Fix NumPy random rand size keyword

src/pyrecest/_backend/_shared_numpy/random.py

@@ -5,8 +5,23 @@
 _allow_complex_dtype = _common._allow_complex_dtype
 
 
+def _rand(*dims, size=None):
+    """Draw uniform samples while accepting the backend ``size=`` contract.
+
+    ``numpy.random.rand`` only accepts legacy positional dimensions, whereas the
+    PyRecEst random backend exposes ``rand(size=...)`` like the JAX and PyTorch
+    implementations.  Use ``numpy.random.random`` internally so keyword and
+    tuple sizes work without dropping support for NumPy's positional form.
+    """
+    if dims:
+        if size is not None:
+            raise TypeError("Specify either positional dimensions or size, not both.")
+        size = dims[0] if len(dims) == 1 else dims
+    return _np.random.random(size)
+
+
 rand = _modify_func_default_dtype(
-    copy=False, kw_only=True, target=_allow_complex_dtype(target=_np.random.rand)
+    copy=False, kw_only=True, target=_allow_complex_dtype(target=_rand)
 )
 
 

tests/backend/test_numpy_random_backend.py

@@ -0,0 +1,25 @@
+import numpy as np
+import pytest
+
+from pyrecest._backend.numpy import random
+
+
+def test_rand_accepts_backend_size_keyword():
+    random.seed(0)
+
+    samples = random.rand(size=(2, 3))
+
+    assert samples.shape == (2, 3)
+    assert samples.dtype == np.float64
+
+
+def test_rand_keeps_numpy_positional_dimensions():
+    random.seed(0)
+
+    assert random.rand(2, 3).shape == (2, 3)
+    assert random.rand(4).shape == (4,)
+
+
+def test_rand_rejects_ambiguous_positional_and_size_arguments():
+    with pytest.raises(TypeError, match="positional dimensions or size"):
+        random.rand(2, size=(3,))