Moreover, older pulse sequences (it was a dark time, when PFG methods were like Scince-fiction) use phase cycling to choose proper coherence pathway transfer, for example COSY without PFG use phase cycling to select P or N type of coherence (p = -1 or 1). PFG introduces some issues, like longer pulse sequences (gradients must be inserted in echo, additional delays for gradient recovery. PFG can induce some currents in coils (especially when you use rectangular gradient pulses), and of course gradient selected experiments are less sensitive than phase cycle selected ones. (two times less sensitive)

In your case (p3919gp) it's a watergate suppresion method with composite 3-9-19 pulse train. Phase cycling in this sequence is made in manner 4*2. So you have here basic block of two scans (it's common in PFG selected experiments) which is repeated 4 times with phase shift of ?/2 pi/2 (90 degrees) for time averaging of spectrum and elimination of artifacts. 3-9-19 pulse train used in this sequence inverts all magnetisation vactors with exlusion of these on-resonance (ones that shows on offset on spectrum). It's just PFG spin echo with refocusing 180 pulse, but insted of selective excitation here you have selective exclusion of signal. First gradient is dephasing all signals, 180 (3-9-19) is inverting all-but-non-water-signal, second gradient is rephasing all signals and further dephasing signal of water.

So why this block is made of two scans, each of them shifted by ? pi (180 degrees) in phase ? PFG methods are not able to distinguish between sign of coherence pathway, so you need two scans if you want to obtain phase modulated spectrum (it's very important in 2D experiments)

So the right answer for your question is:

"Minimum number of scans in case of this sequence is 2" a few dummy scans also could be useful to equalize spin states before aqusition, for example 2 or 4.

Best regards, Arek