I turn on Automation Write, hit play, and play with a parameter.
 
As it's recording, it sounds the way I want it to and seems to be writing the nodes correctly: 
 
Then as soon as I hit stop, the nodes automatically quantize:

 
The result played back sounds awkward, especially for fast, rhythmic effects. Is there any way to prevent Sonar from deleting all my beautiful, nuanced nodes?

And you don't have Input Quantize (or anything similar, if such exists in X-series) on?

You can control the precision of this quantization. I know there's a setting somewhere but I can't remember where. Preferences somewhere... Scook, do you know where it is?

There is a variable in AUD.ini which automatically thins recorded automation called AutomationDecimationMSec. The default is 50 milliseconds.

I tried changing those values from 2 to 1000 ms in X2 (what's on one of my laptops), and didn't see any difference in the on-screen graphic resolution. I tried saving the config file and closing/opening Sonar to adopt the changes. I'll try later with X3.

It has nothing to do with "on-screen graphic resolution". AutomationDecimationMSec removes nodes from recorded (not hand drawn) automation. At least that is my understanding. I don't believe it is documented anywhere.

I was indeed doing fader automation, but whether I moved the fader with decimation at 2ms or 1000ms, I didn't see any difference in what was drawn onscreen. Perhaps the resolution changes internally and the graphic always gives a rough representation? I'll try on X3 and see if there's an audible difference between the two extreme settings, regardless of what shows up on screen.
 
I couldn't find any documentation anywhere either. 

I don't know what the allowable range is for the variable. It may not even work in the X series. FWIW, I just did a quick test with values 20, 50 and 100. They looked the same to me. I don't have a pre-X series version installed to test.

There was a similar thread a few years ago http://forum.cakewalk.com/FindPost/2071724 It gets into discussing the AutomationDecimationMSec midway through the second page. I mostly hand draw automation so it is not an issue.

Okay...I did some tests by generating white noise and doing automation with 2ms and 1000ms decimation. In both cases, the automation graphic indicated a minimum time of around 40ms between nodes. I used Sonar's onscreen faders for automation, moving them as fast as I possibly could - so fast it would definitely be unmusical unless you wanted a really fast random tremolo effect.
 
The following screenshot shows what happens when you render the tracks with the volume automation. It's clear the volume curves follow the automation graphic curves. Give how fast the changes are (the timeline is set to milliseconds), it would certainly seem that Sonar's automation resolution is about the same as what human perception is capable of discriminating.
 
As to whether the automation decimation parameter serves any useful purpose, I have no idea but it doesn't seem to make any difference.