First versus Random Nonfirst Failure Reports for One-Step Code Repair — Scientific Reproducibility Package

Scientific reproducibility package for the completed comparison of canonical first (F) versus uniformly sampled nonfirst (U) wrong-output failure reports for one-step code repair. Contains the fixed 59-task frame, 826 assignments, report pools, exact scientific prompts and request settings, faulty and returned programs, retained observations, outcome tables, specification-example overlap inventory, diagnostic-to-assessment analysis, scientific provenance and standard-library verification scripts. The scripts reproduce frozen assignments and retained-value calculations without model inference or candidate execution. Model weights, backend binaries, private operational records and host-specific containment configuration are excluded. This package supports measurement inspection and recorded-value reproduction, not portable native replication or independently witnessed execution chronology. Original data/documentation are CC BY 4.0; original analysis and verification scripts are CC0 1.0. Third-party material retains applicable MIT/Apache-2.0 terms and notices as explained in LICENSE.txt and SOURCE_AND_PERMISSIONS.md. No blanket ownership of model-generated code is asserted. Manuscript acceptance or publication is not claimed.