Inconsistent behavior of dtype.descr with metadata

We used to the following property: if x is a dtype then numpy.dtype(x.descr) is the same dtype. This property is broken in NumPy 1.8 when one specifies metadata to dtype. Is it possible to at least fix that descr returned by dtype can always be evaluated back to a dtype?

Related issues: https://github.com/h5py/h5py/issues/1307

Reproducing code example:

import numpy as np

# 1. metadata is lost for simple dtype
s_dt = np.dtype('S8', metadata={'msg': 'Hello'})
print(s_dt.descr)
# prints [('', '|S8')]

# 2. metadata is returned for structured dtype, but it cannot be evaluated into dtype
struct_dtype = np.dtype([('a', s_dt)])
print(struct_dtype.descr)
# prints [('a', ('|S8', {'msg': 'Hello'}))]

# next line raises a ValueError exception
# ValueError: invalid shape in fixed-type tuple.
np.dtype(struct_dtype.descr)

Numpy/Python version information:

1.18.1 3.8.1 | packaged by conda-forge | (default, Jan 29 2020, 14:55:04) [GCC 7.3.0]

[mattip]

I am not sure what descr is actually good for

Maybe deprecate/make it private as part of the dtype refactor?

[seberg]

I am not sure what descr is actually good for 😃. We postprocess in a few places, but other than that? Trying to cram this feature in descr is just a bad idea. And it may well be that nobody really should be using .descr at all?