Hi Igor,

there is a quite elementary way to see that star products restrict to open subsets: it's essentially part of the definition of a star product. Here, I will focus on the case of smooth (symplectic/Poisson) manifolds, where you do not have to rely too much on sheaf-theoretic notions. Now, $\star$ being a star product on $M$ means that it consists in each order of $\hbar$ of *bidifferential* operators if you consider a differential star product, or of local operators if you consider a local star product. Local operators are, by a celebrated theorem of Preetre locally differential, i.e. on a sufficiently small open subset the restrict to differential operators. Globally, the only problem may be that their order is infinite (take the real line and a bump function with supp in the unit interval. Now translate the bump function to a bump function $\chi_n$ having supp in $[n,n+1]$ Then taking the sum of all $\chi_n \frac{\partial^n}{\partial x^n}$ is a local but not a differential operator). In any case, bidifferential/local operators restrict to open subsets.

THis induced for every open $U \subseteq M$ a new product $\star_U$ for $C^\infty(U)[[\hbar]]$. The point is now that associativity can be check locally, as in each order of $\hbar$ it is an equation between (multi-)differential operators which localize! So you indeed get a star product.

Unfortunately, with the advent of Kontsevich's formality these things and techniques were mostly forgotten. It was kind of standard arguments and it's also present in all the earlier constructions of/with star products in the 80's... It might be that it is not spelled out here explicitly in the literature, but you just have to look at the early papers of Bayen etc as well as Gutt, Cahen, deWilde, Lecomte and so on, and scan for the words "local operator" or Peetre Theorem ;)

Now for the second question about functoriality. This is of course much more subtle. The naive answer is that there is no thing like a quantization functor (with appropriate continuity properties) as one has the no-go theorem of Gronewold and van Hove. I'm sure you know this. So one has to refine things a bit: the crucial question seems to be what the domain of this functor should be: symplectic manifolds per se is not a good choice. A better choice will be symplectic manifolds with a symplectic connection (this is a huge choice to make, as there are zillions of symplectic connections...)

In this case, one indeed gets a functorial quantization by means of the Fedosov construction of a star product (say for trivial characteristic class) This you can find at many places in the literatur (if you're crazy enough to read german, you can take a look at my book ;) As Theo say, one can modify the construction using a series of closed two-forms, so it might be better to chose this as a domain of you functor...

Simillar things hold in the Poisson case as well: you have to choose a formality for $\mathbb{R}^n$ once and for all, say the Kontsevich one, subject to certain invariance properties ($\mathrm{GL}(n)$-invariant should do the job). Then you can globalize this formality by Dolgushev's construction. This gives a functorial construction of a formality for the price of choosing a connection on each manifold first (this is needed in Dolgushev). The again, one get's functorial star products... This is sort of implicit in a paper by Kontsevich. If I remember correctly, in an appendix... BUt more details on it are in the papers by Dolguushev.