Branching function gets called when time limit hit during LP construction

[dchetelat]

When a time limit is set in Ecole using the scip parameter limits/time, and occurs while the LP is being constructed at a node, the LP construction is aborted, but for some reason the brancher still gets called. This results in weird states where there is no LP, and SCIPgetNLPCols returns 0, yet some variables might have nonzero LP column indices, etc. Quite dangerous.

In particular, this led to a segfault here at line 76 of

ecole/libecole/src/observation/strongbranchingscores.cpp

Lines 70 to 77 in 2c19430

	/* Store strong branching scores in tensor */
	auto const num_lp_columns = static_cast<std::size_t>(SCIPgetNLPCols(scip));
	auto strong_branching_scores = xt::xtensor<double, 1>({num_lp_columns}, std::nan(""));
	for (auto const [var, score] : views::zip(cands, cands_scores)) {
	auto const lp_index = static_cast<std::size_t>(SCIPcolGetLPPos(SCIPvarGetCol(var)));
	strong_branching_scores[lp_index] = static_cast<double>(score);
	}

because lp_index was nonzero but strong_branching_scores was a xtensor of length nb_cols=0.

This is symptomatic of more general issues: what should happen if an observation extraction function gets called after SCIP is in a stopped state? (SCIPisStopped() returns True.) My suggestion would be to add a check, in the code common to all observation functions for all environments, that avoids calling the extraction function if SCIPisStopped()==True. In that case the observation function would return {}/None.