So as explained in #981 (comment), it turns out the fix allows to filter out invalid splits, but also to reach to more valid ones. Here is a comparison of results between master (actually the last working commit 95b30e8) and this PR.

Average gap (%) on all usual HVRP instances

Computing times are really similar, even a tiny bit lower with the PR. The worst gap is greatly reduced also, regardless of the exploration level, it goes down from 43.54 to 19.17%.

That's a significant improvement! 🎉 This operator was an even better idea than we first thought. ;-)