Round 14 — middle-path findings restoration and final precision pass

Kept the paper narrow while restoring three findings with full caveats: the registered one-shot wording null (+0.0063, Holm p=1.00) beside the repeated-game wrapper shift; the P5-2 Bayesian aggregation demoted from “independent corroboration” via the prior sweep (0.138 → 0.205, α=1 crossing 0.20); and four citable Appendix A.2 secondary findings. Five precision corrections, including dropping a wrong-sided RPS interval and the D2 congruence clause (GPT-4.1's swap cell cannot separate word from payoff channels). Deliberately excluded: the post-hoc p13 trait-tension theory.