There might be a problem
Set every 'roll' to 0 and fixed, make all (including that of the anchor image) 'pitch' optimisable, push the button and PT will find the right pitch.
Better, isn't it?
Now freeze 'pitch' and optimize for 'roll'. Then freeze 'pitch' and 'roll' for some anchor images and optimize the others, including 'view' and 'barrel' (EXIF's fov is often not accurate).
One possible cause for the "snake" phenomenon: the incorrect (too high) EXIF fov, whose total angle is more then 360°, cannot fit in one (full) turn unless it is wound.
The max error is now at 1 (down from 5).
(original: 14000x3000)
Colour couldn't be adjusted. 
First try is not always successful......
Second try neither...
Then, applying the recipe given above...
Brightness correction gets the things homogeneously worse. After manual colour processing...
It seems there was a lot of parallax error...
The final result is unexpectedly good
(reduction 20% ; original : 4700x3900; PTstitch+enblend)
(10000x6000)
650 CPs found by Autopano, then adjusted manually: remove points in corners or on the foreground, finetune a few of them. With autopano 1.03b, not a single pair was really wrong; however, I had to add points in areas strangely devoid of points. The max error felle from 4.5 to 2.5.
colour correction (very well) done by PTstitch
360°
resulting error>5; don't know why. I suspect slightly different lens focus lengths. Or too much overlapping?
There is also a lot of parallax error; one way to manage it is to free more parameters like one fov or lens parameter for each image.
