The Hubble Constant from Symbolic Regression: Adversarially Validated with DESI DR2, M(z) Non-Evolution, and External Constraints

We apply symbolic regression via PySR to cosmic chronometer (CC), BAO (SDSS + DESI DR1/DR2), and three independent supernova samples (Pantheon+ full covariance, DES-SN5YR, Union3) to discover the cosmic expansion history H(z) with minimal theoretical priors. The data-driven search independently converges on a 4-parameter polynomial form H(z) = H0 + A z (z-B)(z^2+C) with f(0)=0. Joint fitting yields H0 = 68.0 +/- 0.8 km/s/Mpc (68% CL), consistent with Planck 2018 at 1.2 sigma. DESI DR2 (factor-two improved precision) confirms the DR1 result with H0 unchanged at 68.3. We test for evolution of the supernova absolute magnitude M(z), finding alpha = 0.020 +/- 0.010 consistent with zero. External constraints from GW170817, DES Y3+GW, and TDCOSMO 2025 lensing combine to H0=68.8+/-2.3. A direct LambdaCDM fit gives H0=67.9, joint chi^2=1429.4 vs SR 1430.6 (Delta chi^2=1.2). Fixing M to the SH0ES Cepheid calibration is rejected at Delta chi^2 = +52 to +82 (7-9 sigma). An adversarial validation exercise (2 rounds, 14 challenges, 10 partially sustained, 4 rejected, none fatal) confirms robustness. We conclude that the Hubble tension resides in the Cepheid calibration of M, not the expansion history shape; H0 approx 68 regardless of model choice, data combination, or adversarial scrutiny.