The instruction matcher was missing a few details, which made it faster than it should have been. Right now I'm trying to keep the design of these in sync. Once I fixed that it exposed a legitimate bug that was blocking the github benchmark.
