I feel like I'm taking crazy pills here.
Has nobody else noticed these things work with multiples of 7? It occurred to me while staring at a page of voltage sets I've written down how close many of my numbers were to divisibility by 7, so I went full DaVinci code and wrote a list of them out in ranges near my voltages.
7 x 256 = 1792, which is my exact stock VCCIN voltage in millivolts.
7 x 162 = 1134, which is my exact stock 39x core VID
7 x 172 = 1204, exact stock 39x cache VID
7 x 144 = 1008, stock IOD
Scaling VCCIN in 56mv chunks popped out a few I've previously seen islands of stability around like 2.044v and 1.736v. I randomly pulled 1.680v off the list and it turns out my CPU plays just fine with the correct low VCCIN.
Doing some low multi testing, I noticed that cstkl1 was spot on. My CPU runs 41x core stably in prime95 blend at its 39x VID, and testing in 14mv steps proved 42x stable 42mv later. Following the pattern matches previous experiments' required voltages with 43x 1218mv, 44x at 1260mv, and 45x at 1302mv. 46x would be 1344mv if linear scaling holds true, but considering it tends to break down around the 1.3v mark and mine scales poorly from 1.05v on, I'm probably right on the mark rooting around at 1.35-1.36v.