Local stability qualification for projected receiver fits

Reason for a separate candidate

Removing coordinates creates weakly constrained calibration directions. The previous trust candidate can fit the retained coordinates closely while its final undamped probe proposes a nearly full-range move in weak parameters. Such a move can leave the region where the linearization predicts changes. The completed/active prior studies and their acceptance labels stay unchanged.

This candidate explicitly changes the probe region, not the observations, physical parameter bounds or numerical tolerances. It tests local numerical stability; it cannot certify a global optimum, unique calibration or absence of unscored-axis errors. Paired-start prediction checks remain mandatory.

Defined local gate

Retain all nuisance columns and require the original projected stationarity limit 0.0001 before probing. Solve the bounded linear correction within the original feasible box intersected with a radius of 0.001 per normalized parameter. This radius matches the larger derivative-check step used in the endpoint diagnostic; it is a declared numerical neighborhood, not a physical calibration allowance or sensor-accuracy claim.

Evaluate that correction freshly using native increments. Require the unchanged relative cost-change limit 1e-8, fresh stationarity limit 0.0001, retained coordinate RMS change 1 mm, and maximum change 1 cm. Omitted axes are explicitly unchecked. Damping or a tiny optimization step cannot replace these checks. Preserve the complete Jacobian and numerical rank.

Sequence and finite budget

1. Pass constructed tests covering null directions, active bounds, the original stationarity failure, a nonlinear probe that must fail and a known optimum.
2. Wait for the preceding full trust qualification to finish; inspect every saved outcome and paired-start result before freezing this separate batch.
3. If still warranted, freeze and run all 24 generated-motion cases: three models × nominal/variable intervals × horizontal/height objectives × zero/opposed starts. No retries. Cap each fit at 160 native tangent calls plus two diagnostic predictions; 3,888 native calls maximum, four total threads and 30 minutes. No originals or observed-data fits.
4. Require every fit and all twelve paired-start checks to pass under the same retained-error/objective/repeatability criteria as the preceding study.
5. Continue to full-calibration/active-bound controls only if that gate passes. A failed candidate blocks observed refits. Diagnose saved artifacts without reusing a completed allowance or relabeling prior failures.

Freeze exact sources, tests, environment, parents, identities and allowance; use exclusive locks, pre-call journals and atomic hashed outputs. Preserve every failed start. The existing 53 globe / 0 flat / 0 unresolved comparisons are unaffected by software qualification or this changed local probe.