Test that Tuple.reduce can't remove any indices

[asmeurer]

We can't guarantee that Tuple.reduce is fully canonical like with slices #82, because there are too many combinations. But we can test that it removes any redundant terms. Something like

diff --git a/ndindex/tests/test_tuple.py b/ndindex/tests/test_tuple.py
index c8a53a4..3a02961 100644
--- a/ndindex/tests/test_tuple.py
+++ b/ndindex/tests/test_tuple.py
@@ -10,7 +10,7 @@ from pytest import raises
 from ..ndindex import ndindex
 from ..tuple import Tuple
 from ..integer import Integer
-from .helpers import check_same, Tuples, prod, short_shapes, iterslice
+from .helpers import assert_equal, check_same, Tuples, prod, short_shapes, iterslice

 def test_tuple_constructor():
     # Test things in the Tuple constructor that are not tested by the other
@@ -131,6 +131,12 @@ def test_tuple_reduce_hypothesis(t, shape):
         if isinstance(reduced, Tuple):
             assert len(reduced.args) != 1
             assert reduced == () or reduced.args[-1] != ...
+            args = reduced.args
+            for i in range(len(args)):
+                a_index = a[index.raw]
+                idx = Tuple(*(args[0:i] + args[i+1:]))
+                # Is there a better way to do assert not equal?
+                raises((IndexError, AssertionError), lambda: assert_equal(a_index, a[idx.raw]))
         # TODO: Check the other properties from the Tuple.reduce docstring.

 def test_tuple_reduce_explicit():

The problem is that this leads to some odd cases which aren't straightforward to match, for example, (..., False) and False are equivalent for shape (0,) because the added 0 to the shape from the False is the same on the left or right. Similarly for (..., None) and None for shape (1,).

[asmeurer]

It's not just the ellipsis. There's the same problem for (slice(None), False) and (slice(None), None).

[asmeurer]

And for something like shape (2, 1), (..., None) and (slice(None), None) are equivalent, but not just None. So there's the question of which is better.