Reproduction archive for: When the calibration treatment changes the estimand — a clustered comparison of two lumped-parameter interior-ballistics engines

Frozen dataset, analysis code, source of both interior-ballistics engines, and verification harness reproducing every table, figure and interval of the article. Contents. The frozen derived dataset of 336 evaluated muzzle-velocity records (with the 332-record primary inferential subset identifiable by familia != "fast_pistol") and per-record predictions from both engines; the source of the explicit (v4) engine with a built-in catalogue self-test (9/9) and the source of the simple (v3) engine, both checked to reproduce the frozen per-record predictions (v3 to within 1e-5 m/s over all 336 records); the analysis scripts for Treatments A, B and C, the chambering bootstrap with full cross-fitting re-estimation, the log-scale jackknife, the cluster sign-flip tests and the 159,936-scenario anchor sensitivity; the figure scripts; the pipeline drivers; the consistency verifier and its injection test; LaTeX sources; and a file manifest with SHA-256 checksums. Reproduce with bash run_all.sh. Known limits. The corpus is not exactly regenerable from the original manufacturer sources, as declared in Table 1 of the article. Treatment A's anchor chamberings are not identifiable, so anchor leakage is assessed by a conditionally exhaustive sensitivity analysis rather than eliminated. The underlying manufacturer reloading data are referenced, not redistributed.