Added support for using the object surface as the field stop of a SequentialSystem

[jacobdparker]

Second PR in the stop-solver stack — stacked on #168 (the diff shown here is relative to that branch).

Problem

Marking the object surface (with an angular, dimensionless aperture) as the field stop is the natural configuration for dispersive systems — at a single wavelength of a dispersed spectrum, most of the sensor outline is unreachable, so targeting the sensor wire is ill-posed. But this configuration was broken in three independent ways:

1. The position-variable branch of _calc_rayfunction_stops_only called sag() with a 2-component vector → TypeError before the solve started.
2. The solved outputs were never broadcast over both stop axes → axis error in the field_min/max / pupil_min/max reductions.
3. The hardcoded initial guess (rays at the first stop's origin, direction +z) produces NaN first residuals whenever the next surface is far off that axis (e.g. the feed mirror of a Rowland-circle spectrograph, 174 mm off-axis in FURST) or on steep grazing flanks — and Newton cannot recover from NaN.

Changes

• zfunc builds a proper 3-D vector before calling sag().
• _calc_rayfunction_stops broadcasts position and direction against each other before the stop-axis reductions.
• New _anchor_surface helper: the seed now aims each ray at its own target point on the last stop surface when no powered surface lies between the stops (nearly exact — the grazing solve converges in ~5 iterations), and otherwise at the center of the first powered surface.

Tests

Adds TestSequentialSystemGrazingSpectrograph: a grazing-incidence paraboloid + transmission-grating spectrograph whose source is the field stop, asserting the recovered field equals the source's angular radius (0.25°) to 1e-6 deg. Also verified against the FURST Spectrograph model from Kankelborg-Group/furst-optics#25, which recovers field_max = ±0.26656° (the sunpy solar angular radius).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.50%. Comparing base (e9a4eb5) to head (9edd7b6).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #169      +/-   ##
==========================================
+ Coverage   99.34%   99.50%   +0.15%     
==========================================
  Files         116      116              
  Lines        5974     6042      +68     
==========================================
+ Hits         5935     6012      +77     
+ Misses         39       30       -9     

Flag	Coverage Δ
unittests	99.50% <100.00%> (+0.15%)	⬆️

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.

[roytsmart]

It's not enough for there to be a sag, the material must also have a different index of refraction

[jacobdparker]

You're right — thanks, this was a real bug in the helper's reasoning, not just the test.

I'd been treating the surface's own material as deciding whether it refracts, but the index step is what matters: in Surface.propagate_rays, n1 = rays.index_refraction (carried from upstream) and n2 = material.index_refraction(rays), and Snell's law only bends the ray when n2 != n1. So a curved surface has refractive power exactly when there's an index discontinuity across it — which depends on the previous surface's medium, not this one. The exit face of a glass element makes the point: it's a curved Vacuum-material surface (n2 = 1) that still refracts because the ray arrives with n1 = 1.5. So sag is not None is necessary but not sufficient.

One wrinkle: as far as I can tell no core material currently returns a transmitted index != 1 — Vacuum -> 1, and the mirrors/multilayers/filters all return rays.index_refraction (passthrough) — so n never leaves 1 for a transmitted ray and this branch never legitimately fires today. To make _anchor_surface correct in principle I'd thread the running index through the subsystem and treat a curved surface as powered iff n2 != n1 (which also needs a small refractive material to cover the branch — I'm adding a Sellmeier Glass material + a Cooke-triplet tutorial separately, which would give a real one).

Before I plumb that in: would you prefer the full n2 != n1 tracking now, or a lighter guard (drop the standalone sag heuristic and rely on mirror/rulings until there's an actual refractive material)? Happy to do either.