fix: make TestEmbedText parallel-safe by patching _get_embedding_model

Replace direct mutation of the global EMBEDDING_MODEL with
patch.object on _get_embedding_model, eliminating shared state
that caused flakiness under pytest-xdist parallel execution.

Fixes #454

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

tests/unit/test_memory_embeddings.py

@@ -99,16 +99,9 @@ class TestKeywordOverlap:
 
 
 class TestEmbedText:
-    def setup_method(self):
-        self._saved_model = emb.EMBEDDING_MODEL
-        emb.EMBEDDING_MODEL = None
-
-    def teardown_method(self):
-        emb.EMBEDDING_MODEL = self._saved_model
-
     def test_uses_fallback_when_model_disabled(self):
-        emb.EMBEDDING_MODEL = False
-        vec = embed_text("test")
+        with patch.object(emb, "_get_embedding_model", return_value=False):
+            vec = embed_text("test")
         assert len(vec) == 128  # hash fallback dimension
 
     def test_uses_model_when_available(self):
@@ -116,9 +109,9 @@ class TestEmbedText:
         mock_encoding.tolist.return_value = [0.1, 0.2, 0.3]
         mock_model = MagicMock()
         mock_model.encode.return_value = mock_encoding
-        emb.EMBEDDING_MODEL = mock_model
 
-        result = embed_text("test")
+        with patch.object(emb, "_get_embedding_model", return_value=mock_model):
+            result = embed_text("test")
         assert result == pytest.approx([0.1, 0.2, 0.3])
         mock_model.encode.assert_called_once_with("test")