There are a number of things which are incorrect in your reply, Shawn.
First of all, it always uses the initial transient as a starting point for the shooting interval (the analysis was set up to use shooting because that's the default), regardless of whether you specify tstab. With an oscillator it uses a small number of cycles of the estimated frequency (4, from memory) for this initial transient which is usually not enough for the oscillation to have built up significantly.
Note that for the Shooting Newton method the initial transient does not need to be fully settled - the algorithm can find the settled steady state itself. However, generally you need to be close to the settled response - ideally past any non-linear startup behaviour.
Setting maxstep is not necessary for accuracy - because maxstep is controlled during the shooting interval to be (depending on errpreset and other settings) 5 points per period of maxacfreq, which is set to 4 times the highest harmonic requested, or 40 if fewer than 10 harmonics are requested (so at least 200 points per period, more if higher harmonics are requested). However, this does not apply to the tstab part of the simulation - there is often benefit in setting maxstep to ensure that the oscillator can build up oscillation more quickly - if there are too few points, the oscillation simulation may take a longer time to start. The lack of a driving signal in the circuit tends to result in increased numerical error and limiting the maxstep helps to mitigate that (without going into too much detail as to why).
If the circuit is a more linear oscillator (or high Q) such as a LC tank or crystal oscillator, you may want to try setting the oscic=lin option for the simulation. This can help speed up the starting of the oscillator.
So overall the advice to use maxstep was a good idea, but for the wrong reasons...