Second... If my interpretation of the rollout is correct, should I really believe Gnubg analysis if even 4 ply analysis is this far off from the rollout's?
Bots aren't perfect and the stronger the settings the more likely you are to get the correct result.
I would guess that if you use the supremo (2-ply) or grandmaster (3-ply) analysis level you will get the correct indication more than 90% of the time, so I wouldn't disregard the analysis just because it can be wrong on occassion. Instead I would do exactly what you did. When you don't trust the evaluation you see, then do a rollout. I would recommend 2-ply (supremo or world class) as settings for the rollout.
Can you always trust a rollout? Well no, but you can trust it a lot more than an analysis on pretty much any ply.
Here's a rollout from eXtremeGammon:
                       |
| ||||
| XGID=-aCGBaa----------a-bbbabbA:1:-1:-1:00:0:0:0:7:10 | |||||
| Blue on roll, cube action? | |||||
| Analyzed in Rollout | No redouble | Redouble/Take |
| Player Winning Chances: | 80.67% (G:0.00% B:0.00%) | 82.85% (G:0.00% B:0.00%) |
| Opponent Winning Chances: | 19.33% (G:6.99% B:0.23%) | 17.15% (G:7.28% B:0.81%) |
| Cubeless Equities | +0.536 | +1.225 |
| Cubeful Equities | ||
| No redouble: | +0.799 (-0.201) | ±0.004 (+0.795..+0.803) |
| Redouble/Take: | +1.172 (+0.172) | ±0.009 (+1.163..+1.182) |
| Redouble/Pass: | +1.000 | |
| Best Cube action: Redouble / Pass | ||
| Rollout details | ||
| 1296 Games rolled with Variance Reduction. Moves: 3-ply, cube decisions: XG Roller | ||
| Double Decision confidence: | 100.0% | |
| Take Decision confidence: | 100.0% | |
| Duration: 4 minutes 24 seconds | ||
eXtreme Gammon Version: 2.10, MET: Kazaross XG2
XG evaluates this as a drop on all plys but the lower you go the closer it gets to thinking it's a take.