You may have some lock stability issues resulting from the gradients. I routinely run this experiment at short values for D8 without any problems.

Make sure that your lock phase, lock power and lock gain are correctly set, and optimize your lock loop parameters (loop gain, filter time-constant, etc.) using the "loopadj" command. I have seen these sort of lineshapes during diffusion measurements (DOSY), and they go away when the lock parameters are optimized.

I'd agree with Kirk, this lineshape is a almost certainly caused by a misset lockphase.

One issue is that the lock phase seems to drift quite a bit with repeated calls to loopadj--it seems to drift between 55-65 degrees.

The autophase algorithm is not very precise - +/- 5 degrees is not untypical for older systems (for newer AVIII systems with LTRX you can set the "enhanced lock level" mode in edlock, which is rather more precise). In extremis you can just adjust manually for best lineshape, i.e. run the experiment with just a couple of scans, for different values of the lock phase, and look for best result. But what you see finally looks approximately typical.

N.B. you can use the pulse program selnogpzs, to remove zero-quantum artefacts - you need the prosol entry for Z-spoil (adia) on channel F1 to be set in standard shaped pulses (pulse should be 20000, shape Crp60,20,20.10 and calc to set the power). This can tidy up the lineshape of the NOE peaks in the case of J-coupling (but does not remove artefacts from spin state mixing in the case of strong coupling).

The loop parameters (loop gain, loop time, etc), however, are highly converged and do not fluctuate between multiple calls.

The calculated loop numbers just depend on the lock gain after autogain, so in principle should come out the same each time.

Your lock signal looks normal, there is no significant evidence of saturation (you would see the lock recover after the gradient pulse to a value a bit higher than the equilibrium).