The method admits impossible extreme examples
All 36 tested extreme matrices fail the rotation-invariant check. All six known physical anchors pass. The largest defects are about 0.74–0.87, far above the declared dimensionless numerical allowance of 0.00000001.
After removing the proposed sensor gains, a physical matrix must preserve the original matrix's rotation invariant. The saved extreme examples do not. They are feasible in the simplified prediction region, but fail this necessary physical condition. This explains one source of unnecessary freedom in the current uncertainty calculation.
It does not show that every wide prediction is impossible, or prove what the correct narrower range should be. New constraints must preserve all permitted physical calibrations before any interval can be tightened.
The next derivation will use the allowed orientation range when bounding each matrix column. The existing constraints allow any rotation, losing information about how far orientation can depart from its declared center.
The nine local known-answer checks pass, including nonorthogonal initial matrices, gain removal, stretched and nonorthogonal columns, and the fact that a Gram check alone cannot establish a proper rotation. The frozen diagnostic completed all 42 checks in 0.6 seconds with no LP proposals, integrations or original aircraft reads. The 53-stretch findings remain unchanged.