Scale step limits by a trust-region factor that halves on divergence
(RMS increases > 1%) and relaxes toward full step on good convergence
(RMS decreases > 10%). Prevents oscillation with poor initial guesses
without affecting well-seeded fits. Also stores SVD condition number
for diagnostic use in upcoming covariance output.
Existing 8 OD regression tests + 67 standalone math tests unaffected
(adaptive_factor starts at 1.0, round-trip tests never trigger
divergence).