Qualify the known-calibration benchmark construction

Test both input-error laws on a generated two-second, 600-packet record. Use nine query epochs from 0.1 to 1.7 seconds, with six earlier positions and three future physical-position targets. Supply the true initial state and known sensor bias/gain, antenna lever arm and latency; estimate no parameters.

The native streaming mean must agree with the already checked wide tangent calculation within 0.00001 m per coordinate on this short control. Compare it with the independently generated nominal trajectory as a separate check, at a fixed 0.01 m maximum 3D error. Compare conditional physical second moments with a stable, independent dense calculation using the complete innovation factor. Verify that receiver observation loadings stop at the earlier boundary.

One numerical thread, two compiled libraries and a three-minute cooperative allowance are permitted. Each case allows 700 native calls, 2,000 packet steps and 10,000 packet/query visits. Totals are 1,400 calls, 4,000 steps and 20,000 visits. Construction is bounded at 610 generated samples and 600 constructed intervals per case. No noise draws, optimizer starts, original recordings or protected groups are used. Unused native prepayment remains charged.

This qualifies benchmark construction, not real-world error amplitudes, nonlinear coverage or the longer four-setting execution. That execution still requires its own source/environment freeze and finite allowance.