[antonsoo]

Selecting from an array of observations before computing apparent places currently drops the observer and crashes:

from skyfield.api import load

eph = load('de421.bsp')
t = load.timescale().utc(2025, 1, [1, 2, 3])
position = eph['earth'].at(t).observe(eph['mars'])
print(position[1].apparent().radec()[0])
# Before: AttributeError: 'NoneType' object has no attribute 'xyz'
# After: 08h 17m 38.49s

Keep the ephemeris, corresponding barycentric observer, Earth-relative observer vector and cached light time when indexing an ICRF position. Array observers are indexed with the observation; a scalar observer shared by a star catalogue remains scalar. Barycentric positions retain their self-reference.

This also restores observing from an indexed barycentric position, phase angles after selection, and Earth-occultation checks on selected apparent positions. Regression tests cover integer, slice, repeated/reordered and boolean selections, including a star catalogue with one observer.

The project's test-code.sh passes with the pinned Python 3.8 dependencies, and test-docs.sh and Pyflakes pass. A separate Python 3.12/NumPy 2.5 run retains the existing upstream failures; the added indexing tests pass. Workloads using the shipped JPL ephemeris agree with separately evaluated scalar observations, and complete unselected outputs remain identical on both profiles.

[brandon-rhodes]

I'm busy next week, but will plan to take a look at this in early November. Thanks for the fact that the PR includes a test, that's very helpful.