Re 183 samples versus 734 for SCWF. Makes sense that the bigger number of samples is more accurate for pitch, because when the pitching algorithm wants to pick a new note, it will likely turn out to be wanting a repetition rate that requires a fraction of a sample (say 777.6459 for C#). Rounding that to the actual sample length possible (whole numbers, obviously) introduces an error that is at most about 1.2 cents, wheres, doing the math with 183 samples, the biggest error is about 4.6 cents. Those deviations will of course vary depending on how exactly the M:S pitching algorithm works.
183 for middle C is already about 4.3 cents out, but 734 is out by 0.3 cents.
734 may be optimal ? Not sure if its worth going longer.
EDIT: this is also covered here (I discovered, after doing the math myself)