Trace once for the distortion, vignetting, and sensor angle

[roytsmart]

Depends on #211 and must merge after it. Opened against that branch so the diff shows only this change; retargeted to main immediately, so deleting #211's branch on merge cannot close this PR.

Why

linearize fits the distortion and the vignetting from the same rays, traced with identical arguments, and then traced a third time through rayfunction_default for the cosine of the angle at which light strikes the sensor. None of the three reads the intensity.

What changes

distortion and vignetting keep their behaviour; their fitting bodies move into _fit_distortion and _fit_vignetting, so a caller needing both can trace once and fit twice. direction comes from those same rays.

before: 4 raytraces, 4 stop solves, 3.46 s
after:  2 raytraces, 1 stop solve,  2.64 s

with `num_interpolation` set:      1.66 s -> 1.06 s

What moves, and what does not

distortion      20  IDENTICAL
vignetting      10  IDENTICAL
scene            3  IDENTICAL
sensor           2  IDENTICAL
illumination     1  IDENTICAL
direction        1  CHANGED by 2.31%

direction was evaluated on grid_input.wavelength — a single value, 629.77 Å for ESIS — no matter which band was asked for. It is now averaged over the band being linearized. The old value was worse the further the requested band sat from that wavelength:

today (on grid_input.wavelength):  0.880400+0.044822j

if computed on the grid actually being linearized:
   580-640 A: differs by  5.36%
   300-350 A: differs by 12.27%
   900-1000 A: differs by 22.73%

It feeds sensor.expose, which uses it for the path length through the depletion region, so this is a radiometric correction rather than a refactor artifact.

The trap this had to avoid

The existing averaging line reads:

axis_grid = self.axis_wavelength_ + self.axis_field_ + self.axis_pupil_
direction = direction.mean(axis=tuple(ax for ax in axis_grid if ax in na.shape(direction)))

axis_wavelength_ is derived from grid_input, which holds a scalar wavelength for ESIS, so it is (). That was harmless while direction came from rayfunction_default, which has no wavelength axis either. Fed the shared trace, the axis survives:

shared trace, existing line  shape {'wavelength': 9}      <- the landmine
shared trace, wavelength too shape {}                     0.895272+0.030887j

An array direction is not merely different, it is broken. expose indexes it by the cell centers of the scene's wavelength grid, which has nothing to do with the grid being linearized:

   scene 10 edges ( 9 centres) -> NO ERROR              <- silently pairs unrelated wavelengths
   scene  5 edges ( 4 centres) -> ValueError: shapes ({'wavelength': 4}, {'wavelength': 9}) are not compatible
   scene 20 edges (19 centres) -> ValueError: shapes ({'wavelength': 19}, {'wavelength': 9}) are not compatible

So the average now names the wavelength axis explicitly, and test_linearize asserts direction is a scalar.

Worth knowing while reviewing: _linear_test's sensor is an ideal material that ignores direction entirely (9000 photons in, 9000 electrons out, and the mismatched-shape case does not even raise), so none of LinearSystem's own tests exercise it.

Tests

Full suite: 17768 passed, 304 skipped, 17 xfailed. _sequential.py at 100% line coverage. Six ndfilters tests fail with a numba ReferenceError: underlying object has vanished, identically on main — a pre-existing local flake, not from this branch.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (772540a) to head (8a7083a).

Additional details and impacted files

@@            Coverage Diff            @@
##              main      #212   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files          121       121           
  Lines         7402      7408    +6     
=========================================
+ Hits          7402      7408    +6     

Flag	Coverage Δ
unittests	100.00% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.