Rare float overflow bug in complex_numbers strategy

[Zac-HD]

Hypothesis 3.86.0 made floats(min_value=inf, allow_infinity=False) an error, because the arguments are contradictory and thus impossible to satisfy.

Many test runs later, it turns out that this condition can sometimes be triggered by the following test:

hypothesis/hypothesis-python/tests/cover/test_complex_numbers.py

Lines 96 to 98 in 3c23c07

	@given(st.data(), st.floats(0, allow_infinity=False, allow_nan=False))
	def test_minmax_magnitude_equal(data, mag):
	val = data.draw(st.complex_numbers(min_magnitude=mag, max_magnitude=mag))

Step one is obviously to write a test which can quickly trigger this bug (e.g. by only passing very large magnitude bounds), but by inspection I'm pretty sure it's somewhere in the following lines:

hypothesis/hypothesis-python/src/hypothesis/_strategies.py

Lines 2011 to 2021 in 3c23c07

	if max_magnitude is None:
	zi = draw(floats(**allow_kw))
	rmax = float("inf")
	else:
	zi = draw(floats(-max_magnitude, max_magnitude, **allow_kw))
	rmax = cathetus(max_magnitude, zi)
	# Draw the real part from the allowed range given the imaginary part
	if min_magnitude is None or math.fabs(zi) >= min_magnitude:
	zr = draw(floats(-rmax, rmax, **allow_kw))
	else:
	zr = draw(floats(cathetus(min_magnitude, zi), rmax, **allow_kw))

[Zac-HD]

Three times today 😭