Redefining constraints with outlined result variables reverses their context #260
Description
Example: xy <= 4 can be redefined into a quadratic functional constraint z ~ xy, z<=4. Both z and x*y are in negative context.
z ~ xy can be outlined into a static constraint z >= xy, but here z is in CTX+! Then, e.g., linearizing z >= xy (e.g., x is binary) and propagating context, we get a contradiction, because z ~ xy has already been redefined into z >= x*y with z in CTX-.
Another example (nlmodels/chemeq.mod for a pure-linear solver): assume a quadratic expression is first stored as QuadratricFunctionalConstraint, in CTX_POS. Then it is redefined as res_var <= <expr>. Then assume it is linearized (see mul.h) and the resulting linear constraint is added as AddConstraint_AS_ROOT, propagating context into arguments. Then res_var receives CTX_NEG.
In this example it is possible to skip propagation after linearization, as we don't get any new arguments in the linearized constraint. However it is to check if we ever add new expression arguments tp a redefined functional constraint. Then, probably we need to propagate into the new expressions manually. Yet another example: nlmodels/weapon.mod for MILP/MIQP solver.