Compartment-dependent selection regimes maintain the G-quadruplex paradox across the eukaryotic kingdoms
G-quadruplex (G4) DNA structures are depleted from coding sequences (CDS) and enriched in regulatory regions across diverse genomes (the "G4 paradox"), but the selection regime maintaining this architecture across eukaryotes remains unclear. Analysing 198 reference genomes balanced across the four kingdoms (Animalia 86, Fungi 49, Plantae 33, Protozoa 30), we pro-pose a two-regime architecture in which the dominant selective force depends on the genomic compartment. In CDS, codon-adaptation selection takes priority: naive Nei-Gojobori comparison shows 5-16% synonymous-rate (dS) suppression at G4 codons in six of seven unsaturated species pairs, but the signal is absorbed by per-gene codon-adaptation-index (CAI) adjustment (0/7 pairs retain a G4 odds ratio below 1; five reverse to OR > 1). In introns and promoters, where no codons exist, selection acts directly on G4 structure: phylogenetic generalised least-squares (PGLS) yields five Benjamini-Hochberg-significant kingdom-specific helicase associations (q < 0.10), all at intronic or promoter G4s, including context-dependent FANCJ/BRIP1 ({beta} = +1.60, q = 0.057 for intron G4 in Plantae; {beta} = -2.49, q = 0.004 for promoter G4 in Proto-zoa); three of five remain significant on an independent Open Tree of Life backbone. Intronic G4 are concentrated in deeply conserved orthologue groups ([≥] 87% in every kingdom), consistent with maintenance of regulatory G4 host-gene context since early eukaryotic evolution. Interim-correction notice (bioRxiv v3, 2026-08-03)This preprint version corrects a set of documentation-level errors identified in a coauthor audit of the analysis pipeline (available as AUDIT_REPORT.md in the Zenodo deposit). Several central analytical claims of this manuscript -- the definition of the log2 IRR (density-vs. fraction-ratio), the downstream use of consensus-stable versus candidate G4 motifs, the CAI-controlled dS adjustment model, the filter chain in the stable-G4 classifier, and the reconciliation of PGLS output across main, alternative-tree, robustness, and Supplementary Information files -- are currently under clean-environment re-verification. The outcome of that re-verification may change specific numerical values, the direction of some regional sign summaries, and the framing of downstream conclusions in a subsequent version. Readers should treat the present version as an interim documentation correction, not as a fully verified revision. Users planning to build on the deposited results should consult CORRECTION_NOTES.md and AUDIT_REPORT.md in the Zenodo deposit for the current list of open methodology-level issues.