The example below works differently with numpy:

import numpy as np, dpctl, dpctl.tensor as dpt

np.__version__
# Out: '2.2.4'

dpctl.__version__
# Out: '0.20.0dev0+71.gc5d18f4b6f'

a = np.ones(5, dtype='?')
ia = dpt.ones(5, dtype='?')

np.ceil(a).dtype
# Out: dtype('bool')

np.trunc(a).dtype
# Out: dtype('bool')

np.floor(a).dtype
# Out: dtype('bool')

dpt.ceil(ia).dtype
# Out: dtype('int8')

dpt.trunc(ia).dtype
# Out: dtype('int8')

dpt.floor(ia).dtype
# Out: dtype('int8')

In overall, it seems Python array API standard requires dtype of output array to be matched with dtype of input array:

out (array) – an array containing the rounded result for each element in x. The returned array must have the same data type as x.

Although the support of bool dtypes is not mandated:

x (array) – input array. Should have a real-valued data type.

It sounds like if dpctl provides the support of bool dtype, it has to return out array of bool dtype also.