The linter ecosystem is actively migrating silent failure into an undetectable idiom

In a well-known family of LLM evaluation bugs, the failure doesn't look like a failure. A judge model goes down, and the benchmark keeps publishing scores — because somewhere in the code, a handler li...