Make Instrument indexable via the na.Indexable mixin

[jacobdparker]

Summary

Adds a reusable way to reduce a multi-channel esis.optics.Instrument to a single channel.

• esis.optics.dataclass_select(instance, item) — a recursive helper that indexes every named_arrays.AbstractArray field of a dataclass along a named axis. It recurses into nested dataclasses and into dict/list/tuple containers (e.g. the ruling spacing coefficients dict), and leaves arrays that don't carry the indexed axis untouched (some named-array types such as FunctionArray raise on a missing axis, so the index is filtered to axes each array actually has).
• AbstractInstrument.select_channel(channel) — a copy-safe method that deep-copies the instrument, drops any cached system, and uses dataclass_select to eliminate the channel axis.

This generalizes the manual per-channel indexing in esis.flights.f1.optics.design_single and the ad-hoc copy.deepcopy loop used to model/fit individual channels, and is broadly useful for single-channel modeling, inversion, and diagnostics.

The pattern follows dataclass_select from the sibling multi-slit-solar-explorer repo, extended to handle dict/list/tuple containers and the missing-axis case.

Tests

esis/optics/_select_test.py:

• a fast synthetic test of dataclass_select (axis reduction, missing-axis pass-through, dict/nested recursion, non-array fields untouched);
• a parametrized select_channel test over all four channels covering the dict-stored ruling coefficient, copy-safety (original left unmodified), channel-axis removal, and that the reduced instrument still resolves into an optical system.

Plus test_select_channel added to AbstractTestAbstractInstrument.

pytest esis/optics passes (191 tests); ruff and black clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

Documentation preview available at https://esis-mission.github.io/esis-pr-previews/61-feature/select-channel
Note that it might take a couple seconds for the update to show up after the preview_build workflow has completed.

[roytsmart]

Cross-referencing a parallel approach to the same problem: sun-data/named-arrays#180 adds a generic na.Indexable mixin whose __getitem__ delegates to a recursive na.getitem() (and shape to a recursive na.shape()), indexing every named-array leaf of a dataclass / dict / list / tuple. It overlaps heavily with dataclass_select here.

One important behavioral difference worth reconciling: dataclass_select filters the index per-leaf to only the axes each array actually has, so a leaf that lacks the indexed axis is left untouched. na.getitem does not do this yet — it passes the full index to every leaf. That's fine for ScalarArray (a missing axis is silently ignored), but FunctionArray raises ValueError: item contains axes {...} that does not exist in {...}. So as-is, na.getitem/Indexable would crash on exactly the missing-axis / FunctionArray case this PR handles.

Worth deciding whether dataclass_select should eventually wrap/defer to na.getitem once that per-leaf axis filtering is upstreamed, to avoid two implementations diverging.

[roytsmart]

Follow-up: the FunctionArray missing-axis regression I mentioned above is now fixed upstream in sun-data/named-arrays#181. FunctionArray.__getitem__ now filters out axes it doesn't have (matching ScalarArray) instead of raising, so indexing a function array by an axis it lacks is a no-op.

With that in, na.getitem handles the missing-axis case on its own, so dataclass_select could eventually drop its per-leaf axis filtering and just defer to na.getitem once #181 lands and named-arrays is bumped here.

[jacobdparker]

Reworked per @roytsmart's review: instead of the standalone dataclass_select helper, AbstractInstrument now inherits na.Indexable (named-arrays #180), so it gets a recursive named __getitem__/isel/shape for free, and select_channel(channel) is now just a thin wrapper over self[{self.axis_channel: channel}].

The per-leaf missing-axis filtering the old helper did is no longer needed thanks to #181 (FunctionArray now ignores missing axes), so _select.py is removed entirely. This also makes the capability reusable for any instrument (f1 + f2) and other composite objects.

Since this needs na.Indexable, the dependency pin is bumped to named-arrays~=1.6 (v1.6.0 shipped today and includes #180/#181).

Net diff is now just _instruments.py + its tests + the pin bump.

[jacobdparker]

Per discussion, dropped select_channel entirely rather than keeping it as a wrapper. Since AbstractInstrument now inherits na.Indexable, a channel is selected generically with instrument.isel(channel=i) (or instrument[{instrument.axis_channel: i}]) — a bespoke per-axis method doesn't scale for a vectorized model.

The ESIS-side change is now a single line (inherit na.Indexable) plus a test_isel verifying that selecting a channel via isel returns a model that still resolves into a valid optical system. Pin remains named-arrays~=1.6.