When a new clause with only 1 atom is added with (usually
from simplifying a theory lemma), the usual strategy is to backtrack to
the base level, then propagate the atom. However, when the atom is
already false at base level, Unsat is raised, in which case the
information that the atom must be true is lost. To avoid that, the
single atom clause is simply pushed back in the stack of clauses to add,
so that it will be re-examined later.
Previously, the heap was not grown when adding local assumptions. This
lead to a bug whne backtracking: indeed when a local assumption was
backtracked, it was added to the (too small) heap, which then raised a
Sparse_vec exception.
Proofs require local assumptions to be recognisable.
Keeping the reason of local assumptions as Bcp simplfies
the code, since a proof is a clause, and allows to not have
to recreate the local unit clause that effectively propagates
the local assumptions.
With this fix, simplification of clauses is not required aymore for
levels between 0 (excluded) and base_level, so the partition function
can be simplified accordingly.
Indeed imagine the following case:
"push; assume [c]; pop; push; solve"
since c has user-level 1, in the current state, it would
have been wrongfully added to the solver state when solve
is run.
Apart from new assumptions, clause level can always
be computed from the histoy of the clause, so it is
better to do it in Solver_types when creating clauses.
Aditionally, it seems there was an error in the manual
computing of clause level somewhere, this fixes the bug.
