Heisenbug involving attributed variables, lists and (is)/2

[bakaq]

Stumbled on this one in bakaq/constrained.pl#16 (comment). The most I could isolate it is this:

:- use_module(library(atts)).
:- use_module(library(format)).
:- use_module(library(diag)).
:- use_module(library(lists)).

:- attribute a/1.

verify_attributes(_,_, []).

asdf1([_|Xs], N) :-
    N1 is N - 1,
    asdf1(Xs, N1).

asdf2([_|Xs], N) :-
    true,
    N1 is N - 1,
    asdf2(Xs, N1).

% This errors.
test1 :-
    put_atts(A, a(1)),
    asdf1(A, 1).

% But this loops, as expected.
test2 :-
    put_atts(A, a(1)),
    asdf2(A, 1).

?- test1.
   error(type_error(evaluable,user/0),(is)/2).
?- test2.
   % Loops.

The decompilation of the predicates is as follows:

% asdf1/2
allocate(2).
get_list(level(shallow),x(1)).
unify_void(1).
unify_variable(y(1)).
put_variable(y(2),1).
sub(x(2),1,1).
call(is,2).
put_value(y(1),1).
put_value(y(2),2).
deallocate.
execute(asdf1,2).

% asdf2/2
allocate(3).
get_list(level(shallow),x(1)).
unify_void(1).
unify_variable(y(1)).
get_variable(y(2),2).
call(true,0).
put_variable(y(3),1).
sub(y(2),1,1).
call(is,2).
put_value(y(1),1).
put_value(y(3),2).
deallocate.
execute(asdf2,2).

I don't know enough about the WAM to understand this properly, but I guess the problem is the sub(x(2),1,1). (which gets corrected to sub(y(2),1,1). in the version with the extra true). It seems that only 1 x register is allocated but it's trying to use the second one, while in the version with an extra true it uses the correct y(2) register.