Reproducibility package for rsx: a high-performance streaming toolkit for RAD-seq sex determination

Reproducibility package for the revised rsx software paper, recording rsx v0.2.9 and complete CUDA coverage for the per-marker association tests, the Beta-Binomial Bayes factor with its directional posterior, and the sample-PCA Gram matrix. It includes the raw per-repetition A100 and RTX 5070 rows behind every reported device ratio, the scripts that derive the manuscript tables and crossover figure from them, CPU/GPU parity evidence, literature datasets, binding checks, versioned TOML profiles, hydrated configurations, and pre-execution reproducibility archives. The 35.15 GB literature dataset is retained unchanged from the earlier versions. This version supersedes 0.2.9-resubmission.7, whose Implementation section carried a defence of the development method rather than the disclosure of it; the coding-assistant disclosure stays and the rest moved to the response letter. Earlier, it superseded 0.2.9-resubmission.6, whose manuscript PDF was built over a left-over .aux and froze four pages of float placement that a compile of the accompanying source archive does not reproduce; the build now starts cold and the manuscript here is page-for-page what the archive produces. Earlier, it superseded 0.2.9-resubmission.5, which was packaged before the manuscript's final prose pass; no reported number differs, and it also corrects the path this guide gives for the erfc generator. Earlier, it superseded 0.2.8-resubmission.4. A pre-submission audit reconciled the manuscript and response against the deposited measurements and then found that the erfc polynomial the paper described had never been tested: fitted as one degree-40 basis across [0, 6], its Horner intermediates reach 6^40 and it returned negative probabilities above t = 4, far from the 8.22e-17 its generator reported. Nothing called it, so no result was ever affected. rsx v0.2.9 refits it as 24 panels of width 1/4 in a centred variable, measured at 4.4e-16 relative error against the platform function where a threshold falls and at parity on cost, and the paper now reports evaluated error rather than fitted bounds throughout. Every reported result was produced by rsx v0.2.8 and is unchanged. Paper source is pinned at commit 16dbc3db80cf634367e50b1a14c9bff02b720386 (tag submission-v2); the reproducibility tree at 98d52c5d8a75a568ec210844f0acf1f1beff0faf. RSX_DEPOSIT_README.org describes every component and how to verify it.