Madelon: original UCI challenge splits (Datamonger representation, version 1)

Lossless CSV representation of Madelon from the UCI Machine Learning Repository, created by Isabelle Guyon for the NIPS 2003 feature selection challenge. Preserves all 500 integer features, row and feature order, and the original training (2,000), validation (600), and unlabeled test (1,800) splits. The synthetic benchmark has five informative dimensions, 15 redundant combinations, and 480 distractors. Datamonger applies no scaling, filtering, imputation, feature selection, or label remapping. The three gzip-compressed CSV files have positional feature columns x1 through x500. Training and validation append response with original -1/+1 labels; the test file has no target. All fields decode as int64. Seven files are included: train.csv.gz, validation.csv.gz, test.csv.gz, README.md, prepare.py, provenance.json, and LICENSE. The recipe requires the Datamonger repository and its Python development environment. Source size and SHA-256 are pinned. All 2,200,000 feature values and 2,600 labels match an independent NumPy source read; source checksums pass, and Rust and reference Python canonical encoders agree. Three preparations produced identical data hashes. These are conversion checks, not an independent registry-release attestation. CC BY 4.0 applies to the data, following UCI's explicit license statement. Credit Isabelle Guyon and UCI, link CC BY 4.0, and identify the lossless format conversion and compression. MIT applies only to the preparation code and accompanying documentation.