Every source midpoint and endpoint now has a checked reference row
All six generated sources pass. The comparison checks every packet midpoint and endpoint, including all 27 physical columns: 9,600 compiled source steps, 9,600 independent steps and 3,225,600 scalar comparisons in 18.82 seconds. The largest value difference is about 2.9 × 10⁻¹⁴ m/s²; the largest coefficient difference is about 4.6 × 10⁻¹⁵.
The native calculation uses a fixed 10,768-byte working state. Packet inputs remain immutable, and derivative rows are emitted as needed rather than stored for the entire recording. Independent quaternion/Jacobian calculations agree with the native arithmetic. Local complex-step matrix-exponential controls also check nonuniform timing, exact packet boundaries and all coefficient columns.
This removes a reference-map and memory obstacle to using the tighter source interval throughout a recording. It does not yet connect that interval to velocity and position calculations or qualify final prediction widths. The next saved-baseline screen checks what the already-qualified transport layer improves; the remaining source integration uses these streamed rows.
No aircraft result changed, and no aircraft data or protected ESA data was used.