Robustness of CPL dark-energy constraints from Pantheon+, Union3, and DES-SN5YR Dovekie Type Ia supernovae and DESI DR2 BAO

We test the compilation robustness of constraints on the Chevallier–Polarski–Linder (CPL) dark-energy parameterization using three alternative Type Ia supernova likelihoods: Pantheon+, Union3+UNITY1.5, and DES-SN5YR Dovekie. Each is combined separately with DESI Data Release 2 baryon acoustic oscillations and a compressed Planck 2018 distance prior. For every compilation we fit spatially flat LCDM and CPL to an identical within-pair data vector and covariance. We retain the release observable conventions and full covariance information, profile the common supernova distance scale, and restrict information-criterion differences to valid within-compilation comparisons. The compilations share supernovae and are not statistically independent; they are alternative likelihood pipelines, not factors in a joint likelihood. The separate compilation-specific CPL comparisons give Pantheon+: delta AIC 0, delta BIC 9.435; Union3: delta AIC 0, delta BIC 0; DES-SN5YR Dovekie: delta AIC 0, delta BIC 7.125. The registered audit does not permit a detection claim. This study is therefore a reproducible robustness cross-check, not an independent claim of evolving dark energy. Every result is tied to input checksums, frozen configurations, and quantitative chain-quality gates. The deposited package contains the analysis code, the frozen configurations, the provenance manifests recording the release and checksum of every public input, the complete LaTeX source, the test suite, and the compact result summaries and figures behind every reported number. Raw external data releases and the sampler chains are not redistributed; ARCHIVE_CONTENTS.md states exactly what is included, what is excluded, and why. A SHA-256 checksum is supplied alongside the archive. The figures in the deposited PDF are raster images at publication resolution; the full-resolution vector originals rebuild from the archived code. Preprint. Not peer reviewed. This work was produced with substantial AI assistance (Claude, Anthropic). The author takes full responsibility for the content.