Hi @wjrforcyberm, thanks for your effort in improving mockturtle!
Unfortunately, the change in your PR leads to suboptimal performance in the cut_rewriting algorithm with compatibility graphs.

If you run the tests, you'll likely see a failure in quality.cpp at line 128:

Expected: { 0, 20, 80, 49, 160, 79, 195, 131, 506 (0x1fa), 2, 258 (0x102) }
Got:      { 0, 20, 74, 47, 34, 77, 103, 87, 275 (0x113), 2, 202 }

These numbers reflect node reductions in the MIG. The reason is that best_gain tracks the best replacement per cut at a given root. Your change skips cuts that could be valid just because they don't beat the best gain from a different root, which breaks the intended behavior.

Let me know if this makes sense, if you can reproduce the failure, and if we can go ahead and close the PR.