Yes that's expected, Irrespectively its the range of 3, its a bit problematic overall. Anyway, looks to me your not actually wanting to use a lattice deformer at all. But your actually want to use a wrap deformer instead so do use that then.
See 3 points away get to affect the patch 3 happens to be sortof bad odd combination because its slightly unevenly discontinuous. For best results the ffd value should be a even number. because you WANT the ffd to look quite evenly both ways. A uneven number especially 3 which is the lowest possible number does get this behavior.
Since one moment the cloest point to interpolate form is on the other side then suddenly its nolonger closest and shifts so its now using different set of 3 points. With 4 the points usually space quite evenly around the point in question.
In any case it does seem like maya has a rounding priority upward, exact fractional values of 0.5 have to round somewhere maya just rounds them up (yes you did that wrong in school too). Might as well a statistical rounding would be even more funky. This may be consdered a bug but frankly if you can find a good solution for this then please tell me, ive gota few hunderd thousand statsticians who would love to know the middleground.
By the way this exact same effect happens with splines too, where the data has to shoot off inorder to meet the target.
But it's in number 3 where I find the perfect amount of control needed.
Explain to me again the again the logical conclusion here. 
Its the best solution because it doesnt work out at all?
so use a warp deformer.
