test: make the MGE bitwise sigma-ladder guards portable across CPUs

[Jammy2211]

Closes #550.

What changed

test__mge_model_from__default_sigma_list_is_bitwise_unchanged and
test__mge_point_model_from__default_sigma_list_is_bitwise_unchanged (added with #549)
asserted exact equality between two different numpy code paths. The implementation
builds each sigma with a per-element scalar power
(gaussian.sigma = 10 ** log10_sigma_list[i], autogalaxy/analysis/model_util.py:190
and :271); the tests built their expectation with a vectorised 10 ** np.linspace(...).
numpy does not guarantee its scalar and SIMD power loops agree bit for bit, so the two
disagree by 1 ULP on AVX-512 hardware — green on GitHub's runners, red on an AVX-512
developer machine. The regression guard was not portable.

This builds the expected ladder element by element, so both sides take the same
numpy path. The comparison stays bitwise; only the code path producing the expectation
changes.

Test-only. No library source is touched.

What was deliberately not done

pytest.approx(rel=1e-8) is still not used. The test docstring rules it out and the
reasoning holds — it only fails once the ladder has moved by a relative ~1e-7, already
past the point where the PyAutoFit identifier (RESOLUTION = 1e-8) changes. That
reasoning is kept verbatim; both docstrings now additionally carry a PORTABILITY TRAP
note explaining why the expectation must stay element-wise, so a later tidy-up does not
re-vectorise it.

Verification

The host used for this change is itself AVX-512 (avx512f/bw/cd/dq/vl/vnni) with numpy
2.4.6, so the failure was reproduced, not assumed — both tests fail on the unmodified
tree and pass after.

The reported footprint was narrower than the real one. Drift occurs at:

test	case	index
mge_model_from	mask_radius=3.0, total_gaussians=20	18 (the reported one)
mge_model_from	mask_radius=3.5, total_gaussians=30	8
mge_point_model_from	pixel_scales=0.1, total_gaussians=10	4, 9
mge_point_model_from	pixel_scales=0.05, total_gaussians=5	3

Both guards were broken, not just the first.

Control test — an element-wise expectation risks becoming vacuous (the test merely
restating the implementation), so this was checked rather than assumed: perturbing the
implementation defaults by a relative 1e-7 (sigma_min 1e-4 → 1.0000001e-4,
0.01 → 0.010000001) still fails both tests. The exactness guarantee survives the
change. The implementation was restored afterwards; only the test file is modified here.

Also verified: np.log10(1e-4) == -4.0 and np.log10(0.01) == -2.0 exactly, so the
tests' literal endpoints remain a faithful stand-in for the implementation's
np.log10(sigma_min).

Suites: test_autogalaxy/analysis/test_model_util.py 29 passed; full
test_autogalaxy/ 1004 passed, 3 skipped. Run locally on Python 3.11 (the only
interpreter with a stack in that sandbox) — CI grades 3.12/3.13, and the change is
pure-Python test code with no version-sensitive surface.

Scope

Test-only, so no downstream workspace impact and no pending-release gate. The
neighbouring pytest.approx(..., 1.0e-8) assertions (L129, L237) are tolerance-based by
design and are untouched. PyAutoLens has no equivalent exact-equality
10 ** np.linspace assertion.

🤖 Generated with Claude Code

https://claude.ai/code/session_011tJFsEesnF7rZmn2xvfxUe

---

Generated by Claude Code