Preserve training information and trace dynamic direction loss

The saved audit establishes that the physical-only static bound adds no tightening in all 39 completed cases. Simply intersecting it with the existing training-derived bound would retain the old acceleration limits. Do not launch another static replacement or finish the timed-out study: neither can repair the already-failed width target in the complete globe groups.

The next investigation is the dynamic propagation of force and rotation uncertainty. The existing vector bounds use one Euclidean radius, and extracting a coordinate assigns that radius to the coordinate. This is conservative but can lose the difference between horizontal and vertical uncertainty. It is a specific implementation feature to investigate, not yet a demonstrated complete explanation of the remaining prediction widths.

Develop direction-preserving value and Taylor-remainder operations separately from the frozen engines. Retain the existing training-conditioned bounds as valid fallback information. Derive each operation before using it, and test independent known motions that separate horizontal deflection from vertical change. Preserve native centers, physical bounds and every shared-error column.

Before any further complete comparison, show that the new operation tightens the bound actually used by the current method on the same source/domain—not just a broader physical prior. Then qualify the unchanged native control matrix and freeze a separate finite comparison. No new observed fits follow from the isolated component tests. Recording-scale uncertainty and observational coverage remain open gates.