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

[Jammy2211]

Overview

test__mge_model_from__default_sigma_list_is_bitwise_unchanged and
test__mge_point_model_from__default_sigma_list_is_bitwise_unchanged
(test_autogalaxy/analysis/test_model_util.py, added with #549) assert 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), while the tests build their
expectation with a vectorised 10 ** np.linspace(...). numpy does not guarantee the
scalar and SIMD power loops agree bit for bit, so on AVX-512 hardware they differ by
1 ULP and the tests fail; on GitHub's runners they agree and the tests pass.

The effect is cosmetic, not functional — the PyAutoFit identifier quantises floats at
RESOLUTION = 1e-8 and this drift is ~1e-16 relative, so no identifier moves and no
archived fit is orphaned. But the regression guard added with #549 is not portable:
green in CI, red on an AVX-512 developer machine.

Confirmed reproducible on an AVX-512 host with numpy 2.4.6 — both tests fail, at
mask_radius=3.0, total_gaussians=20 index 18 (the index pytest reports) and at
pixel_scales=0.1, total_gaussians=10 index 4.

Plan

• Keep the exact-equality guarantee. pytest.approx(rel=1e-8) is explicitly ruled out
  by the test docstring and its reasoning still holds — it would only fail once the
  ladder has already moved past the identifier's quantisation.
• Build the expected ladder element by element, the way the implementation does, so
  both sides take the same numpy scalar-power path. The comparison stays bitwise; only
  the code path used to produce the expectation changes.
• Keep the existing docstring reasoning intact and add a note recording the portability
  trap, so a future tidy-up does not re-vectorise the expectation.
• Verify: both tests fail before the change and pass after, on the same AVX-512 host;
  full test_autogalaxy/ suite stays green.

Detailed implementation plan

Work Classification

Library

Affected Repositories

• PyAutoGalaxy (primary) — test-only change, no library source touched

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoGalaxy	claude/pyautogalaxy-mge-sigma-test-3neq07 (at origin/main = 13d3023)	clean

Branch: claude/pyautogalaxy-mge-sigma-test-3neq07 (session-mandated)
Worktree root: n/a — cloud session, worked in the canonical /home/user/PyAutoGalaxy checkout

Implementation Steps

1. test_autogalaxy/analysis/test_model_util.py,
   test__mge_model_from__default_sigma_list_is_bitwise_unchanged (~L169): replace

   assert sigma_list == list(
       10 ** np.linspace(-4, np.log10(mask_radius), total_gaussians)
   )

   with a named log10_sigma_list = np.linspace(-4, np.log10(mask_radius), total_gaussians)
   and an element-wise expectation [10 ** log10_sigma_list[i] for i in range(total_gaussians)].

2. Same transformation in
   test__mge_point_model_from__default_sigma_list_is_bitwise_unchanged (~L220) for the
   np.linspace(-2.0, np.log10(max_sigma), total_gaussians) ladder.

3. Extend both docstrings: keep the existing "why not pytest.approx(rel=1e-8)" reasoning
   verbatim, and add the portability note (scalar vs SIMD power loops are not bit-identical;
   the expectation must be built element-wise).

Key Files

• test_autogalaxy/analysis/test_model_util.py — the two exact-equality guards (only file changed)
• autogalaxy/analysis/model_util.py:190, :271 — the implementation's per-element
  10 ** log10_sigma_list[i], unchanged; it defines the code path the test must mirror

Notes / non-goals

• The two neighbouring pytest.approx(..., 1.0e-8) assertions (L129, L237) are tolerance-based
  by design and are left alone.
• np.log10(1e-4) == -4.0 and np.log10(0.01) == -2.0 exactly, so the test's literal
  endpoints remain a faithful stand-in for the implementation's np.log10(sigma_min).
• No library source changes, so no downstream workspace impact.

Original Prompt

Click to expand starting prompt

# Bug in PyAutoGalaxy: the MGE bitwise sigma-ladder tests fail on

Type: test
Target: PyAutoGalaxy
Repos:
- PyAutoGalaxy
Difficulty: medium
Autonomy: safe
Priority: high
Status: formalised

Bug in PyAutoGalaxy: the MGE bitwise sigma-ladder tests fail on AVX-512 hardware. In PyAutoGalaxy, test__mge_model_from__default_sigma_list_is_bitwise_unchanged and test__mge_point_model_from__default_sigma_list_is_bitwise_unchanged in test_autogalaxy/analysis/test_model_util.py assert exact equality between two different numpy code paths. The PyAutoGalaxy implementation builds each sigma with a per-element scalar power, gaussian.sigma = 10 ** log10_sigma_list[i] at autogalaxy/analysis/model_util.py line 190, while the test builds its expectation with a vectorised 10 ** np.linspace(...). numpy does not guarantee the scalar and SIMD power loops agree bit for bit; on an x86-64-v4 / AVX-512 CPU they differ by 1 ULP at index 18, exactly the index pytest reports. Reproduced in isolation with numpy 2.4.6. The effect is cosmetic, not functional: the run identifier quantizes at RESOLUTION 1e-8 and this drift is ~1e-16 relative, so no identifier moves and no archived fit is orphaned. But the tests are green on GitHub runners and red on AVX-512 developer machines, so the regression guard added with PyAutoGalaxy#549 is not portable. Fix in PyAutoGalaxy by building the expected ladder element by element the way the implementation does, or by comparing with a tolerance far below the 1e-8 identifier resolution.

Session context added by the human at /start_dev:

Do NOT weaken this into pytest.approx(rel=1e-8) — the test docstring rules that out
explicitly and gives the reason. The guarantee worth keeping is that the ladder stays
stable well inside the identifier's RESOLUTION = 1e-8 quantisation. Preferred fix is to
build the expected ladder element-wise so both sides take the same numpy path; keep the
docstring's reasoning intact and note the portability trap in it.

[Jammy2211]

Fixed on branch claude/pyautogalaxy-mge-sigma-test-3neq07 (91eb878). No PR opened — not requested.

Three things worth recording beyond the plan:

The session host is itself AVX-512, so this was reproduced rather than assumed — pytest failed on the unmodified tree with numpy 2.4.6 and passes after. The footprint is wider than the report: the point-model test drifts too (pixel_scales=0.1, total_gaussians=10 at indices 4 and 9; 0.05, 5 at index 3), and mge_model_from also drifts at mask_radius=3.5, total_gaussians=30 index 8, not only the reported 3.0, 20 index 18. Both guards were broken.

The guarantee was control-tested, not assumed. An element-wise expectation could in principle have become vacuous — comparing the implementation against a restatement of itself. Perturbing the implementation defaults by a relative 1e-7 (sigma_min 1e-4 → 1.0000001e-4, 0.01 → 0.010000001) still fails both tests, so the exactness guarantee survives the change. The implementation was restored afterwards; only the test file is modified.

pytest.approx(rel=1e-8) was not used, per the docstring's existing reasoning, which is kept verbatim. Both docstrings now carry a PORTABILITY TRAP note explaining why the expectation must stay element-wise, so a later tidy-up does not re-vectorise it.

Validation: test_model_util.py 29 passed; full test_autogalaxy/ 1004 passed, 3 skipped. Run on Python 3.11 (the only interpreter with a stack in this sandbox) — CI grades 3.12/3.13, and the change is pure-Python test code with no version-sensitive surface. Test-only, so no downstream workspace impact; PyAutoLens has no equivalent exact-equality 10 ** np.linspace assertion.

---

Generated by Claude Code