I believe you can take one of the low rank examples, such as Ringel's, and simply take the dual. Since Ringel's is on 9 points, with rank 3, the dual is on 9 points with rank 6.
@RicardoAndrade, I wrote some code to test this on a range of oriented matroids. My results are as follows:
ceva: good = 31; not basis = 4; bad = 0
πάππος: good = 75; not basis = 9; bad = 0
ringel: good = 84; not basis = 0; bad = 0
tsukamoto13[+1]: good = 267; not basis = 19; bad = 0
tsukamoto13: good = 266; not basis = 20; bad = 0
tsukamoto13[-1]: good = 267; not basis = 19; bad = 0
circularsaw3: good = 35; not basis = 0; bad = 0
Ω14[+1]: good = 334; not basis = 30; bad = 0
Ω14: good = 333; not basis = 31; bad = 0
Ω14[-1]: good = 334; not basis = 30; bad = 0
suvorov14: good = 331; not basis = 33; bad = 0
For each of the dual oriented matroids, I took each r element subset of the n elements. If they were independent forming your A and the remaining set, your B, was also independent, I counted this as good; if the r element subset were dependent I added to the not basis count. The bad count, always 0, is when the A set was a basis, and the remainder was not independent.
So I conclude that while your observation that while every r element subset for a uniform oriented matroid will satisfy your conditions, the restriction to uniformity is pretty strong and unnecessary.
@Ricardo I am sorry to keep posting up here, but I want an image this time:
This one is all bad ... i.e.
wheel12: good = 0; not basis = 165; bad = 55
I think my main observation is that my interest is pseudoline stretching, i.e. rank 3 oriented matroids, typically 'interesting' ones. This wheel is not very interesting for me. Whenever a lot of pseudolines meet at a single point, then the dual of the rank 3 oriented matroid will tend to have small dependent sets, and so not have the property you are interested in. But the oriented matroids of interest to me, are rank 3, and typically at each vertex have only two or three lines, maybe four meeting, so that every convector has at least n-4 elements, and all sets of size 3 are independent (in the dual), so that your property is trivial for any basis of the dual. In summary I think the fact that the bad column is always 0 is just an artifact of my interest versus yours.