Relate solver-in-the-loop results to Jacobian conditioning on the Kolmogorov setup #277

The solver-in-the-loop benchmark (#274) reports, per solver and unroll K, the error gap each training arm closes and the value of the solver gradient (the stopped-gradient arm's error over the full arm's). It is an open question whether those numbers follow from properties the gradient suite already measures, such as the singular-value spectrum of the solver's Jacobian or how FD agreement degrades with horizon.

Answering it needs gradient-suite runs on the Kolmogorov configuration (64², dt = 0.01, forcing and drag, horizons of 1, 4 and 16 correction intervals) next to the campaign results. Each SIL run already records:

• the training-loss FD error at zero and nonzero readout;
• the gradient norm at the identity model for the stopped and full arms, per K.

Those are a starting point.

With four eligible solvers, any correlation is suggestive at best. The aim would be to see whether a solver's gradient-suite profile anticipates its SIL behaviour, not to fit a model.