numpy apply_along_axis drops values after casting incorrectly

Hello,
I have found that when I make the following call using apply_along_axis and return an int rather than a float, the field with the incorrect type is cast to the right type and put into the array properly, but the value of the subsequent field is set to zero:

def f0(m,l,val):
    return np.apply_along_axis(f1,0,m,m,l,val)

def f1(v,m,l,val):
    return np.apply_along_axis(f2,0,m,v,l,val)

def f2(v1,v2,l,val):
    l[0] += 1
    if l[0] <= 2:
        return 1.
    if l[0] == 3:
        return val
    if l[0] == 4:
        return .34151

print f0(np.array(np.ones((1,2))),list([0]),0) <-- int
print f0(np.array(np.ones((1,2))),list([0]),0.) <-- float
print f0(np.array(np.ones((1,2))),list([0]),1) <-- int
print f0(np.array(np.ones((1,2))),list([0]),1.) <-- float
print type(f0(np.array(np.ones((1,2))),list([0]),1)[0][1])
print numpy.__version__

output:

[[ 1.  0.] <-- correctly zero
 [ 1.  0.]] <-- value dropped
[[ 1.       0.     ] <-- correctly zero
 [ 1.       0.34151]] <-- value correct
[[ 1.  1.] <-- correctly 1
 [ 1.  0.]] <-- value dropped
[[ 1.       1.     ] <-- correctly 1
 [ 1.       0.34151]] <-- value correct
<type 'numpy.float64'> <-- type is float64, this was correctly cast
1.8.2 <-- numpy version

discovered with @asna1005

Start by running the supplied f0/f1/f2 reproduction with NumPy 1.8.2 and compare the integer and floating-point cases. Trace the nested apply_along_axis calls to identify where the subsequent value is lost; done means the reproduced cases retain the expected value while preserving the demonstrated casting behavior.