On taking a second look at this, I think the
apparent failure of the the lowpass filter to be as effective as expected is likely due to the relatively small "block" size being used to calculate the spectrum using Fourier analysis.
Without getting too deeply into it, in order to calculate the amplitude at a given frequency you have to sample a "block" of time that encompasses at least one full cycle at that frequency, and the more you have, the more precise the result is. At these very low frequencies, a typical block size of 2048 samples won't even include a full cycle of 20Hz sampled at 44.1kHz (1/20 * 44,100 = 2205).
Voxengo Span uses a default block size of 2048 samples, and the spectrum roll-off with the LP filter enabled looks very similar to your screenshot. But if I crank the sample size up to 8192 or 16384 in the Spectrum Mode Editor, the filter starts to look a lot more effective with a much steeper roll-off of the spectrum "shadow". The downside is that the spectrum can't be updated as frequently with a larger block size so the display isn't as smooth.
So, basically, this is another case where you need to use your ears instead of your eyes. The actual audible result will be much better than what's being displayed.