Artifact-Certified Long-Support Prime Exponential Sums with Adversarial Null Forensics

This record contains the manuscript, source files, figures, and reproducibility materials for Artifact-Certified Long-Support Prime Exponential Sums with Adversarial Null Forensics. The work studies finite prime-only and prime-power-complete von Mangoldt exponential sums under a shared hard-cutoff, endpoint-subtracted localization protocol. Its central positive result is a precommitted known-answer benchmark on \(t ∈ (90,120)\): inclusion of prime-power von Mangoldt terms reduced ordinal top-13 localization mean absolute error relative to the prime-only arm at each of three finite cutoffs, \(P=100{,}000\), \(1{,}000{,}000\), and \(5{,}000{,}000\). The full-\(Λ\)/prime-only MAE ratios were approximately 0.70, 0.94, and 0.79, with a geometric mean of approximately 0.80. The record also preserves negative and fail-closed evidence, including a failed position fixture, a coordinate-coverage diagnosis for an earlier low-frequency band, an endpoint-origin seam analysis, and audits of unsuitable null families. Reported computational objects and evidence states are bound to explicit manifests and cryptographic hashes. The package distinguishes finite numerical agreement and provenance evidence from analytic truncation control or theorem-level certification. This is a computational-methodology and artifact-forensics record. It does not claim a new zeta zero, an asymptotic theorem, a new explicit-formula theorem, progress toward the Riemann Hypothesis, or an RH proof. For the larger raw corpus, this record includes a manifest-bound inventory and reproducibility references. The raw payload itself is retained separately because of its size and file count.