Regrouped propagation passes its qualification

All 24 generated identities pass across the three physical candidates and both processing cases. The construction preserves every one of the 27 nuisance directions. It combines force and gravity before their common duration is bounded and uses one correlated clock for the reference position advance.

• All 162 tangent-column checks pass. The largest finite-difference discrepancy is 13.7 micrometres per normalized parameter, below the frozen 20 micrometre limit.
• Centers agree with the unchanged native engine to within 8.9 nanometres in these fixtures. That is numerical agreement, not instrument accuracy.
• All packet-boundary and broad-region probes stay within their Taylor bounds.
• All six independently generated stationary-pose controls retain the known fixed position, including nonzero field/removal cases.

The study spent 78 native calls and 348 bound evaluations in 45.8 seconds. Every generated input, native output/Jacobian and bound is preserved. The completed freeze cannot be restarted or expanded.

This qualifies the finite generated comparison. It does not yet show useful long-record uncertainty, certify all floating-point behavior or establish the actual receiver-error law. The next separate experiment repeats the complete orientation-cover comparison with the same original inputs, physical limits, sectors and improvement target. It will preserve the earlier broad results and save its training constraints for subsequent metadata-only diagnosis.

Controls and frozen limits · Regrouped orientation-cover plan · Method · Roadmap.