The expected results that are generated for the random test cases are being rounded before being converted into strings. Since the expected result is a string, it isn't possible to pass by returning a float. When you convert a float to a string it's rounded automatically for you. Passing this in python would require replicating the exact same rounding method at a bit level in order to produce exactly the same string in every case.

The kata could be altered to accept a float value as the result. But as it is written now it requires concatenating with 'is freezing temperature' first

Test description for python mentions rounding to 1 decimal place, but randomized tests require rounding to a changing number of decimal places.

'-28.333333333333332 is freezing temperature' should equal
'-28.333333333333336 is freezing temperature'
'103.88888888888889 is above freezing temperature' should equal
'103.8888888888889 is above freezing temperature'

The tests appear to be generated by a typed library using a double. Rounding in this manner is incompatible with untyped languages. You should change the kata to round to 1 decimal place per your description to provide logical compatability.