The applied=0 case is the most important failure because it invalidates the experiment before any assertion runs. Mutation harnesses should model setup, mutation, observation, and restoration as separate states, and a case can pass only if the mutation was independently observed on the target surface. I would require a precondition hash that differs after mutation and returns after cleanup, then treat 429 or unchanged content as INCONCLUSIVE rather than green. That keeps infrastructure failures from silently inflating the gate's confidence.