Follow-up to #678, which fixed the tilted-stop, back-surface-stop and folded cases but left the offset-stop case out. That case is a real solver bug, not the system-definition inconsistency the earlier file suggested.

The bug

The entrance pupil is the image of the stop, so a decentered stop has a decentered entrance pupil. RealReferenceStrategy measured the stop radius by launching a probe at y = EPD/2, on the axis, and reporting where it landed relative to the stop centre. When the pupil is off axis that probe never goes near it, so the reported radius is unrelated to the stop.

On the system reported in the follow-up comment (a small R=5 lens, r_max 1.0, decentered to y=+5 behind a singlet):

launch y -> local stop (x, y)      [R=5 sphere: |local r| must be < 5]
  y= 0.000 -> (nan, nan)             the on-axis probe misses the stop entirely
  y= 1.131 -> ( 0.0000, -4.0392)     EPD/2 probe: 4 mm from the stop centre
  y= 5.673 -> ( 0.0000,  0.0000)     the actual pupil centre
  y= 6.804 -> ( 0.0000,  0.9773)     the actual pupil edge

So r_stop came out 4.039 against a true radius of ~1.0. The aimer then targeted the full pupil four times too wide: the outer rays missed the stop surface (NaN) and the rest were vignetted by its physical aperture, leaving only the chief ray. That is the degenerate spot in the report.

Note the aperture is self-consistent here: EPD 2.2622 puts the paraxial marginal ray at exactly 1.0, which is the stop's r_max. The system is correctly defined, unlike the earlier file (EPD 20 against a radius-5 stop) that #678 set aside.

The fix

Solve for the pupil centre first, the axial ray landing on the stop's own origin, which needs no stop radius, then offset EPD/2 from there. The solve reuses IterativeRayAimer._solve_core against target (0, 0) and is seeded with the first-order pupil decenter (stop decenter times pupil magnification), which predicts 5.6555 against the true 5.6727 and converges immediately.

For a centered system the seed is zero and the solve is a no-op, so the measured radius is unchanged by construction.

After the fix

Aimed rays land at local stop y [-0.977, -0.489, 0.000, 0.489, 0.977]: linear in Py, centred on the stop to 1e-9, all inside r_max 1.0. The full disk pupil traces unvignetted.

r_stop is bit-identical on every system that already worked:

No system falls back to the paraxial path, so the real-ray measurement is exercised throughout.

Tests

Two regression tests in tests/test_ray_aiming.py, on both backends: the real-ray radius must match the paraxial reference for a decentered stop, and every pupil ray must aim inside the stop's clear aperture and reach the image unvignetted. Both fail on master and pass here. The existing ray-aiming, initialization, cached-aimer and image-simulation suites are unchanged (226 passing).

Refs #654