Jim,
This is part philosophical, and part objective.
I have no particular bias against mechanically starting an idler, but I far prefer electrically starting an idler as it is exquisitely easy to do for idlers equal to or less than 5 HP (Steveco 90-66, only), and it is only slightly more complicated for much larger idlers, larger or very much larger than 5 HP (Steveco 90-66 for sensing and control, plus a definite purpose contactor).
A polyphase motor, an idler in this case, will not start without first achieving a significant imbalance in its magnetic circuit, after which the rotor will continue to rotate without further intervention. It is this imbalance which effects the starting.
A rope or pony motor starting means, both of which are a mechanical starting means, can achieve the required imbalance.
Now, an imbalanced idler of the Fitch Williams type, which is usually imbalanced to about 60/40, still requires a "significant imbalance" to start, except that it is already significantly imbalanced by the balancing caps, so it takes much more effort to further significantly imbalance it in order to effect a good, clean start.
My "rule of thumb" is somewhat imprecisely, yet very practically stated as "1, 2, 3 and 30".
Which means, that for each quarter HP of idler capacity:
1) 1 µF for power factor correction, Cac,
2) 2 µF for Ccb phase-shifting,
3) 3 µF for Cab phase-shifting, and
4) 30 µF for Cs momentary phase-shifting and starting (in parallel with Cab).
There can be no ambiguity if 30 + 3 = 33 µF is placed between A and B, whereas 2 µF is placed between C and B.
33 µF >> 2 µF, therefore the idler starts quickly and positively, usually in 1-2 seconds, after which the 30 µF is removed and the idler thereafter operates with a 3 µF/2 µF ratio, which is the same as 60/40 percent.
About the opposite case, where the 30 µF is placed across the 2 µF capacitor, 30 + 2 = 32 µf and 32 µF is >> 3 µF, so the idler will still start quickly and positively, but it is rotating in the "wrong" direction.
I offer without proof that it is better to rotate as in the case first-mentioned, above.
Objective tests have proved this to be true.
Yet, both achieve the required "significant imbalance" for starting, whereas only one of these is that configuration which aids dynamic performance in actual operation, that is, after starting has been effected.
Peter