DBFC-1X: Corrected Database Forensic Classification Corpus, analytic likelihood module and evaluation harness

A corrected version of the DBFC-1 database forensic classification corpus (doi:10.5281/zenodo.20503388), supplied with an analytic module that computes exact class-conditional likelihoods and therefore the true likelihood ratio for every event — a property no real forensic corpus possesses, and the reason for this release. Corrections to v1.0: (1) a deterministic single-feature separator for the exfiltration class, where clipping bounds left non-exfiltration scan_amplification at a maximum of 5.000 against an exfiltration minimum of 16.296 across 300,000 clean events, so that a single threshold separated the class with precision and recall both equal to 1.000; (2) class-specific categorical and clipping supports that made some observations logically impossible under some classes, yielding structurally infinite likelihood ratios; (3) label-noise semantics under which the nominal 3.5% error rate was realised as 3.00%; (4) pre-noise ground-truth labels, previously discarded, now retained. Contains 24,000 reconstructed database events with 18 features in seven forensic groups, official stratified 70/15/15 splits, an analytic 18×6 ground-truth relevance matrix, a deterministic generator (seed 20260115, with a legacy flag that reproduces v1.0 byte-identically), the analytic likelihood module, the full evaluation harness, and machine-readable results for every table and figure in the accompanying paper on the sensitivity of the log-likelihood-ratio cost to reference-label error. Version 1.0 is retained rather than withdrawn so that published results derived from it remain reproducible. Data, figures and documentation are licensed CC BY 4.0; code is licensed MIT (see LICENSE-CODE-MIT.txt inside the archive).