So I decided to finally stop being lazy and do some actual hard number data on the correlation between agility and iFrame counts. I will probably put this into a video at some point, but for now here is a bit of data. FYI I did this on PC but knocked my FPS down to 30 so that the frame count would cover all platforms. I also double checked at 60 fps to verify there weren't any strange rounding errors, and it was identical.
AGI - iFrames
85 - 8
90 - 9
95 - 10
100 - 12
105 - 13
110 - 13
115 - 15
120 - 16
For reference, here are the numbers I got from Dks 1 awhile back.
Slow Roll - 9
Medium Roll - 11
Fast Roll - 13
DWGR - 15
So in a nutshell, 120 agi is superior to DWGR in regards to iFrames, though the flip still had faster recovery for unadulterated spammage. 105 agi is equivalent to the fast roll and unless you're willing to go to 115 for the extra 2 iFrames, it's not worth it. I didn't test at 1 increment steps, but I'm quite sure somewhere between 110 and 115 would give you 14 iFrames, but again it's a steep cost for little gain. 85 agi is actually 1 iFrame less than fat roll, while 100 agi is 1 iFrame more than medium roll.
As you can see, the scaling is not linear. I went back and verified the 100 agi number more than once, and it is correct. Either the scaling is purposely flattened in that area to provide a good break point, or it's some sort of bug.
How did I test this? The same way I did in Dks 1, using a long duration AOE attack so that I could easily see at what point in the roll I became vulnerable. In Dks 1 I used the 4 Kings AOE because the hit box was longer than even the DWGR, ;IE impossible to roll through. In Dks 2 I used Licia's WoG which again is longer than even the highest iFrame count possible. Even on an absolute perfectly timed roll, meaning the first frame of my roll coincided with the first frame of her WoG becoming active, it still hit me at the end. They may have nerfed player WoG's, but Licia's is running at full tilt. What's strange is the light from the WoG ended long before my iFrames ran out, but there was this massive lingering phantom hit box afterwards. You could probably roll away from it if you were naked and not directly in front of her, otherwise you're toast.
Take from this what you will, I thought I'd finally get around to ending any speculation and just giving some hard data numbers. I'd like to get around to making an actual video explaining it all and showing how I validated these numbers, but it's going to take some time to do it right.
Edit:
Did a little more testing and have come to a few more conclusions.
Weight only affects roll distance. It has no affect on iFrames or roll duration. Whether you are at 0% burden, or 70%, your entire roll takes about 25 frames to complete.
Agility only affects iFrames. As long as you aren't fat rolling, you get as many iFrames at 70% burden as you do butt-naked.
16 iFrames out of 25 is actually very, very strong and only a couple frames off of the DWGR 15/22(at max burden limit). DWGR had 15/19 while naked, but that wasn't realistic.
120 agi gives you invincibility for 64% of the roll animation. DWGR at 50% burden was 68% of the roll animation.
Most people aren't going to go for the full 120, but even at only 12-13 iFrames you're basically invulnerable for about 50% of the roll animation(starting from frame 1).