Gauss triplet selection is blind to close encounters, where the method cannot work

Summary

_build_sequence splits observations into segments on observation-time gaps alone
(orbitfit.py:677, sep_dt=90.0) and takes the longest-span chunk as seq[0]. gauss_iod
then uses that chunk's first, middle and last observation (iod.py:102-104). Nothing in the
selection knows any dynamics.

Gauss's method assumes Keplerian heliocentric motion across the triplet. If the span contains a
close planetary encounter that assumption fails outright: the pre- and post-encounter
heliocentric orbits are different orbits, so no root satisfies all three observations. The
method returns nothing, or returns a spurious root which is then integrated through the
encounter — which is the grind apply_ias15_min_dt's comment describes.

The selection makes this likely rather than rare. Observing around a close approach is
continuous — inbound, closest approach, outbound — so those observations contain no 90-day
gap and form a single chunk, and being densely sampled they are a good candidate for seq[0].

Where

• src/layup/orbitfit.py:677 — _build_sequence, splits on gaps > sep_dt, picks longest span
• src/layup/iod.py:102-104 — gauss_iod, first/middle/last of seq[0]

Observed

(367943) Duende passes 34,053 km from Earth on 2013-02-15 (0.09 lunar distances) and its
optical arc runs 2012-02-23 to 2013-02-21, so any triplet drawn from that span straddles the
encounter. It is one of two numbered objects the full-catalogue re-fit could not fit; the
other, (308635) 2005 YU55 at 0.85 LD, now converges.

Suggested

Exclude a segment containing a close approach from seq[0], or draw the triplet from one side
of it. Layup already loads the ephemeris needed to know where close approaches are.

Related: #465, #486.

Start with _build_sequence in src/layup/orbitfit.py:677 and trace how the loaded ephemeris is available during sequence selection. Then inspect gauss_iod in src/layup/iod.py:102-104 and the related issues #465 and #486. Done means triplets containing a close encounter are excluded or selected from one side, allowing the Duende refit to converge without the reported spurious root or integration grind.