Controlled Diagnostics of w0CDM Support Conditioning with Planck 2018 FULL and DESI DR2 BAO: A Structured Early–Late Likelihood Redistribution

We perform controlled diagnostics of w_0CDM support conditioning using Planck 2018 FULL CMB likelihoods combined with DESI DR2 BAO measurements. Three inference branches are compared using identical datasets, likelihoods, and numerical infrastructure: (i) a neutral control run with broad prior support in w_0, (ii) a non-phantom restricted-support run enforcing w_0 > -1, and (iii) a phantom restricted-support diagnostic run enforcing w_0 < -1. The component-wise likelihood decomposition reveals a structured early-late split: the non-phantom branch improves the lowest-redshift BAO component (LRG1, z ≈ 0.14) by Δχ² ≈ +1.3 while incurring a modest penalty in high-ℓ CMB (Δχ² ≈ -0.6), whereas the phantom branch exhibits the opposite pattern. These component-level effects largely cancel in the total goodness-of-fit (|Δχ²_total| < 0.2), indicating redistribution along the existing (w_0, H_0) degeneracy rather than decisive model preference. Early-universe parameters remain stable across all branches (shifts < 1σ). We interpret the results as a Stage Zero diagnostic establishing numerical admissibility and internal likelihood structure under support conditioning, serving as a baseline for incorporating genuinely late-time probes in future work. All analyses use Cobaya with CLASS in a four-chain MPI configuration. Complete MCMC chains, configuration files, and reproduction notebooks are publicly archived on Zenodo.