A problem with compressors/limiters is that, even with an “zero” attack time, nasty-sounding transients and overshoot can occur when there is a signal with a fast rise time. By delaying the signal path the compressor attack precedes the the signal e.g. you don't miss any of the peaks.
Now to the esoteric side ... if you know what is going to happen before it happens couldn't you do a better job of handling it? To me, in my limited testing, the vocal sounds better, more natural with the side chain delayed, the level is better controlled without sounding squashed and being pushed back into the mix.