Check the source reference map and its usable error bound

Use the same six generated sources and sector-4 domains as the completed value preflight. No aircraft data, ESA confirmation data, optimizer starts or full trajectory predictions are used.

For A=R f and B=R(P−I)f, expose a source-only calculation using the unchanged native tangent code's Dual arithmetic, calibration and exponential primitives. Independently compute all 27 columns using quaternion rotation values and the analytic right Jacobian of the rotation exponential. The native exponential's small-angle polynomial remains unchanged; record numerical discrepancies rather than normalizing matrices or silently changing the frozen implementation.

Use the native clock branch: at an exact packet endpoint, use the preceding packet's partial-step derivative. Record the selected packet, query time and distance to its boundaries. A branch-specific coefficient at a kink is not a unique two-sided derivative. Unused columns remain present as zeros.

Local deterministic tests compare against a third implementation based on complex-step matrix exponentials, check a discontinuous packet slope, nonuniform periods, support rejection, signed calibration and an exact rational affine bound. Six frozen source identities then each receive one native source reference and one independent reference calculation. Compare 162 value/coefficient entries per identity, plus its six values; coefficient agreement uses absolute tolerance 2e-10 and relative tolerance 2e-8. Value agreement uses absolute tolerance 1e-10 m/s². These numerical tolerances do not certify whole-program floating-point enclosure or observational uncertainty.

For the already sealed vertical interval [l,u], native center c and native row J, compute with exact rational arithmetic

max(|l−c|, |u−c|) + Σ |Jj| max(|lowerj−centerj|, |upperj−centerj|).

This bounds deviation from that specified affine row for the mathematical source map. It does not require assuming the row is an exact derivative. Compare against the saved original body-departure remainder before developing another full propagation candidate. Preserve both the value interval and baseline. A failed improvement target is a useful negative result, not permission to retry.

Finite execution allowance: six identities, six reference setups, six compiled source evaluations, six independent reference calculations, and six exact affine-radius comparisons. Reserve 1,600 packet visits for each native and each independent prefix (19,200 total). Native source calls are charged separately from full-trajectory calls; the latter remain zero. Four workers, one numerical thread each, cooperative wall ceiling 180 seconds, no retries or extensions. Compiler, source, inputs, environment and allowances are frozen before launch.

Passing coefficient checks qualifies this source component only. Native integration and a complete prediction comparison remain separate gates, and observed fitting remains blocked until the broader prediction method qualifies.