I think the "Angles" removal demonstrates the optimization is unsound. The "optimized" code gives out a result when a correctly optimized one would give a ReferenceError. This optimizer ignores invalid code (that references a non-existent thing) and happily short-circuits it.

The only reason to trust this code to do other optimizations well is all the evidence it's targeted at one specific benchmark.