I've tested things related to #977 as follows:

• generate optimized routes for a set of random instances;
• remove various fractions of the routes to generate a set of initial solutions that are more or less "complete";
• solve with partial solution in input.

Compared to master, this PR yields much better solutions as expected since applying heuristics to partial solutions broadens the search. In particular in many situations, master is unable to assign all tasks back when starting with a single local search on the partial solution, while PR does reach to "full" solutions with no unassigned task. When solutions are full on both sides, PR reaches smaller cost solutions.

Also I've generated specific instances to reproduce the embarrassing situations mentioned here in order to confirm that this PR does fix it.

Now for the refactoring part from #1181, this PR significantly reduces the number of lines and heuristic stuff is now much more readable.

I've confirmed that we still get the exact same executions and searches on the usual benchmarks. The caveat being that I've notice a slight slowdown for the heuristics computing. I've done some profiling but I have no explanation other than the fact that the core heuristic parts are now factored in helper functions. Those seem to be inlined by the compiler but probably the added boilerplate prevent some further optimization that was happening before.

Turns out the computing times differences I reported above are similar to the differences that can be observed by just switching compiler, so I don't thing there is a specific flaw in the refactoring.