Datum A as a curved surface (no datum targets)

If you have datum targets, use an iterative alignment. If not, take a good spread of points over the whole surface (enough points to capture the true shape and form) and then do a 3D best-fit.

If you don't have datum targets on curved surface, it's up to programmer to choose targets

If there are no datum targets, then the entire surface should be used, as I said in my previous comment. You should take enough hits to accurately capture the shape and form of the surface. It's the same philosophy as if it were a plane, only taking three points would result in a perfect plane with no form deviation. To capture the flatness, you would need to spread your points over the whole plane and choose a sufficient point density to capture the expected amount of deformity.

Sorry but for me you are not right, because in real life curved surface will touch opposite part in 3 highest places, if it is datum A, it will never touch complete surface

And GD&T will take only touching points anyway from whole surface

if it is a curved surface BY DESIGN, then you have no idea what the 'high' (or 'low') points are unless you measure the 'entire' surface. It isn't like putting it on a flat plate and letting it 'level' to the 3 contact points like for a designed flat surface that came out with curves in it.

I would make iterative alignment with points choosen by you, then measure whole surface with many points or better scan it and then you will know which ones will be best for datums. If drawing is not complete or clear (what must be) it's up to programmer to choose best way and nobody can say nothing. Or ask designers to add this

FYI - Per ASME Y14.5.1 for Y14.5:2009 and later, you are not correct.

The feature is to be solved by least squares fitting AND then moved to the outside of the actual material envelope, so it might not touch in three places.

Also, you can not do Constrained Least Squares with three points (datum targets for a primary), so unless they are specifically ON the engineering document, your customer CAN complain about your inspection decision -IF- the drawing is invoking ASME Y14.5:2009 or later.

If it is from prior to 2009, you can do what you are suggesting and then get into an arguement with your customer about if you are allowed to decide what the design intent of their print was. You are not allowed to, you are supposed to ask the CDA for guidance, not decide you know best, but you can get in the arguement, and on one hand win, but your customer may put you on the list of suppliers they ween off their ASL.

Or they may, in some weird way, fall in love with you for it. Who knows.

But, post 2009, in ASME, your statement is factually incorrect. Doesn't mean I might not do what you are suggesting, but I would know where I was afoul of the standard if I made such a decision. Having had issues with point distribution in the past with a couple customers, I would likely ask their inspection what they were going to do, so I could do it the same way, right or wrong (wrong here not meaning I believe it could cause harm to humans while in use, I'm not trying to crash a plane because a customer's inspector is silly).

Yes you are right but I did mean same. datum simulator of curved surface is whole countersurface, you can must use at least 3 points to establish datum. But you can use more, best whole surface as I said. But it's up to you how many, what's what i meant.