History dependent semi_lex_compare/3

[j4n_bur53]

Native compare/3 has violations for both ground and non-ground terms.

@kuniaki.mukai

Borrowing the Monkey test, I have tested @<.

% ?- b_setval(compare_history, []), aggregate_all(count,
% (between(1,1000000,_), random_cyclic(A), ground(A),
% random_cyclic(B), ground(B), random_cyclic(C), 
% ground(C), A @<B, B @<C, \+ A @< C), F).
%@ F = 2.

If you filter ground/1 you get fewer tested cases, and you get a smaller count.
So that it drops from F≈7 to F≈2 is not unlikely. Also every test run might show
a different number, since the random number generator might not hit

the same test cases in each run. But for use cases such as term_factorized/3,
which uses rbtrees in turn, I guess you need a compare/3 that can also
deal with non-ground terms.

Do you have a semi_lex_compare/3 that also works for non-ground terms?
The proposals of @ridgeworks and my proposal are supposed to both work
for non-ground terms as well, but they depend on term_factorized/3,

which seems to be buggy at the moment:

Glitch in term_factorized/3, stack overflow
https://swi-prolog.discourse.group/t/glitch-in-term-factorized-3-stack-overflow/6433

Edit 01.04.2023
Currently the fuzzer does not sample non-ground terms very well, for
example it would not generate s(X,X) as a test case, i.e. multiple occurence
of the same variable. Maybe should work on a new version of the fuzzer.

Variables are rather generated by accident as a special case of a circular term.
X = X, is a circle of length zero, if I remove randomly generated subterms,
that are variables, the fuzzer would generate only ground terms.