Short native bounded-uncertainty controls

This finite software test follows the separately qualified native-interface checks. It evaluates supplied calibration points; it performs no optimization. It does not repeat the completed two-minute fitting study.

The matrix has 24 cases: both input-error laws, three candidates, both processing assumptions, and constant versus height-dependent gravity. Each uses a generated 160-packet record with 12 receiver query times. Only the first ten receiver positions enter initialization and error conditioning. A supplied gyro-offset parameter is exactly at a bound. All 26/27 mean columns remain present.

Compare the returned bounded envelope with a separate dense calculation from the same fixed covariance. Verify its conditional mean, every parameter contrast, shared-error response norms and error allocation. Preserve the original unrestricted diagnostic's boundary abstention. No future receiver observations are passed to the provider.

The zero mean-remainder allowance here tests the linearized mathematical problem only. This does not bound the nonlinear native map or establish physical truth coverage. A second calculation with missing nonlinear allowances must explicitly remain unsupported. Do not refit any failed case.

Before compilation, freeze all sources and the environment in a fresh request. Use the shared prepaid test fixtures with one numerical thread and a three-minute cooperative wall limit. Allow two compiled libraries in total. Each case permits at most 450 native calls, 15,000 packet steps and 300,000 packet/query visits. Totals are 10,800 calls, 360,000 steps and 7,200,000 visits; unused prepayment is not returned. There are zero optimizer starts, noise draws, original-recording reads and protected-data reads. All failed checks and journals are retained.