chore: stop installing pynufft in Hands/Heart CI and CTI install docs

[Jammy2211]

Overview

pynufft is no longer a dependency of any PyAuto library — @PyAutoLabs/PyAutoArray#475 deleted TransformerNUFFTPyNUFFT and dropped pynufft from both the optional and dev extras. Four CI recipes and one install doc still install it. This is phase 3 of 3 of the pynufft-removal residue sweep (phases 1 and 2 shipped 2026-08-23); it is the CI + install-doc tier and is independent of the other two.

Severity is low and this is not urgent: every site pins 2025.1.1, not the broken 2022.2.2, so nothing is hitting the scipy.linalg.pinv2 failure and no build is red. What it costs is install time and unnecessary resolver surface on every release/validation run.

Plan

• Drop pip install pynufft==2025.1.1 from PyAutoHands release.yml (3 sites — one bare, two alongside numba, where numba is kept).
• Drop the same line from PyAutoHeart workspace-validation.yml (mode=release install step; numba kept).
• Delete pip install pynufft from PyAutoCTI docs/installation/source.rst, having first verified PyAutoCTI's suite genuinely passes without it.
• Confirm the affected workflows green afterwards — every run and every matrix leg.

Detailed implementation plan

Work Classification

Workspace / CI-and-docs. No PyAuto* library source changes, so there is no library-first merge gate — the three PRs are independent of one another and can merge in any order.

Affected Repositories

• PyAutoLabs/PyAutoHands (primary — 3 of the 4 CI sites)
• PyAutoLabs/PyAutoHeart
• PyAutoLabs/PyAutoCTI

Branch Survey

Repository	Current Branch	Dirty?
PyAutoHands	main	clean
PyAutoHeart	main	clean
PyAutoCTI	main	clean

Suggested branch: feature/pynufft-removal-phase-3 (all three repos)

Mobile session — no worktrees; work is done against fresh clones and pushed as three feature branches.

Implementation Steps

1. PyAutoHands/.github/workflows/release.yml
   • release_test_pypi → Tests step (~:296): delete the bare pip install pynufft==2025.1.1, leaving pytest and numba.
   • run_smoke_tests → Install from TestPyPI at pinned version (~:355): pip install pynufft==2025.1.1 numba → pip install numba.
   • release_workspaces → Regenerate API audit baseline (~:774): same rewrite.
   • The neighbouring "matplotlib deliberately unpinned" comments are about matplotlib, not pynufft — left untouched.
2. PyAutoHeart/.github/workflows/workspace-validation.yml (~:302), the mode=release "Install TestPyPI wheels" step: pip install pynufft==2025.1.1 numba → pip install numba. The nufftax>=0.6.1,<0.7.0 install two lines down is the live NUFFT backend and stays.
3. PyAutoCTI/docs/installation/source.rst:58: delete pip install pynufft under "For unit tests to pass you will also need the following optional requirements".

Verification that PyAutoCTI's suite does not need pynufft

Verified rather than assumed, three ways:

• PyAutoCTI/.github/workflows/main.yml is a thin caller of PyAutoHeart's reusable lib-tests.yml, which installs pip install ./PyAutoCTI[optional] (plus the PyAutoNerves/PyAutoFit/PyAutoArray chain) and never installs pynufft in either leg.
• PyAutoCTI/pyproject.toml's optional extra is ["numba"] — pynufft is not in it.
• The latest main run of that workflow (run 31134701968, head f40b2f7) is success across every leg: unittest on Python 3.12 and 3.13, plus the unittest-nojax leg.
• A full-tree grep of PyAutoCTI finds pynufft only in paper/paper.bib, files/citations.tex (both published-record material, deliberately out of scope) and the doc line being deleted — no import, no test, no config.

Key Files

• PyAutoHands/.github/workflows/release.yml — 3 sites
• PyAutoHeart/.github/workflows/workspace-validation.yml — 1 site
• PyAutoCTI/docs/installation/source.rst — 1 line

Green-CI caveat (read before closing this issue)

Every changed line lives on a release path, not on a PR-triggered path:

• release.yml is workflow_dispatch-only; the Tests step runs in rehearsal and live mode, while run_smoke_tests and release_workspaces are live-only.
• workspace-validation.yml's pynufft line is inside the mode=release install step, reached only via release-integrate.yml (Brain Release Agent dispatch, after an M1 TestPyPI rehearsal). The continuous mode=smoke channel never executed it.

So ordinary PR CI cannot exercise these lines. Full green confirmation therefore lands on the next release rehearsal / nightly release run; each repo's own PR checks are the pre-merge signal. Heart is currently RED on "release validation FAILED" for unrelated reasons.

Out of scope, noted while verifying

• PyAutoCTI/docs/installation/source.rst still lists pip install pylops==1.11.1 in the same block, and CTI's CI installs no pylops either while passing — the same evidence falsifies that line too. Not touched here (this task is scoped to pynufft); worth a follow-up.
• The same file points at PyAutoCTI/requirements.txt and PyAutoArray/optional_requirements.txt, neither of which exists any more (packaging moved to pyproject.toml extras). Separate doc rot, separate task.
• PyAutoCTI/files/citations.tex still asks users to cite PyNUFFT. Published-record class, deliberately untouched.

Original Prompt

Click to expand starting prompt

# Phase 3: stop installing pynufft in Hands/Heart CI and PyAutoCTI install docs

Type: maintenance
Target: ci
Repos:
- @PyAutoHands
- @PyAutoHeart
- @PyAutoCTI
Difficulty: low
Autonomy: supervised
Priority: normal
Status: draft
Filed: 2026-08-23

Phase 3 of 3. Parent: `pynufft_removal_downstream_residue.md`. Independent of
phases 1 and 2.

`pynufft` is no longer a dependency of any PyAuto library (@PyAutoArray#475
dropped it from both `optional` and `dev`), but four CI recipes and one install
doc still install it.

## Sites

- `@PyAutoHands/.github/workflows/release.yml:296,355,774` —
  `pip install pynufft==2025.1.1` (once bare, twice with `numba`)
- `@PyAutoHeart/.github/workflows/workspace-validation.yml:302` — same
- `@PyAutoCTI/docs/installation/source.rst:58` — `pip install pynufft`, listed
  under "For unit tests to pass you will also need the following optional
  requirements". **Confirm PyAutoCTI's suite genuinely has no such need before
  deleting the line** — verify, do not assume.

## Severity

Not urgent. These pin **2025.1.1**, not the broken `2022.2.2`, so they are
**not** hitting the `scipy.linalg.pinv2` failure and no build is red. This is
wasted install time and unnecessary resolver surface.

Worth knowing while working: these recipes are the only reason the local dev
environment still has `pynufft 2025.1.1` installed at all — removing them
changes what a fresh local env contains.

## Acceptance

- No PyAuto CI workflow installs `pynufft`.
- PyAutoCTI's install doc no longer instructs users to, with evidence its tests
  pass without it.
- The affected workflows are confirmed green afterwards — **every run and every
  matrix leg**, not just the first one reported.

[Jammy2211]

Shipped

Merged 2026-08-24 across all three repos:

PR	Change	CI at merge
#259	release.yml — 3 sites	3/3 legs (3.12 / 3.13 / 3.14)
PyAutoLabs/PyAutoHeart#162	workspace-validation.yml — 1 site	2/2 legs (3.12 / 3.13)
PyAutoLabs/PyAutoCTI#107	docs/installation/source.rst	3/3 legs (3.12 / 3.13 / no-jax)
PyAutoLabs/PyAutoCTI#108	unplanned — see below	3/3 legs

No PyAuto CI workflow installs pynufft any more, and CTI's install doc no longer tells users to.

Scope change: pylops went too

The CTI stanza held pip install pynufft and pip install pylops==1.11.1. The maintainer confirmed mid-task that pylops is defunct, so both went and the sentence went with them — a heading over an empty code-block is invalid RST. Same evidence covers both: no import or reference anywhere in PyAutoCTI (only paper/paper.bib), optional extra is ["numba"], CI installs neither and the suite passes. This closes the "out of scope, noted while verifying" item in the plan above.

The detour: PyAutoCTI main was already red

Opening #107 surfaced a failure on all three CTI legs that was not the diff — a single deleted .rst line, run on the PR merge ref, so the failure was main's:

test_serial_eper.py::test__region_list_from__array_2d_list_from
  autoarray.exc.MaskException: shape_native[1] must be a positive number of pixels; got 0. (3, 0)

lib-tests.yml clones PyAutoArray main at run time, and PyAutoArray#440's zero-length shape_native guard (2026-08-09) landed after CTI's last main run (2026-08-07), so the break sat dormant for two weeks and would have blocked every CTI PR.

The guard didn't break a working test — it exposed one that checked nothing: pixels=(2, 3) on region (0, 3, 5, 8) of a 3×10 fixture resolves to columns 10–11, extracting (3, 0), and (empty == [[10.0], [10.0], [10.0]]).all() is vacuously True for any expected value. 10.0 isn't even in the fixture. #108 fixed it test-only and deliberately left open whether such a window should raise or clip — tracked in PyAutoMind as draft/bug/autocti/serial_eper_zero_width_region_vs_autoarray_guard.md.

The green-CI caveat in the plan still stands

Every Hands/Heart line changed here is on a release path (release.yml is workflow_dispatch-only; Heart's is in the mode=release leg), so ordinary PR CI cannot exercise them. Pre-merge signal was each repo's own checks plus the diff; full confirmation lands on the next release rehearsal / nightly run. The CTI leg is fully exercised and green.

Heart was RED on "release validation FAILED" throughout — pre-existing, unrelated, acknowledged at launch.

Still open (untouched, as planned)

• CTI's Building All Projects section still points at PyAutoArray/optional_requirements.txt and PyAutoCTI/requirements.txt — neither file exists any more. Separate doc rot.
• paper/paper.bib and files/citations.tex still cite PyNUFFT and PyLops — published-record material, deliberately out of scope.

---

Generated by Claude Code