fix: requirements for tutorials 6/7 + revert workshop Colab stopgap

[Jammy2211]

Overview

HowToFit tutorials 6 and 7 construct af.BlackJAXNUTS (tutorial 6 also af.Nautilus), but neither blackjax nor nautilus-sampler is on the documented local install route: PyAutoFit declares them only in its optional extra and HowToFit/requirements.txt is just autofit + numba. CI runs at PYAUTO_TEST_MODE=2, which bypasses the sampler, so the smoke gate is green while a reader hits ModuleNotFoundError: No module named 'blackjax'.

The Colab leg of this prompt already merged as PyAutoNerves#168 (_SHARED_EXTRAS now carries blackjax>=1.6.2 and nautilus-sampler==1.0.5) and goes live with the next autonerves PyPI release — PyPI is still 2026.9.15.1, which predates it. This task does the remaining local-install leg and, at the human's request now that the 2026-09-15 workshop is over, reverts the Colab stopgap that was merged for it (HowToFit #63/#64, HowToGalaxy #77, HowToLens #85), each of which was labelled "REMOVE once autonerves > 2026.9.15.1".

Merge condition: reverting the stopgap before autonerves > 2026.9.15.1 is on PyPI re-opens the Colab gap on every notebook until that release lands. Merge these PRs only once pip index versions autonerves shows a newer version, or with the human explicitly accepting the interim gap.

Plan

• Add blackjax and nautilus-sampler to HowToFit/requirements.txt so the local route installs what the tutorials run.
• Revert the four workshop-stopgap merges with plain git revert -m 1 commits (no history rewrite) in HowToFit, HowToGalaxy and HowToLens.
• Revert HowToFit fix: point Colab links at main so the #63 dependency stopgap reaches readers #64's README badge repointing too: the release tool only re-pins version-shaped badge tags, so blob/main/ links would never be re-pinned.
• One PR per repo, all cross-referencing this issue; no smoke run is meaningful for a pure revert to released notebooks plus a requirements line.
• Out of scope: cutting the autonerves release; a smoke-profile import guard for constructed searches (separate prompt); the /prm close-out of the colab-workshop-dep-stopgap task.

Detailed implementation plan

Work Classification

Workspace

Affected Repositories

• HowToFit (primary)
• HowToGalaxy
• HowToLens

Branch Survey

Repository	Current Branch	Dirty?
./HowToFit	main @ f4bca40	clean
./HowToGalaxy	main @ 95eb176	clean
./HowToLens	main @ 5c3c727	clean

Worktree guard: HowToFit/HowToGalaxy/HowToLens are claimed by colab-workshop-dep-stopgap (all PRs merged, stale claim) and HowToFit by howtofit-mode (README/AGENTS prompt propagation, 0 own commits). Human-approved parallel-claim override 2026-09-15; file sets disjoint.

Suggested branch: feature/blackjax-reqs-stopgap-revert
Worktree root: ~/Code/PyAutoLabs-wt/blackjax-reqs-stopgap-revert/

Implementation Steps

1. HowToFit: git revert -m 1 f4bca40 (PR fix: point Colab links at main so the #63 dependency stopgap reaches readers #64: 20 README badges blob/2026.9.15.1/ → blob/main/, stopgap cell in start_here.ipynb), then git revert -m 1 1b85e12 (PR fix: install autofit's lazily-imported deps in the Colab setup cell #63: stopgap cell in 18 notebooks/**/*.ipynb + one README paragraph).
2. HowToFit: edit requirements.txt to autofit, numba, blackjax, nautilus-sampler (unpinned, matching the file's and its siblings' convention; pinned literals would recreate the drift PyAutoNerves#168 fixed). Separate commit from the reverts.
3. HowToGalaxy: git revert -m 1 95eb176 (PR #77: stopgap cell in 32 notebooks).
4. HowToLens: git revert -m 1 5c3c727 (PR #85: stopgap cell in 50 notebooks).
5. Verify per repo: git diff 2026.9.15.1 HEAD --stat is empty for HowToGalaxy/HowToLens and, for HowToFit, shows only bd08343's docs change plus requirements.txt | 2 +; HowToFit README has 20 blob/2026.9.15.1/ and 1 blob/main/ link (the tagged counts); every touched .ipynb parses as JSON; a scratch venv pip install -r requirements.txt followed by python -c "import blackjax, nautilus" succeeds.
6. ship_workspace: one PR per repo against main, PR bodies carrying the merge condition above.

Key Files

• HowToFit/requirements.txt — the local install route the tutorials document.
• HowToFit/README.md, HowToFit/start_here.ipynb, HowToFit/notebooks/**/*.ipynb — restored to the 2026.9.15.1 content by the reverts.
• HowToGalaxy/notebooks/**/*.ipynb, HowToLens/notebooks/**/*.ipynb — same.
• PyAutoHands/autohands/bump_colab_urls.sh:30 — why the blob/main/ badges must not stay.
• PyAutoNerves/autonerves/setup_colab.py _SHARED_EXTRAS — the already-merged Colab fix (#168).

Original Prompt

Click to expand starting prompt

HowToFit tutorials 6 and 7 die at the NUTS fit — blackjax is on no install route

Type: bug
Target: HowToFit
Repos:

• HowToFit
• PyAutoNerves
  Themes:
• tutorials
• colab
• dependencies
  Difficulty: small
  Autonomy: safe
  Priority: high
  Status: formalised
  Consequence: judge
  Review-minutes: 10
  Unattended: ready
  Witness: with only pip install autofit numba (or a Colab setup_colab.setup("howtofit")), python3 scripts/chapter_1_introduction/tutorial_6_gradients.py and tutorial_7_the_details.py both run to completion instead of raising ModuleNotFoundError: No module named 'blackjax'.
  Filed: 2026-09-15

Symptom

Measured 2026-09-15 against HowToFit main @ b174c03 (PR #62), PyAutoFit main,
Python 3.12.3, at real settings (no PYAUTO_TEST_MODE, visualization on):

tutorial_6_gradients.py    FAIL (11.0s)  ModuleNotFoundError: No module named 'blackjax'
tutorial_7_the_details.py  FAIL (67.1s)  ModuleNotFoundError: No module named 'blackjax'

Installing blackjax>=1.6.2 and nautilus-sampler and re-running, nothing
else changed: tutorial 6 rc=0 in 43s, tutorial 7 rc=0 in 73s, both completing
their real BlackJAXNUTS (and, in 6, Nautilus) fits. The missing
dependency is the whole defect.

Why no gate sees it

CI runs PYAUTO_TEST_MODE=2 (config/build/profile_smoke.yaml), which
bypasses the sampler, so af.BlackJAXNUTS is never constructed and blackjax
is never imported. Both tutorials PASS the smoke gate — 18/18 green — while
being unrunnable by a reader. This is the same blind spot config/build/no_run.yaml
already documents for tutorial_5_expectation_propagation ("only passes today
because CI runs at PYAUTO_TEST_MODE=2, which bypasses the sampler").

Why it is missing on every route

• HowToFit/requirements.txt is autofit + numba.
• PyAutoFit declares blackjax in the optional extra, not the base
  dependencies (PyAutoFit/pyproject.toml:83, [project.optional-dependencies] optional), so pip install autofit does not bring it. Same for
  nautilus-sampler, which tutorial 6 also runs (:727).
• Colab: autonerves/setup_colab.py installs _SHARED_EXTRAS with --no-deps
  (:280). _SHARED_EXTRAS (:37-46) lists pyvis, dill, jaxnnls,
  nautilus-sampler, timeout_decorator, anesthetic, emcee, dynesty — and no
  blackjax. Exactly the gap PyAutoNerves #166 closed for emcee and dynesty,
  one package short.

Searches actually run: tutorial 6 — LBFGS, MultiStartAdam, Emcee,
BlackJAXNUTS(:685), Nautilus(:727); tutorial 7 — DynestyStatic x3,
Emcee x2, MultiStartAdam x2, BlackJAXNUTS(:1086).

Tutorial 7's NUTS fit is new tonight: PR #62 moved the Hamiltonian
diagnostics out of tutorial 6 and gave tutorial 7 its own short BlackJAXNUTS
fit, so tutorial 7 is a fresh breakage, tutorial 6 a pre-existing one.

Fix

1. PyAutoNerves/autonerves/setup_colab.py: add "blackjax>=1.6.2" to
   _SHARED_EXTRAS, tracking PyAutoFit/pyproject.toml's specifier as the
   emcee/dynesty entries already do.
2. HowToFit/requirements.txt: add blackjax and nautilus-sampler so the
   documented local route installs what the tutorials run.
3. While there: _SHARED_EXTRAS pins nautilus-sampler==1.0.4 where
   PyAutoFit's optional extra pins ==1.0.5. Reconcile.

A follow-up worth its own prompt: the smoke profile's PYAUTO_TEST_MODE=2
cannot see a missing sampler dependency at all. A cheap guard is an import
check over the af.<Search> classes each script constructs, run at smoke
settings.

[Jammy2211]

Workspace PRs Created

• HowToFit: fix: requirements for tutorials 6/7 + revert workshop Colab stopgap #66
• HowToGalaxy: revert: workshop Colab dependency stopgap (#77) HowToGalaxy#78
• HowToLens: revert: workshop Colab dependency stopgap (#85) HowToLens#86

Merge condition (all three): autonerves newer than 2026.9.15.1 on PyPI (PyAutoNerves#168 released), or an explicit decision to accept the interim Colab gap. /prm per PR; the issue closes with the last one.

[Jammy2211]

Shipped — 2026-09-15

PRs

• fix: requirements for tutorials 6/7 + revert workshop Colab stopgap #66
• revert: workshop Colab dependency stopgap (#77) HowToGalaxy#78
• revert: workshop Colab dependency stopgap (#85) HowToLens#86

Summary

HowToFit/requirements.txt now carries blackjax>=1.6.2 and nautilus-sampler==1.0.5 (PyAutoFit's own specifiers), so the documented local install route brings what tutorials 6 and 7 construct. The 2026-09-15 workshop Colab stopgap was reverted in all three HowTo repos with plain git revert -m 1 commits (HowToFit #63/#64, HowToGalaxy #77, HowToLens #85); notebooks are byte-identical to the 2026.9.15.1 release and the HowToFit README badges are back on release-tag pinning, which the release tool can re-pin.

Session Notes

• The Colab leg of this prompt had already merged as PyAutoNerves#168 under colab-bootstrap-lazy-deps; only the local-install leg remained.
• Merged by human decision while PyPI autonerves was still 2026.9.15.1: Colab readers hit the missing-dependency error until the next autonerves release publishes PyAutoNerves#168. That release is the outstanding obligation (release-gate: PyAutoNerves on the completion record).
• Smoke at PYAUTO_TEST_MODE=2 cannot see a missing sampler backend; the follow-up guard is filed as PyAutoMind/draft/feature/pyautohands/smoke_profile_cannot_see_a_missing_sampler.md.
• colab-workshop-dep-stopgap (the task this reverts) was closed out the same day: record PyAutoMind/complete/2026/09/colab-workshop-dep-stopgap.md.