Earlier observations stay in the prediction
All six generated cases pass. Two training stages retain every earlier constraint, all 27 calibration directions and all 24 shared expression terms. Independent exact checks confirm the saved prediction endpoints and retain the known answer. The run completed 60 native queries and 30 bounds in 94 seconds.
These are deliberately small software examples, not aircraft evidence. They qualify the staged path before applying it to complete recording histories. The earlier loader-failure study remains preserved with all six failed starts; this separate version changes only the three imports required by that loader.
The full-history feasibility test uses eight-minute inputs, the broad calibration bounds and explicit clock uncertainty. It must produce useful ranges, not merely contain the answer.