Herget's convergence tolerance is ~30x tighter than useful, costing seeds on near-Earth arcs

herget_iod passes tolerance=0.0001, so herget_with_assist accepts no seed until the range step falls below 1e-4 au (15,000 km). An initial-orbit seed only has to land in the right basin -- the differential correction refines it -- so objects that already have a usable seed instead exhaust 100 iterations on each of the three rho rungs and return nothing.

Measured on 163 numbered NEOs, scoring the raw seed's semimajor axis against JPL ("good" = within 10%):

Total seeds rise monotonically, but usable seeds peak near 1e-2 and collapse at 1e-1, where unconverged seeds start being accepted.

Suggested value: 3e-3 -- the loosest threshold with no regressions, and the fraction within 10% is unchanged (50% vs 49%). The main belt is unaffected: tested at the more aggressive 1e-2, seeds went 156 -> 158 and good seeds 154 -> 155. Main-belt seeds are accurate to ~0.2% anyway, so the threshold rarely binds there; this is a near-Earth-regime effect.

One-line change. Note this is a parameter herget_iod passes, not a change to herget_with_assist itself.

Against #554, where this code currently lives.

Start at the herget_iod entry point and trace the tolerance passed to herget_with_assist; issue #554 identifies where this code currently lives. Change the passed threshold to 3e-3, then verify that near-Earth seed counts improve without accepting the unconverged seeds described at 1e-1, while main-belt results remain unaffected.