FORGE code, trained models and source data for RNA backbone sequence-recovery diagnostics

Code, trained model artifacts, processed result tables, public-input manifests and figure source data supporting the manuscript "FORGE decomposes RNA sequence recovery into memorisation, family transfer and residual prediction error". The release contains the FORGE feature-extraction and analysis code, trained XGBoost model files for nucleotide identity and auxiliary atlas tasks, locked result artifacts used by the manuscript figures and tables, and manifests for the public PDB/Rfam/OpenKnot-derived inputs. Raw PDB/mmCIF coordinate caches and full third-party software repositories are not redistributed; they are identified by accession/source manifests and can be retrieved from their original public repositories. Version 0.3.0 adds the Nature Communications manuscript source, the final Figure 1 pipeline, the sequence-recovery decomposition and confidence-transport analyses, the family-disjoint in-domain evaluation, and the final AlphaFold 3 and design-audit source data. It also updates the text, figure source data and release documentation to the current submission.