array[:] = value silently coerce value to array type even when unsafe

[gdementen]

Unsure if this should be a warning or an error, but doing this silently is clearly not acceptable to me for unsafe conversions.

class CS(CheckedSession):
    v: CheckedArray(a, dtype=int)

>>> arr = ndtest(a) * 1.6
>>> arr
a   a0   a1
   0.0  1.6
>>> cs = CS(v=arr)
>>> cs.v
a  a0  a1
    0   1

Surprisingly Numpy has the issue too (I suppose we inherit it from there):

>>> import numpy as np
>>> np.__version__
'2.3.5'
>>> target = np.zeros(2, dtype=np.int64)
>>> source = np.array([0., 1.6])
>>> source.dtype
dtype('float64')
>>> np.can_cast(source.dtype, target.dtype)
False
>>> target[:] = source
>>> target
[0 1]

This has been a known issue in Numpy for a looong time but even though the developers seem to agree it is unfortunate, it will be a while until that behavior can be changed (if ever) in Numpy. See numpy/numpy#7730 and numpy/numpy#8733

[gdementen]

Silent string to int conversion happens too, which makes me really sad 😢 :

>>> source = np.array(["1", "2"])
>>> target[:] = source
>>> target
[1 2]

[gdementen]

I just realized this problem is not specific to CheckedSession:

>>> target = zeros(a, dtype=int)
>>> source = ndtest(a) * 1.6
>>> source
a   a0   a1
   0.0  1.6
>>> target[:] = source
>>> target
a  a0  a1
    0   1
>>> source = Array(["1", "2"], a)
>>> source
a  a0  a1
    1   2
>>> target[:] = source
>>> target
a  a0  a1
    1   2

The only thing specific to CheckedSession is that

>>> cs.v = arr
>>> cs.v
a  a0  a1
    0   1

because it actually does cs.v[:] = arr behind the scene

[gdementen]

Even though the problem lies with numpy, I think that LArray as the more-user-friendly-less-performance project should tackle this issue directly without waiting for numpy to fix it