fix: verify_install Check F loads stale dataset + resolves mixed dev/released stack

[Jammy2211]

Overview

PyAutoHeart's verify_install Check F (Colab-simulation) driver FAILs even though the real new-user install path (Checks A/C/D) is green, so the whole verify_install_release job reports FAIL. Surfaced by the 2026-07-13 release-validation Stage-3 run (workspace-validation run_id=29266305445, TestPyPI 2026.7.13.1.dev65501). This is a harness gap, not a library defect — the core install path is healthy.

Two distinct problems, both isolated to heart/checks/verify_install.sh's check_f():

1. Stale dataset load (primary). The generated F_driver.py loads dataset/imaging/simple/{data,noise_map,psf}.fits — a path the current imaging/start_here.py no longer uses (and which isn't bundled) → FileNotFoundError. The driver comment says it runs "one real cell from the imaging start_here", but start_here.py now loads the bundled cosmos_web_ring JWST dataset (pixel_scales=0.06) — which is exactly why Check A (which runs start_here.py) passes. cosmos_web_ring is one of the few example datasets that ship committed with the workspace, so it's present after the setup_colab clone.

2. Mixed dev/released stack (secondary). Check F's pip install "$PIP_INSTALL_TARGET" jax installs autolens at the pinned dev version but lets autoarray/autoconf/autofit/autogalaxy resolve to the latest final on PyPI (2026.7.9.1) — because dev versions are pre-releases pip won't pick for a floor-only dependency. So Check F isn't exercising a coherent dev stack like A/C/D do.

Plan

• F_driver dataset: change the from_fits cell to load cosmos_web_ring (data.fits/noise_map.fits/psf.fits, pixel_scales=0.06), mirroring the current start_here.py cell. No simulator needed — the dataset is bundled. Keep the dataset/ existence assertion.
• Coherent stack: in check_f(), when $TARGET_VERSION is set, pin all five PyAuto packages (autolens autoarray autoconf autofit autogalaxy) to it in the emulated-Colab pip install, so the deps come from the same dev wheels as A/C/D. (The later verbatim pip install autoconf --no-deps inside the driver then finds the pinned dev autoconf already satisfied → no-op, so it can't silently pull a released autoconf.) When unpinned (released mode), behaviour is unchanged.
• Leave Checks A/C/D untouched (green). Final PASS confirmation requires a TestPyPI rehearsal re-run (network + a fresh five-package dev upload) — out of band for the code change.

Detailed implementation plan

Work Classification

Library (harness only — PyAutoHeart). No workspace/library-source change.

Affected Repositories

• PyAutoLabs/PyAutoHeart (heart/checks/verify_install.sh, check_f() only)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoHeart	main	clean

Suggested branch: feature/verify-install-check-f-fix
Worktree root: ~/Code/PyAutoLabs-wt/verify-install-check-f-fix/

Implementation Steps

1. heart/checks/verify_install.sh check_f() — build an F-specific install list that pins all five PyAuto packages to $TARGET_VERSION when set (else keep $PIP_INSTALL_TARGET), then pip install "${PIP_INDEX_ARGS[@]}" "${f_targets[@]}" jax.
2. Same function's F_driver.py heredoc — swap the simple paths for cosmos_web_ring at pixel_scales=0.06; update the print/assertion text.
3. Verify: bash -n the script + python -m py_compile the extracted driver; confirm the cosmos_web_ring fits are bundled in autolens_workspace.

Verification limits

Check F needs live network + a fresh TestPyPI dev upload of all five packages to run end-to-end; that confirmation belongs to the next release-validation rehearsal. The code change is verified by shell/Python syntax + path/logic review here.

Autonomy

--auto, effective level supervised (bug work-type cap). Plan written here; run parks at ship sign-off. Merge/close remain human.

Original Prompt

Click to expand starting prompt

verify_install Check F: assumes a bundled dataset (no auto-simulate) and resolves a mixed dev/released stack

Type: bug
Target: health_fixes
Repos:

• PyAutoHeart
  Difficulty: small
  Autonomy: supervised
  Priority: normal
  Status: formalised

PyAutoHeart's verify_install Check F driver fails on a missing dataset, so the whole
verify_install_release job reports FAIL even though the real new-user install path is
healthy. Surfaced by the 2026-07-13 release-validation Stage-3 run (workspace-validation
run_id=29266305445, TestPyPI 2026.7.13.1.dev65501). Check status was:

A  PASS  pip install + start_here.py + welcome.py
C  PASS  conda(python=3.12) + start_here + welcome
D  PASS  autolens[optional]==2026.7.13.1.dev65501 resolved + imports
F  FAIL  driver rc=1

Two distinct problems in Check F only:

1. Missing auto-simulate (primary). F_driver.py:36 calls
   Imaging.from_fits('dataset/imaging/simple/data.fits') from its CWD, but never simulates
   that dataset first, so it crashes FileNotFoundError: dataset/imaging/simple/data.fits.
   Datasets are meant to be auto-simulated at run time, never bundled/committed (only a few
   select example .fits ship with a workspace) — this gap was exposed by the 2026-07-13
   release dataset -f leak fix (PyAutoBuild#150) that stopped force-adding dataset/.
   Check A's start_here.py passes because it simulates its own dataset; Check F must do the
   same (run the matching simulator, or simulate to the exact folder it then loads) instead
   of assuming a bundled file. This is a HARNESS gap, not a library defect.

2. Mixed dev/released stack (secondary). Check F's env resolved
   autolens 2026.7.13.1.dev65501 (the rehearsal dev wheel) but its dependencies came from
   released PyPI — autoarray/autoconf/autofit/autogalaxy 2026.7.9.1 — so Check F is not
   even exercising a coherent dev stack. Pin all five PyAuto packages to the TestPyPI dev
   version (or install with the TestPyPI index for the whole family) so Check F tests the
   same wheels as Checks A/C/D.

Fix both in the Check-F driver, then re-run verify_install against a TestPyPI rehearsal
version to confirm F PASS. Keep the core new-user path (A/C/D) untouched — it is green.

Diagnosis note

The prompt's primary framing was "auto-simulate the simple dataset". Investigation shows the current start_here.py loads the bundled cosmos_web_ring (not a simulated simple), which is why Check A passes — so the faithful fix is to load cosmos_web_ring, no simulator required.

[Jammy2211]

⏸️ Parked at ship sign-off — supervised --auto (needs your go-ahead to open the PR)

Both Check-F fixes are implemented; under the supervised bug cap the run stops here for your review before commit/push/PR. Nothing committed or pushed. Merge/close stay human.

What changed (worktree ~/Code/PyAutoLabs-wt/verify-install-check-f-fix, branch feature/verify-install-check-f-fix)

heart/checks/verify_install.sh — check_f() only (+34/-7):

1. Stale dataset → bundled cosmos_web_ring (primary). The F_driver.py heredoc now loads dataset/imaging/cosmos_web_ring/{data,noise_map,psf}.fits at pixel_scales=0.06 — the exact dataset the current imaging/start_here.py loads (and why Check A passes). It's a bundled example that ships committed with the workspace, so it's present after the setup_colab clone; no simulator needed. (The prompt guessed a simple auto-simulate, but current start_here.py doesn't use simple — it loads the bundled cosmos_web_ring.)
2. Coherent dev stack (secondary). When $TARGET_VERSION is set (a TestPyPI rehearsal), the emulated-Colab install now pins all five PyAuto packages (autolens autoarray autoconf autofit autogalaxy) to it, instead of pinning only autolens and letting the deps resolve to the latest final on PyPI. The later verbatim pip install autoconf --no-deps inside the driver then finds the pinned dev autoconf already satisfied → no-op (can't pull a released autoconf). Released (unpinned) mode is unchanged.

Checks A/C/D untouched.

Ship evidence

• bash -n verify_install.sh → OK; extracted F_driver.py → py_compile OK.
• cosmos_web_ring/{data,noise_map,psf}.fits confirmed committed in autolens_workspace.
• Existing tests/test_verify_install_script.py → 5 passed.
• API/Scripts changed: none — harness-only (one shell function). No library/workspace source touched.

Verification limit (honest)

Check F only runs end-to-end with live network + a fresh TestPyPI dev upload of all five packages, so a true F PASS can only be confirmed by the next release-validation rehearsal — I can't reproduce that here. The change is verified by shell/Python syntax, the dataset-presence check, path/logic review, and the unit test. Worth re-running verify_install --testpypi against a fresh rehearsal version once one is uploaded.

On your go-ahead

Reply here (or relaunch) and I'll run the full ship_library gate (repo has a pytest suite → runs it + review + Heart), commit, push, and open the PR on PyAutoHeart.

[Jammy2211]

✅ Shipped (PR open, pending-release). Human ship sign-off given ("ship all").

• fix: verify_install Check F stale dataset + mixed dev/released stack #71 — pytest tests/ 263 passed; bash -n + driver py_compile OK; cosmos_web_ring bundled confirmed

Full F PASS still needs a live TestPyPI rehearsal re-run (network + fresh 5-package dev upload) — worth running verify_install --testpypi against the next rehearsal version. Merge/close remain yours.