Factored transfer selection of binarization schemes: code, protocol, manifests, tables and per-run results

This record is the material behind the manuscript of the same name: everything needed to rerun the study, and everything needed to recompute its numbers from the runs already made. It holds the code, including the reference implementation the patch is applied to, and the protocol with its two dated amendments. It holds the exact command line of every run, so a campaign can be repeated as it was executed rather than as it is described. It holds the trained policies, one per problem family and one per family and optimizer, and one csv per instance, optimizer and scheme with a row per seed, which is the level at which every table of the manuscript is computed. It also holds the per-generation traces of the convergence sample and the outputs of the analysis scripts. The runs come from nine campaigns. Three reference factorials over the 34 binarization schemes were produced here, on the railway level of set covering, on feature selection and on unit commitment; the remaining four families reuse the factorial of the earlier calibrated layer study, deposited separately and cited in the manuscript. The rest are the calibration phase, the main evaluation over 115 instances of six families, six optimizers and 31 seeds, the extreme-scale study on two railway instances of about one million columns, the transfer study across optimizers, an ablation with the paid oracle disabled, and the sample of traced runs. Together they are roughly 143 000 runs. Recomputing the numbers takes minutes and rerunning a campaign takes days, so README.md gives both routes and says which scripts serve each. SHA256SUMS.txt carries a checksum for every file, and MANIFEST.json records what was collected and whether any campaign was short of runs when the archive was built.