Damped trust-region solver qualification

The completed 200-evaluation continuation left all three target comparisons unfinished. Saved histories show many rejected corrections, sensitivity to starting points and, in one case, a stalled primary despite unused evaluations. This separate candidate changes how the optimizer proposes and accepts a step.

It limits each correction to a trust region, adds scaled damping in weak directions, and checks agreement between predicted and actual objective improvement. Physical models, calibration bounds, receiver weights and integration clock are unchanged. The final acceptance test uses the existing undamped correction and fresh stationarity, objective and actual-position checks. Small damped steps cannot establish convergence by themselves.

Prepared; qualification has not yet run. The finite plan defines 48 checks, with zero NASA-data fits. A metadata-only waiter starts the checks after the active nine-stretch sequence finishes and releases its four-thread allocation. It does not pause, restart or change that active study. Test execution is charged once; interrupted or failed qualification is not retried automatically.

Passing these controls would justify a separately frozen matched archival trial. It would not establish statistical significance, a global optimum or the cause of the background field. The current public finding remains 41 globe preferences, zero flat and 12 unresolved pending audited results from the active recovery.