fix: drop stale astropy <=7.2.0 cap + stale cosmology docstrings

[Jammy2211]

Overview

astropy is capped at >=5.0,<=7.2.0 in PyAutoArray and PyAutoGalaxy, so a fresh pip install downgrades users to astropy 7.2.0 even though the dev stack runs 8.0.1 and passes. Since the cap was filed, the stack's dependency convention moved to floors-not-pins (PyAutoArray#433 / PyAutoLens#687); PyAutoFit and PyAutoNerves already declare astropy>=5.0 uncapped. Added scope: autogalaxy/cosmology/model.py was reimplemented natively (JAX-compatible, no astropy import) but two docstrings still claim to wrap astropy functions — corrected in the same pass.

Plan

• Drop the astropy cap in PyAutoArray and PyAutoGalaxy pyproject.toml, leaving the floor astropy>=5.0 (matches the PyAutoFit / PyAutoNerves convention).
• Fix the stale astropy claims in the cosmology docstrings; keep the accurate parity statements.
• Verify a clean-venv resolve pulls astropy 8.x, run both library test suites, smoke an imaging fit that exercises FITS/WCS.
• Ship library-first: PyAutoArray, then PyAutoGalaxy.

Detailed implementation plan

Affected Repositories

• PyAutoArray (primary)
• PyAutoGalaxy

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoArray	main	clean
./PyAutoGalaxy	main	clean

Suggested branch: feature/astropy-cap-bump
Worktree root: ~/Code/PyAutoLabs-wt/astropy-cap-bump/ · Work Classification: Library
Parallel-claim note: PyAutoGalaxy is also claimed by active task yang24-sidm-gravothermal-profile; human approved parallel run 2026-08-06 (file overlap nil: pyproject + cosmology docstrings vs new dark-mass profile).
Sizing note: Brain scored 13/too-large with a 4-phase split — prose-driven artifact; actual scope is 2 repos / 3 files / no API change. Kept as one task.

Implementation Steps

1. PyAutoArray/pyproject.toml line 32: "astropy>=5.0,<=7.2.0" → "astropy>=5.0".
2. PyAutoGalaxy/pyproject.toml line 34: same edit (keep the two specifiers identical).
3. autogalaxy/cosmology/model.py docstring fixes:
   • angular_diameter_distance_to_earth_in_kpc_from (~line 126) and angular_diameter_distance_between_redshifts_in_kpc_from (~line 142): remove "returns only the value of the astropy function it wraps" — state the native implementation returns a plain value in kpc.
   • critical_density (~lines 215-221): replace the conversational porting note with a direct statement: native xp (NumPy/JAX) implementation returning Msun/kpc^3, value-matching astropy.cosmology.FLRW.critical_density(z) after unit conversion.
   • KEEP accurate parity statements (lines ~42, 61, 102, 454, 517, 756) — they document a real compatibility contract.
4. Tests: pytest test_autoarray/, pytest test_autogalaxy/ (docstring-only source change; suites must pass).
5. Clean-venv verify (fresh venv, PYTHONPATH cleared): resolve pulls astropy 8.x, not 7.2.0.
6. Smoke: one imaging-fit script loading FITS/WCS under astropy 8.0.1.

Key Files

• PyAutoArray/pyproject.toml — dependency floor
• PyAutoGalaxy/pyproject.toml — dependency floor
• PyAutoGalaxy/autogalaxy/cosmology/model.py — stale docstrings

Original Prompt

Click to expand starting prompt

Bump astropy cap to allow 8.x (drop the stale <=7.2.0)

Type: maintenance
Target: PyAutoArray
Repos:

• PyAutoArray
• PyAutoGalaxy
  Difficulty: small
  Autonomy: supervised
  Priority: normal
  Status: formalised

Filed 2026-07-12 from a /hygiene deps audit.

Why

astropy is capped at >=5.0,<=7.2.0 in both PyAutoArray and
PyAutoGalaxy pyproject.toml. The cap is demonstrably too tight: the
development venv already runs astropy 8.0.1 with the full stack passing, and
PyPI latest is 8.0.1. A fresh pip install today would downgrade astropy to
7.2.0 because of the cap. This is the clearest stale cap in the audit — the
installed environment is the proof it works.

Scope

• Raise the astropy cap in PyAutoArray/pyproject.toml and
  PyAutoGalaxy/pyproject.toml. Convention update (2026-08-06): since this was
  filed, the stack moved to floors, not pins (PyAutoArray#433 /
  PyAutoLens#687 dep-floors work; PyAutoFit and PyAutoNerves already declare
  astropy>=5.0 uncapped) — so drop the cap entirely: astropy>=5.0.
• Keep the two repos' astropy specifiers consistent (they match today; the
  audit found no cross-repo conflict — don't introduce one).
• No source changes expected; astropy 8 already imports and runs in the venv.
• Added scope (human request 2026-08-06): stale astropy docstrings in
  autogalaxy/cosmology/model.py. The cosmology was reimplemented natively
  (JAX-compatible) and no longer imports astropy, but two docstrings still say
  "returns only the value of the astropy function it wraps" (lines ~126, ~142)
  and critical_density carries a conversational porting note ("Astropy
  returns g/cm^3, but in AutoLens you were immediately converting…",
  lines ~215-221). Rewrite these to state the native implementation and units
  directly. KEEP the accurate parity statements ("matches
  astropy.cosmology.arcsec_per_kpc_proper", "values match astropy Planck15",
  interface "follows the astropy FlatLambdaCDM class") — they document a real
  compatibility contract.

Verify

• Library-first ship: PyAutoArray then PyAutoGalaxy (Galaxy inherits Array).
• Smoke: an imaging fit that touches astropy (WCS / units) still runs.
• Confirm a clean-venv resolve pulls astropy 8.x, not 7.2.0.

Out of scope (recorded, not this task)

Other stale caps from the same audit — route separately if wanted: scipy
<=1.17.1→1.18.0, scikit-learn <=1.8.0→1.9.0, xxhash <=3.4.1→3.8.1
(Tier 1); networkx ==3.1→3.6.1, psutil ==6.1.0→7.2.2, corner
==2.2.2→2.3.0, nufftax <0.5.0→0.6.1 (Tier 2). Tier-4 sampler pins
(dynesty/nautilus/zeus/getdist/colossus) intentionally left alone.

[Jammy2211]

Shipped and merged. PyAutoArray#435 (merge 828d5c1) and sibling PyAutoGalaxy#558 (merge bfc766e2) are both on main: astropy cap dropped to astropy>=5.0 (floors, not pins) plus the cosmology docstring corrections. Both PRs carry pending-release — the change reaches PyPI with the next release. Merge authorized by human 2026-08-06.

[Jammy2211]

Smoke Test Results — 2026-08-06 (post-merge confirmation)

Workspace	Passed	Failed	Skipped	Total
autofit_workspace	7	0	1	8
autogalaxy_workspace	13	0	0	13
autolens_workspace	34	0	1	35
autolens_workspace_test	14	6	2	22
euclid_strong_lens_modeling_pipeline	6	0	0	6
HowToLens	50	0	0	50
Total	124	6	4	134

All 6 failures are autolens_workspace_test jax_likelihood scripts (imaging/{rectangular,mge,lp,smbh}.py, interferometer/rectangular.py, multi_dataset/mge.py) hitting hardcoded-DESIRED-literal assertion mismatches. Classification: pre-existing local-environment artifacts, not caused by this change — they reproduce identically on main, match the documented local-vs-CI jax_likelihood artifact class (lp.py passes in CI, fails locally), and the diff is dependency metadata + docstrings with no functional code. Skips are the standing no_run.yaml entries.

Run under astropy 8.0.1 with the merged commits (b90a284f / e3a8c600) on the import path. CI on the PRs never ran due to the 2026-08-06 GitHub Actions outage; this local sweep + the 929/1027 unit-test passes stood in for it per the human's merge-without-CI authorization.