NeuroState-D as a Matrix-Determined Drift Baseline: Cross-Architecture Stress Testing Under Social Engineering

v2 (erratum, 2026-07-05): This paper's Section 4.2 cites, from 論文1 (NeuroState-D), a Transformer G_thresh_D comparison table including Gemini=29.3 and Codex=12.5. 論文1's own 2026-07-05 erratum (DOI 10.5281/zenodo.20937328, v2) shows these values are success-detector false-positive artifacts: the S5 scenario's disclosure request exists only in its final turn, at which point G is deterministically 0 for every model (this paper's own model-independence proof, Section 4.1/Appendix A, establishes the mechanism). Reconstructing 論文1's archived data reproduces the Gemini/Codex figures exactly as the mean G at turns where no disclosure was requested. This paper's own primary contribution -- cross-architecture validation that the D-condition G trajectory is instrument-determined, not model-determined, across RNN/SSM/LSTM/Transformer backends -- is unaffected, since it uses only the deterministic G trajectory and does not depend on any success label. Only the imported Transformer comparison table (Section 4.2) and the now-moot OQ2 (Section 5.4) are withdrawn. Full derivation in errata_2026-07-05_gthresh_d_artifact.md; fully corrected draft (Draft v0.7) attached as paper3_draft_v0.7_corrected.md. Coordinated with parallel corrections to 論文1 (DOI 10.5281/zenodo.20937328) and 論文2/NeuroState-R (DOI 10.5281/zenodo.20937329).