The following prolog code establishes a very simple grammar for sentences (sentence = object + verb + subject), and provides some small vocabulary.
% Example 05 - Generating Sentences
% subjects, verbs and objects
subject(john).
subject(jane).
verb(eats).
verb(washes).
object(apples).
object(spinach).
% sentence = subject + verb + object
sentence(X,Y,Z) :- subject(X), verb(Y), object(Z).
% sentence as a list
sentence(S) :- S=[X, Y, Z], subject(X), verb(Y), object(Z).
When asked to generate valid sentences, swi-prolog (specifically swish.swi-prolog.org) generates them in the following order:
?- sentence(S).
S = [john, eats, apples]
S = [john, eats, spinach]
S = [john, washes, apples]
S = [john, washes, spinach]
S = [jane, eats, apples]
S = [jane, eats, spinach]
S = [jane, washes, apples]
S = [jane, washes, spinach]
Question
The above suggests that prolog always backtracks from the right to the left of conjunctive queries. Is this true for all prologs? Is it part of the specification? If not, is it common enough to be relied upon?
Notes
For clarity, by backtracking from the right, I mean that Z is unbound and rebound to find all possibilities, given the first matches for X and Y. Then after these have been exhausted, Y is unbound and rebound, and for each Y, different Z are tested. Finally it is X that is unbound then rebound to new values, and for each X the combinations of Y and Z are generated again.
Short answer: yes. Prolog always uses this same well defined scheme also known as chronological backtracking together with (one specific instance) of SLD-resolution.
But that needs some elaboration.
Prolog systems stick to that very strategy because it is quite efficient to implement and leads in many cases directly to the desired result. For those cases where Prolog works nicely it is pretty much competitive with imperative programming languages for many tasks. Some systems even translate to machine code, the most prominent being the just-in-time compiler of sicstus-prolog.
As you have probably already encountered, there are, however, cases where that strategy leads to undesirable inefficiencies and even non-termination whereas another strategy would produce an answer. So what to do in such situations?
Firstly, the precise encoding of a problem may be reformulated. To take your case of grammars, we even have a specific formalism for this, called Definite Clause Grammars, dcg. It is very compact and leads to both efficient parsing and efficient generation for many cases. This insight (and the precise encoding) was not that evident for quite some time. And the precise moment of Prolog's birth (pretty exactly) 50 years ago was when this was understood. In the example you have, you have just 3 tokens in a list, but most of the time that number can vary. It is there where the DCG formalism shines and still can be used both to parse and generate sentences. In your example, say you also want to include subjects with unrestricted length like [the,boy], [the,nice,boy], [the,nice,and,handsome,boy], ...
There are many such encoding techniques to learn.
Another way how Prolog's strategy is further improved is to offer more flexible selection strategies with built-ins like freeze/2, when/2 and similar coroutining methods. While such extensions exist for quite some time, they are difficult to employ. Particularly because understanding non-termination gets even more complex.
A more successful extension are constraints (constraint-programming), most prominently clpz/clpfd which are used primarily for combinatorial problems. While chronological backtracking is still in place, it is only used as a last resort either to ensure correctness of solutions with labeling/2 or when there is no better way to express the actual problem.
And finally, you may want to reconsider Prolog's strategy in a more fundamental way. This is all possible by means of meta-interpretation. In some sense this is a complete new implementation, but it can often use a lot of Prolog's infrastructure thereby making such meta-interpreters quite compact compared to other programming languages. And, it may not only be used to implement other strategies, it is even used to prototype and implement other programming languages. The most prominent example being erlang which first existed as a Prolog meta-interpreter, its syntax still being quite Prologish.
Prolog as a programming language contains also many features that do not fit into this pure view, like side effecting built-ins like put_char/1 which are clearly a hindrance in meta-interpretation. But in many such situations this can be mitigated by restricting their use only to specific modes and producing instantiation errors otherwise. Think of (non-constraint based) arithmetics which produces an error if the result cannot be determined immediately, but still produces correct results when used with sufficiently instantiated arguments like in
?- X > 0, X = -1.
error(instantiation_error,(is)/2).
?- X = -1, X > 0.
false.
?- X = 2, X > 0.
X = 2.
Finally, a word on non-termination. Often non-termination is seen as a fundamental weakness of Prolog. But there is another view on this. Also other much older systems or engines suffer (from time-to-time) runaways. And they are still used. In the case of programming languages, runaways are a fundamental consequence of their generality. And a non-terminating query is still preferable to an incorrect but terminating query.