Crumple: derived dataset, QC pipeline and analysis code for "Crumple: An Empirical Crush and Force Envelope for the Modern Australian Vehicle Fleet"

Derived dataset, quality-control pipeline, analysis code, reports, and figures behind the paper "Crumple: An Empirical Crush and Force Envelope for the Modern Australian Vehicle Fleet". Contains: the flattened NHTSA bulk pull (10,876 crash-test records); QC outputs with per-row reject reasons; the Australian-fleet-matched cohort; dynamic-crush derivations from accelerometer double-integration; barrier load-cell force results; impact-height (AHOF) results; fitted quantile-envelope design tables; all analysis scripts; machine-generated QC/validation reports; and the paper's figures including the Annex C design-artefact sheets. Raw cached NHTSA records are not redistributed — they are public data re-buildable with the included fetch scripts. Version 1.2 is a corrigendum, and backs Version 3 of the paper. It corrects a data-acquisition defect that contaminated the derived dynamic crush in versions 1.0 and 1.1. NHTSA's static file store served the same plain-text accelerometer file for different per-curve requests of a test (the first curve's record, truncated at a 4,096-byte boundary) while answering every distinct URL with HTTP 200 and a cleanly parseable payload, so the acquisition step did not detect it. 1,069 of 2,279 cached channel files were affected, reaching 170 of the 614 rows (27.7 %) of the fitted cohort. Channels are now read from the intact per-curve binary UDS-1992 payload in the same API response, behind a hash-based acquisition guard, and dynamic crush has been re-derived for the full corpus. Effect on the fit: 594 rows in place of 614; velocity exponent 1.272 → 1.110; mass exponent 0.182 → 0.023. Design force at the stiff 5th percentile moves only +0.2 % to +7.6 %, but crush at the soft 95th percentile falls 17 % to 23 %, so crush and stand-off values taken from versions 1.0 and 1.1 were over-generous. The barrier load-cell force and impact-height (AHOF) results are unaffected: that acquisition route was checked and found clean. The estimator, the QC thresholds and the cohort definition are unchanged, so the data correction is not confounded with a method change. Versions 1.0 and 1.1 remain public and carry the defect. Full census, repair trail and per-coefficient impact tables are in this deposit: README.md, reports/channel_cache_audit_report.md, reports/dyn_crush_repair_report.md, reports/repair_impact_report.md Data, reports and figures: CC BY-NC 4.0. Code: PolyForm Noncommercial 1.0.0.