feat(loader): add load_zaya_model wrapper around mlx-lm.load

tests/test_loader.py

@@ -0,0 +1,25 @@
+import pytest
+from markovian_rsa_mlx.loader import resolve_model_path
+
+
+def test_resolve_local_path_passthrough(tmp_path):
+    # local dir is returned as-is
+    d = tmp_path / "weights"
+    d.mkdir()
+    assert resolve_model_path(str(d)) == str(d)
+
+
+def test_resolve_hf_id_format_unchanged():
+    # HF repo id is returned as-is (load_model handles download)
+    assert resolve_model_path("kyr0/zaya1-base-8b-MLX") == "kyr0/zaya1-base-8b-MLX"
+
+
+@pytest.mark.integration
+def test_load_real_zaya_q4_smoke():
+    """Hits the actual Q4 weights — slow, requires HF cache."""
+    from markovian_rsa_mlx.loader import load_zaya_model
+    model, tokenizer = load_zaya_model("kyr0/zaya1-base-8b-MLX")
+    assert model is not None
+    assert tokenizer is not None
+    assert hasattr(tokenizer, "encode")
+    assert hasattr(tokenizer, "decode")