eps(63396.105f0) = 0.00390625f0 > 0.0016288757f0
You are worrying about an imaginary component that is smaller than the floating-point resolution of your real component. Your result is in fact real-valued (within the resolution of the number type used).