I agree with the original poster--nothing ruins the illusion like a jerky start. It's a good, valid question and there are a LOT of variables...
A permag 12V motor having a high stall current, and without some kind of closed-loop feedback speed control will tend to give a jerky start; it may also stall after starting once the slack of the trailing consist has been stretched, because the increased current draw will tax the power supply and reduce the total voltage in the circuit. Substituting a motor with 24V winding will start more smoothly and be more resistant to stalling. But it will make fewer RPMs at the NMRA 12V reference point. Using a 24-volt control system designed for G-gauge will restore the original top speed and give finer control at all speeds. The use of a cut-wave or "pulse" power supply may also yield improved operation at low speeds. The tradeoff is increased motor heating and/or electrical noise.
The gear ratio is important- nearly all of the locos with the so-called "China drive" are capable of more than 100 scale mph, which compromises their ability to maintain slow speeds without stalling. In the case of steam locos with direct drive to a worm wheel on the driving axle, the worm wheel should be of the largest possible diameter to provide more pull-in torque for the worm.
Any loco with self-locking worm gears may stutter at very slow speeds if there's insufficient torque to overcome the internal friction. Depending on the gearbox characteristics, it may also "buck" or surge going downhill. Furthermore, a loco with two independent motors and two locking worm gears won't start until BOTH motors are making enough torque to overcome the static friction, and then the loco will tend to start abruptly. The performance of *ANY* China-drive loco can be improved simply by removing one motor (whichever one has the "lazier" start characteristic when tested independently.) Another technique is to wire both motors in series. This is somewhat equivalent to substituting a 24-volt motor as described above.
Gears should be inspected carefully--burrs, malformed or chipped teeth, bent wheels or axles, misaligned bearings, dirty wheels or poor electrical pickup, pinch points in the gauge, kinks and dirty track will all contribute to stalling or jerking at slow speeds.
One more thing to think about--prototype locos are free-rolling. Three men can push a loco on level track, and without brakes it would go downhill like a roller coaster! Unfortunately physics don't scale down. The ratio of total static friction (in the motor, gearbox, and trailing car axles) to rolling friction once underway is much higher in a model compared to the prototype. Think about a real freight train starting on a cold day (especially before the widespread use of roller bearing trucks on freight cars.) There's a lot of static friction in the journals, but even after that static friction has been overcome, the train won't suddenly accelerate due to its enormous mass. Even with heavily weighted cars, there's a lot less mass to accelerate in a model (lower inertia), and this reduced mass also translates to relatively insignificant momentum compared to the prototype.
Closed loop speed control (which use an optical tach or back-EMF) helps to maintain a steady speed on curves and grades, and prevent stalling as the slack is being taken out. But in my experience these electronic systems can't duplicate the delicious hysteresis, the precious "moments" when a prototype loco strains from a dead stop to one... two... three mph. Some experienced loco builders will disagree, but I'm a fan of heavy flywheels, non-self locking gears, and generally a lot of mass. One of these days I would like to fabricate a Hobbytown-style clutch drive scaled up for O-gauge duty. Taken together, all of these things help preserve the illusion, and make our trains more forgiving and fun to operate! -Ted