Default the IAS15 integrator to the modern (mode 2) adaptive step controller

[matthewholman]

Summary

Change the default of the IAS15 adaptive step-size controller from the legacy controller (mode 1) to the newer controller of Pham, Rein & Spiegel (2024) (mode 2). Concretely this flips a single global default in orbit_fit.cpp (g_ias15_adaptive_mode, -1 -> 2) and updates its comment. The knob itself and its per-simulation application were added in #324 and are unchanged.

Motivation

assist_attach() sets IAS15 to the legacy adaptive controller, which drives the integration timestep toward zero during close encounters with the Earth. A single such integration can then fail to return, hanging the orbit fit and consuming the wall-clock budget. This was the entire "timeout / hang" class encountered while re-fitting the full Minor Planet Center (MPC) catalog: close-approaching near-Earth objects (NEOs) that never completed under the legacy controller.

The mode-2 controller steps through these encounters efficiently. Crucially, correctness is independent of the controller -- the step controller only decides whether and how quickly an integration completes, never the resulting orbit. In the full-catalog run, the 40 hardest close-NEO cases (perihelion 0.24-1.02 AU) that hung under the legacy controller all completed in under 0.5 s under mode 2, with fitted states identical (to numerical tolerance) across controllers. Hanno Rein suggested this controller in the #324 review as a cleaner alternative to the minimum-timestep floor.

Change

• src/lib/orbit_fit/orbit_fit.cpp: g_ias15_adaptive_mode default -1 -> 2 (+ comment).
• No API change. set_ias15_adaptive_mode() / get_ias15_adaptive_mode() still let a caller override (e.g. -1 to restore the legacy behavior, 1 to force it explicitly).
• The fit path, both predict paths, and the IOD picker already call apply_ias15_adaptive_mode() after assist_attach, so all of them pick up the new default.

Tests

• test_ias15_settings.py: updated the default-value test to assert the new default (2).
• Added test_fit_is_mode_independent (guarded by requires_ephem): a well-behaved object fit under the legacy controller (mode 1) and the modern controller (mode 2) returns the same state -- verified locally that the two fits agree bit-for-bit.

Scope / caveats

• This addresses the integrator hang class only. It does not fix the separate 2020 CD3 case, whose fit fails from a numerical breakdown in the observation model / variational partials during a sub-0.0003 AU topocentric pass -- a distinct, lower-priority item.

Related

• Builds on Pluggable IOD layer, multi-root picker, residual prefilter, and IAS15 min-dt floor #324 (introduced the adaptive-mode knob and its per-simulation application).

[kjnapier]

looks good