There is a related (in fact, identical) question from statistical mechanics from where the answer may be better understood: When averaging, what is better, ensemble average or time average? Time average is usually faster since the system is sampled in "typical" situations, which are linked by a trajectory in phase space. In contrast, when averaging an ensemble with realisations chosen by chance, certain realisations of extremely small probability may be chosen. This is in particularly true when the statistical ensemble is small. For large statistical ensemble (large time) both averages should be the same, of course. (This is generally assumed, however, not mathematically proven as far as I know.). So time averaging is usually more efficient than ensemble averaging, HOWEVER, this comes with a precondition: It works only for ergodic systems, that is, the available phase space is completely sampled. There are simple systems which are non-ergodic, where time averaging would fail. 

So my answer is: trajectory search is usually more efficient, provided the system is such that starting from an arbitrary initial situation the optimum can be reached.

I don't think we can judge that SA is better than other algorithms. The choice of algorithm or resolution technique is based on our mathematical formulation of the problem (objective function (linear non-linear, convexity), variables (number and bounds), Constraints and space research).

given a specific problem instance, it is impossible to assess if a heuristic A will be better than any other B. If population-based methods are properly used (i.e. crossover works as expected) then, in a single run you may have a "set" of good enough solutions to analyze instead of a single one (if you run a trajectory based methods). In my opinion, there is no sense in searching for the "best" method. Combine them, play with them and take the best solution.