EvalMonkey debuts local agent chaos testing

This is useful because it measures the failure mode most evals ignore: how much an agent degrades once the happy path stops being clean. The local-only design removes the biggest adoption blocker for teams that do not want to send eval traffic to a third-party service. Pairing a capability score with a resilience score is more actionable than a single benchmark number, especially for tool-using agents. The chaos profiles map to real production failures, which makes this relevant for orchestration, middleware, and agent reliability work. The maintainer-wanted framing suggests the project is early, but the problem it targets is real and under-served.