Saved-only diagnosis of the remaining width
Inspect all 16 specified-field regions in the completed saved-bound comparison. Verify the evidence receipt and independently recalculate both east-coordinate support certificates and outward endpoints in each region: 32 support checks. Read only sealed generated results. No new bounds, integration, LP proposals, optimizer starts, aircraft reads or future-truth inputs are allowed.
Separate the saved affine-support term, future remainder and arithmetic allowance. Identify the regions determining each complete group's lower and upper endpoints and the providers supplying their future remainders.
Also remove the future remainder algebraically while keeping the saved training constraints and supports unchanged. This is a sensitivity diagnostic, not a valid tighter prediction interval. It does not prove the remainder can be removed, an optimum was found or a physical model fits better. Do not use it in scientific decisions or relabel the completed width comparison.
Record source and input hashes, all 16 regions and original target widths. Use one guarded numerical thread and disk-backed scratch. The arithmetic is standard-library rational arithmetic; the finite work is 32 saved supports.