diff --git a/tests/test_lite.py b/tests/test_lite.py
index daea2a2..6bfb969 100644
--- a/tests/test_lite.py
+++ b/tests/test_lite.py
@@ -333,6 +333,9 @@ def test_the_foreground_is_never_asked_about_in_lite(self):
         self.assertIsNone(pill.paste_target)
         asked.assert_not_called()
 
+    @unittest.skipUnless(sys.platform == "win32",
+                         "full mode resolves the foreground window through inject.py, "
+                         "which binds user32 at import")
     def test_full_mode_still_tracks_it(self):
         import flow.ui as ui
 
diff --git a/tests/test_main.py b/tests/test_main.py
index c9fb850..7810b31 100644
--- a/tests/test_main.py
+++ b/tests/test_main.py
@@ -371,6 +371,7 @@ def launch(self, argv) -> int:
         import flow.asr
         import flow.diag
         import flow.profile
+        import flow.refine
         import flow.ui
 
         import flow.__main__ as mod
@@ -445,7 +446,21 @@ def launch(self, hotkeys, argv=(), platform="win32") -> tuple[int, str]:
         path = self.dir / "profile.json"
         path.write_text(json.dumps({"schema": 1, "hotkeys": hotkeys}), encoding="utf-8")
         out = io.StringIO()
+        # `refine.resolve` is stubbed, and not only for speed. It calls
+        # `shutil.which`, and `shutil.which` reads `sys.platform` - which this
+        # helper has just lied about. On a real macOS runner that sends the stdlib
+        # down its Windows branch, where `_winapi` is None: `AttributeError:
+        # 'NoneType' object has no attribute 'NeedCurrentDirectoryForExePath'`,
+        # raised from a line no part of Flow wrote. Which agent CLI happens to be
+        # installed has nothing to do with what this class asserts.
+        #
+        # Stubbed at `resolve` rather than at its callers, which was the first
+        # attempt and was whack-a-mole: `main` reaches the same line through
+        # `available()` *and* `unverified()`, and patching one left the other. It is
+        # the single choke point - the only `shutil.which` in the module, and what
+        # every lookup goes through.
         with mock.patch.object(sys, "platform", platform), \
+                mock.patch.object(flow.refine, "resolve", return_value=None), \
                 mock.patch.object(flow.profile, "DEFAULT_PATH", path), \
                 mock.patch.object(flow.diag, "Diag"), \
                 mock.patch.object(mod, "Session"), \
diff --git a/tests/test_pill.py b/tests/test_pill.py
index e096a97..afdda73 100644
--- a/tests/test_pill.py
+++ b/tests/test_pill.py
@@ -111,6 +111,15 @@ def test_the_advance_is_the_one_the_real_font_measures(self):
             root.withdraw()
             ui._load_fonts()
             f = tkfont.Font(root=root, font=ui.FONT_TRACE)
+            # `_load_fonts` registers the bundled Plex weights through GDI and returns
+            # early off Windows, so anywhere else Tk quietly substitutes its own
+            # monospace and this measures *that* - 10 px on a macOS runner against the
+            # 7 the real face gives. Skipped on the substitution rather than on the
+            # platform, because the question is whether the font resolved, and a
+            # Windows machine missing the file deserves the same answer.
+            if f.actual("family") != ui.FONT_TRACE[0]:
+                self.skipTest(f"{ui.FONT_TRACE[0]} not installed; "
+                              f"Tk substituted {f.actual('family')!r}")
             self.assertEqual(f.measure("M"), ui.LABEL_ADV)
             # Monospaced, which is the assumption behind one advance for every glyph —
             # including the space in `NO INPUT`.