Supplemental Materials for Beyond the manifest: a case study on learning test-code co-changes in JUnit 3.x/4.x migrations

Replication package for the study "Beyond the manifest: a case study on learning test-code co-changes in JUnit 3.x/4.x migrations". Contents. The mining and corpus-construction pipeline, the corpus of 4,435 unique input–output pairs mined from 123 GitHub repositories, the repository-level partition (3,608 training pairs from 49 repositories, 827 test pairs from 74), model predictions for every system reported in the paper, licence and provenance tables, and the results of the build-and-test validation. MANIFEST.json records the SHA-256 and byte count of all 110 files. Verification. REPLICATION.md describes two tiers. Tier A reproduces every number reported in the paper from the shipped artefacts, on CPU alone and without network access. Tier B rebuilds those artefacts from scratch and needs a GPU and a GitHub token. Changes from the previous version. The corpus is rebuilt and partitioned at repository level, and all model results, statistics and figures are recomputed accordingly. The manuscript sources are not distributed with this version. Licensing. The data are released under CC BY 4.0. The code under revision_code/ is released under the MIT License.