Perfectly Consistent, Still Wrong — LLM Judges Keep Failing the Same Way, Plus Cheaper Alternatives and Unified Observability