They optimize away unnecessary monitoring inside the traces once the types have been propagated there. So only slow path code does the monitoring. All the active traces should get stable types without extra monitoring at some point.

Essentially it's moving doing type propagation in "batch" to a "dynamic trigger" system. How effective that is really depends on the programs (or the programing style) and how often they really change types.

I hope it's not just optimized for the usual benchmarks, but some real code.

It varies. Remember that generational GC's used to write barrier and process every memory write into GC space, and that wasn't truly horrible for performance.