This is due to the fact that Cosmic Ray Glitches can cause a change in the responsivity and one would like to move across as many detector elements as possible before that happens.
Each of the above rules can cause the realized S/N to surpass the requested S/N and cause the observation to take longer than naively anticipated by the observer. The reason for the first two rules is to insure the highest possible data quality: by taking three or four integrations (ramps) at a grating or FP setting, spurious results can be eliminated by polling. (Since FAST MODE is now required, this polling possible no longer exists). By sweeping the grating and FP back and forth three times, memory and directional effects can be identified and removed (these are known, at a low level, to exist). Finally a frame is filled with the maximum amount of data, this being preferable to nothing.
A Fast Scan is one where the first rule is dropped, with the intention of taking the fastest grating or FP scan as possible and doing one and only one integration (ramp) per step. Thus on a single sweep or the grating of FP, there will be no indication as to the reality (repeatability) of a measurement. Only by comparing the three or more results of sweeps of the grating of FP, will a polling be possible. This illuminates the danger of using fast scanning - of not having sufficient statistics to tell if a blip is real or not.
Note that the second and third rules are never dropped.