eval: expand benchmark dataset from 6 to 15 fixtures covering race conditions, TOCTOU, async misuse, N+1, path traversal, insecure deserialization, and missing auth

feat: add evaluation framework with precision/recall metrics and model comparison