test: pin FitEllipse masked-points-loop behaviour

[Jammy2211]

Summary

Pin the current behaviour of FitEllipse.points_from_major_axis_from's 300-iteration mask-rejection loop with five new unit tests in test_autogalaxy/ellipse/test_fit_ellipse.py. Step 3 of 7 in the ellipse_fitting_jax feature decomposition — prompt 6 will rewrite this loop with a JAX-friendly oversample-then-NaN-mask approach; these tests are the regression target it must reproduce.

Each test pins one of the four loop branches plus a multipole variant. Where the loop converges deterministically (zero-masked, under-masked-trim, multipole), the full (N, 2) returned array is pinned at rtol=1e-12. Where the iteration count varies more (over-masked-extra-points), the first and last point are spot-checked at the same tolerance.

Test	Loop branches exercised
zero_masked	equals-branch every iteration (mask is far from ellipse perimeter)
under_masked_trim	extra-points (i=1) → trim (i=2) → equals (i≥3); one perimeter point is masked
over_masked_extra_points	extra-points (i=1, 2) → equals (i≥3); 3×3 mask block straddles perimeter
unreachable_raises	extra-points fires every iteration to i=300, then raises ValueError
with_multipole_under_masked	same path as under_masked_trim with EllipseMultipole(m=4) — pins the inner multipole branch at lines 113-120

All tests numpy-only per PyAutoGalaxy/CLAUDE.md's "Never use JAX in unit tests" rule. JAX parity will be checked in the workspace_test scripts shipped in prompt 2 once prompt 7 lands.

Issue #394.

API Changes

None — test additions only.

Test Plan

• python -m pytest test_autogalaxy/ellipse/test_fit_ellipse.py -v passes (16/16: 11 pre-existing + 5 new)
• Full python -m pytest test_autogalaxy/ passes
• Reference values byte-stable across runs (scipy RegularGridInterpolator is deterministic)

🤖 Generated with Claude Code