While working on #1196, I've come across situations where two different heuristic solutions have exactly the same indicators.

Example on instance X-n251-k28 from the CVRP benchmarks:

Of course this is related to the fact that those benchmarks use euclidean distances with rounding and this is much less prone to happen with real-life instances: it would require different solutions with "real" indicators to be equal to the second and meter.

Yet the consequence is that in this situation one of the solution "shadows" the other and one of the local search phase is skipped, potentially resulting in a different overall solution. This was already a thing before but used to be deterministic since the deduplication was done from heuristic solutions stored in a std::vector. After #1196, this is now depending on which heuristic solution we get first., resulting in different solutions across runs in those edge cases.