Hmmm, there's still a bad corner case as follows:

>>> import numpy as np
>>> a = np.arange(6, dtype='f4')
>>> t = a % 2 == 1
>>> t
array([False,  True, False,  True, False,  True], dtype=bool)
>>> t*a
array([ 0.,  1.,  0.,  3.,  0.,  5.], dtype=float32)
>>> (1-t)*a
array([ 0.,  0.,  2.,  0.,  4.,  0.])

Since 64-bit Linux has the integer as 64-bits, and t is a boolean kind, (1-t) becomes a 64-bit integer array. Since it's an array, now it doesn't trigger scalar_array rules, so (1-t)_a becomes float64 instead of float32.

I think the categories for this should be reduced further from the three I did into just two: boolean/uint/int and float/complex, to fix this.

Integer is still 32 bits on 64 bit Linux, it's the long that is 64. That doesn't affect the problem though ;) Numpy 1.5.x also behaves this way, so it is backward compatible. Something like (1 - t) should really be done as ~t to preserve the boolean type.

I didn't do a change for the (1-t)*a corner case, since getting it to behave right in general requires some additional trickiness. I did put in a commented-out test for it, however.