dscoyne
I would not be doing this in real time.  I would be correcting the vocal after the track has been recorded.
 
So if I understand correctly, if I raise the buffer to 1024 when doing this, it should not have an effect on latency.
 
Should I leave it there, or lower it back down when I am done pitch correcting with Melodyne?  For example, if I then process the track with the V-64 Vocal Strip and/or double it, or apply other effects.  And what about further recording?
 
Thanks.

Increasing buffer size will increase latency. But if you do it as part of mixing/editing (as opposed to tracking) it won't matter. The ASIO panel is your friend. You don't have to decide on a buffer setting at the beginning and keep it unchanged throughout the project. Lower it when recording soft synths or playing guitar through a simulator, raise it (if necessary) when pitch-correcting later.
 
Or, see how far you can get with Melodyne while ignoring their warning. That can work too. Most of all, have fun.

Latency only affects real-time input from the user. This is an important point that is often overlooked. It will not mess with existing tracks, offline edits or really anything except what you're playing or manipulating in real-time. Even then, a soft synth you're playing will still record the notes at the time you hit the keys rather than the moment you hear the sound.

So you can comfortably up the buffer size/latency for editing even if you plan to track more later at lower settings. Or, yes, you can ignore the warning. I've always done that.

Sanderxpander
I've never bothered to up the buffer that much for Melodyne. Truth be told, if you leave a lot of clips "active" (rather than bouncing them down) Sonar seems to eventually get less stable.

I've also experienced increasing instability if I'm "careless" with Melodyne. Leaving an active instance in the Project when closing is a risk to me. I only edit a few bars at a time and bounce them before proceeding.

Sanderxpander
Melodyne's algorithms can shift timing by a few samples as part of its efforts to keep a neutral sound during pitch and timing changes. This is a known "issue". So that could conceivably affect phase coherence although I have to wonder in what kind of situation you'd hear that on vocals. Did you multitrack a backing vocal group with a lot of crosstalk?

This effect, by the way, is separate from ARA integration so using Melodyne externally and then importanting tuned WAVs should suffer from the same thing.

Tracking multiple vocals from the same vocalist through a modular multitrack deck, then importing to X2 for mixing. There is some headphone leakage on a few of the tracks which I know is a source of trouble, but I tend to only push tuning to short of the audible artifact point. I fix timing issues with editing. When I disable the Melodyne for the fx bins, phase returns to normal coherence.