exclude-min/exclude-max not reliable when drawing arrays of floats

[rsokl]

hypothesis-4.14.0

I'm not sure if this is necessarily a bug or something that should be documented in the arrays strategy.

import numpy as np
import pytest
import hypothesis.strategies as st
import hypothesis.extra.numpy as hnp


@pytest.mark.parametrize("dtype", [np.float16, np.float32, np.float64])
@pytest.mark.parametrize("max_", [-1., 0., 1.])
@given(st.data())
def test_bad_exclude_min_in_array(dtype, max_, data):
    x = data.draw(
        hnp.arrays(
            dtype=dtype,
            shape=(1,),
            elements=st.floats(max_ - 1., max_, exclude_max=True)),
        label="x"
    )
    assert np.all(x < max_)

@pytest.mark.parametrize("dtype", [np.float16, np.float32, np.float64])
@pytest.mark.parametrize("min_", [-1., 0., 1.])
@given(st.data())
def test_bad_float_exclude_max_in_array(dtype, min_, data):
    x = data.draw(
        hnp.arrays(
            dtype=dtype,
            shape=(1,),
            elements=st.floats(min_, min_ + 1, exclude_min=True)),
        label="x"
    )
    assert np.all(min_ < x)

(Sorry to leave a terse issue - I caught this at work and wanted to document it ASAP. Gotta run!)

[Zac-HD]

The problem is that downcasting rounds to an excluded endpoint (fine so far...) which is silently added to the array when it should be emitting a warning about lossy conversion.