user.email", "ci@example.invalid")
+            # Git trees can contain these bytes even when the host filesystem
+            # cannot. Never require a checkout or create an invalid OS filename.
+            def write_object(kind, contents):
+                return subprocess.check_output(
+                    ["git", "hash-object", "-w", "-t", kind, "--stdin"],
+                    input=contents, cwd=repository,
+                ).decode("ascii").strip()
+
+            empty_tree = write_object("tree", b"")
+            base = subprocess.check_output(
+                ["git", "commit-tree", empty_tree], input=b"baseline\n", cwd=repository,
+            ).decode("ascii").strip()
+            blob = write_object("blob", b"Ghostty\n")
+            tree = write_object("tree", b"100644 bad-\xff.rs\0" + bytes.fromhex(blob))
+            head = subprocess.check_output(
+                ["git", "commit-tree", tree, "-p", base],
+                input=b"invalid path fixture\n", cwd=repository,
+            ).decode("ascii").strip()
+            result = subprocess.run(
+                [sys.executable, str(CHECKER), "range", "--base", base, "--head", head],
+                cwd=repository, text=True, encoding="utf-8", capture_output=True,
+                env={**os.environ, "PYTHONUTF8": "1"},
+            )
             self.assertNotEqual(result.returncode, 0)
             self.assertIn("utf-8", result.stderr.lower())