Feasible initialization resolves the saved linear failure

All four matched saved-matrix checks pass. The revised calculation no longer worsens either failed case, and the earlier controls remain satisfactory. Four new linear solves were charged in 1.35 seconds; no native calculation, nonlinear fit or new noise sample was used.

Independent convex controls establish the reference answer without aircraft data. The previous method fails ten of 24 such cases. The revision passes all 24, a nonoptimal-start check, a test that rejects misleading solver success and the remaining source-bound checks: 56 in total. Another 90 related software tests pass separately.

The change starts BVLS at a feasible zero step instead of an unconstrained solution. It also rejects a worsening result even when the solver reports success. Every parameter direction, physical bound and scientific acceptance limit stays unchanged. The version-specific SciPy entry point is covered by the frozen numerical environment and requires requalification after an update.

This qualifies the tested linear calculation, not the complete optimizer or its native trajectory predictions. The next gate is a fresh full-optimizer qualification, followed by a separately frozen native trial. The closed calibration still supplies no threshold; validation remains blocked and the 53-stretch aircraft findings are unchanged.

The earlier rejected task envelopes remain preserved. This new comparison used its own finite allowance.

Verified report · Results CSV · Frozen plan · Next analyses.