I wouldn't call 63.7% the "expected result" (unless you got fooled by the progress window in the same way that I did, see the explanation that precedes the final diagram of this post). However, inspired by your 20k attempt, I did a rollout with the XGR++ settings until I finally hit the coveted 50.0%. I had been reluctant to try before because I overestimated how long it would take.
The same jumpy behavior you describe (though your hypothesis was probably better than mine as to the reason) occurred at the XGR++ settings level as at the 4-ply level, except (a) less so, (b) XGR++ is more resistant to coming to a 100% conclusion prematurely, and (c) the 50% figure in the bgonline-result window that you see here matched the 50% reported in the progress window (unlike what happened to me at the link above).
Below is my XGR++ rollout, which finally hit pay dirt.
Nack
                     |
| ||||
| XGID=--AC--A--------aaaadcca---:1:-1:1:61:0:0:3:0:10 | |||||
XGR++ moves and cube decisions, 362k trials | |||||
| 1. | Rollout1 | 6/Off 3/2 | eq: +1.486 | |||
| ||||||
| 2. | Rollout1 | 6/Off 2/1 | eq: +1.486 | |||
| ||||||
| 1 362880 Games rolled with Variance Reduction. Dice Seed: 88888888 Moves and cube decisions: XG Roller++ | ||||||
eXtreme Gammon Version: 2.10