The long-source screen remains incomplete
Two of six generated cases completed. All four globe cases reached the unchanged time allowance. The study is terminal; no failed case will be retried under this allowance. All six bounds and 38,400 world/source steps remain charged.
The two completed disc cases preserve the reference centers, coefficient matrices and training bindings. Their height allowances are 29–37% smaller than the best saved bounds. Their fresh horizontal allowances are wider, so the compatible saved north/east allowances remain the retained minimum.
This is a numerical bound-development result, not a comparison of aircraft fits or model support. It changes none of the 53-stretch findings. It does not pass the complete prediction-width gate, which remains four of six groups.
Two implementation issues now need separate work: runtime on the globe cases, and horizontal error bounds that retain useful cancellation. The finite shortened profiles diagnose current costs. Wider source-row bounds must be improved through a justified Taylor/reference calculation, not by dropping columns or loosening targets.