Diagnose cost and preserve horizontal Taylor information

The screen is complete and terminal: two compatible bounds, four time failures. Never restart it or increase its allowance. The earlier native qualification remains valid; the complete prediction comparison remains blocked.

First complete the shortened runtime profiles under their separate finite freeze. Use measured costs before changing the implementation, and preserve any partial profiling failures.

The completed disc bounds also show horizontal widening. The current adapter converts a value interval about an arbitrary supplied row; that can lose linear cancellation. Develop a source-specific Taylor or derivative-error certificate before retaining an earlier remainder about that row. Keep all 27 physical and 24 shared columns, source identities, training constraints and explicit center/row differences. A compatible per-axis minimum may retain an old bound, but cannot invent a new narrower one.

Any revised implementation needs separate local controls and a new finite native/long-source qualification. Do not run the complete prediction comparison or aircraft fits merely because one component or one axis is tighter.