Looking at overall usage of c_, s_ and r_ (which is a lot as I checked github search) I decided to keep them in the main namespace, therefore there's no need for local index_tricks namespace.

That sounds like the right decision to me.

The code changes look good to me and the changes to the lib namespace look correct:

In [5]: diff = set(oldall) - set(np.lib.__all__)

In [6]: diff.issubset(np.__all__)
Out[6]: True

In [7]: diff
Out[7]: 
{'c_',
 'diag_indices',
 'diag_indices_from',
 'fill_diagonal',
 'index_exp',
 'ix_',
 'mgrid',
 'ndenumerate',
 'ndindex',
 'ogrid',
 'r_',
 'ravel_multi_index',
 's_',
 'unravel_index'}

I think is good to merge as soon as the conflict is fixed.

Rebased and resolved conflicts!

Looking on github search, it seems there are downstream usages of np.lib.index_tricks.IndexExpression and np.lib.index_tricks.nd_grid. Both I think need to be available for type annotations and isinstance checks. I think that indicates that index_tricks needs to remain public with only those two symbols remaining inside it, everything else should be available in the main namespace. @rgommers do you agree?

Sorry for missing that on my previous review.

Looking on github search, it seems there are downstream usages of np.lib.index_tricks.IndexExpression and np.lib.index_tricks.nd_grid. Both I think need to be available for type annotations and isinstance checks. I think that indicates that index_tricks needs to remain public with only those two symbols remaining inside it, everything else should be available in the main namespace. @rgommers do you agree?

In my opinion both of them should be private...
[EDIT] I think they can stay for the purpose of type annotations, as CClass and others are available.

Ah, you mentioned type annotations - then sure, I can keep them.

Yeah, isinstance checks too.

I'm not sure about some of these. Using https://cs.github.com/, I find a single usage of index_tricks.CClass and 14 usages of index_tricks.IndexExpression. Each funky np.X_ object seems to have a separate class associated with it. nd_grid is the only one for which I get a lot of hits.

I'm also not sure I understand the type annotations I'm seeing, they seem wrong for s_. E.g:

>>> np.s_[:, :-1]
(slice(None, None, None), slice(None, -1, None))
>>> type(np.s_[:, :-1])
<class 'tuple'>

That's a regular tuple so annotating it as IndexExpression is incorrect.

nd_grid is the only one for which I get a lot of hits.

For completing the record: I rechecked those hits, and a large majority are outdated forks of the numpy repo or its docstring contents under See Also, which had a usage of np.lib.index_tricks.nd_grid. There wasn't anything that looked relevant.