- move to jbuilder
- use a functorial heap (with indices embedded in lit/var)
- update Vec with optims from mc2
- change semantics of Vec.shrink
- use new Log module
When analyzing an mcst conflict clause and looking at a semantic
propagation in the trail, the last resolved clause was looked at again,
which caused an invalid history to be generated (the computation of the
backtrack clause was not affected because the second resolution of the clause
was basically a no-op thanks to the 'seen' field), thus it did not
introduce any soundness bug, just a faulty clause history which was
caught during proof expansion.
Following a bugfix for doublon elimination, doublons are now eliminated
in add_clause, and as such, in case no simplification is made, the
clause without doublons should be used (and not init). This was a typo
missed during the refactoring of doublon elimination.
Since the dimacs functor has an internal state (relating to the output
in iCNF format), it is desirable to have a generative functor, inc as
esomeone wants to output select parts of a problem to two distinct iCNF
files).
When adding clauses that conatins duplicates, the checking
of some proof would fail because there would sometime be multiple
littrals to resolve over. This fixes that problem.
When theory raises a conflict, it is analysed, and the backtracck clause
that result is added to the solver, however, I didn't find yet a
satisfying answer as to wether the original clause is implied (or not)
by this backtrack clause, so in order not to lose information, we also
add the original conflict clause when it comes from the theory (because
if not, then it comes from a conflict detected during propgation, so the
conflict clause is actually already attached).
In add_clause, a conflcit at level 0 raised Unsat, but was forgotten,
which is obviously incorrect as succesive solving of the same problem
would yield different results.
Indeed, the previous strategy was that late propagations didn't need to
be propagated since they already have been, however that may not be the
case as a conflict might arise during propagation. It manifested as a
bug when the conflict did *not* depend on local hyps, and was tragically
lost during popping.
