Pin lean container's science stack to the Anaconda 2025.12 baseline (#28)

[mmcky]

Closes #28.

Root cause

The quantecon-build (lean) image listed its scientific stack unpinned, so numpy/scipy/pandas/etc. resolved to latest at build time and drifted ahead of the Anaconda 2025.12 set that every lecture repo pins via anaconda=2025.12 (and that the full quantecon image uses). That drift produced non-reproducible images that diverged from what the lectures are tested against, and surfaced downstream as a CellExecutionError in un_insure.md's sp.optimize.fsolve(Vu_error_Λ, 15000, ...) path for repos that build on the lean image (e.g. lecture-dp). The full image already pins anaconda=2025.12 and was unaffected.

Fix

Pin the lean image's core scientific stack to the exact Anaconda 2025.12 versions, so the lean image stays reproducible and consistent with the full image and the source repos.

package	was	now (= Anaconda 2025.12)
numpy	unpinned	2.3.5
scipy	unpinned	1.16.3
pandas	unpinned	2.3.3
matplotlib	unpinned	3.10.6
seaborn	unpinned	0.13.2
sympy	unpinned	1.14.0
numba	unpinned	0.62.1
networkx	unpinned	3.5
statsmodels	unpinned	0.14.5
scikit-learn	unpinned	1.7.2

Pinning the whole core stack (not just numpy/scipy) eliminates the drift class entirely and keeps numba/numpy in their tested pairing. Dependabot (conda ecosystem) will surface future bumps for review; bump the set when the lecture repos move to a newer Anaconda release.

On the exact breakage — honest caveat

I could not reproduce the original un_insure.md failure on current packages. The actual calibration cell runs cleanly on four combinations I tested — numpy 2.3.5+scipy 1.16.3 (2025.12), numpy 2.4.0+scipy 1.15.3, numpy 2.4.0+scipy 1.16.3, and numpy 2.4.6+scipy 1.17.1 (2026.06). numpy 2.4.0 did make float()/int() of a 1-d array a hard TypeError, which is the most likely trigger class, but the specific combination that broke it in March 2026 appears to have been a transient regression that's since been patched upstream. The fix here stands regardless: it addresses the root cause (unpinned drift / non-reproducible builds that diverge from the source-repo baseline), not the transient trigger.

Why not Anaconda 2026.06

2026.06 exists and is installable for py313, and un_insure.md does run on its stack (numpy 2.4.6 / scipy 1.17.1). But it jumps to pandas 3.0 (a major release no lecture repo is validated against) and would make the containers diverge from the source repos (all still on 2025.12), recreating the drift this PR fixes. The newer baseline should be a separate, coordinated bump of the lecture repos + both containers.

Verification

• Confirmed the pinned set solves on linux-64 (the container platform) via conda env create --dry-run with CONDA_OVERRIDE_GLIBC=2.39; numpy resolves to 2.3.5, numba to 0.62.1, nodejs stays at 20.x (within >=20,<25).
• Version data taken directly from conda search anaconda=2025.12=py313_mkl_0 --info against pkgs/main.
• Ran the actual un_insure.md calibration cell (extracted helpers + brentq→fsolve) on the pinned baseline and on the current-latest stack.

Notes for reviewers

• Testing does cover this. test-containers-lectures.yml runs after each container build (workflow_run) and its Stage 1 "Build HTML (Execute Notebooks)" executes notebooks for lecture-python-advanced.myst (which contains un_insure.md) on both images. Caveat: that workflow is non-blocking (runs after :latest is pushed) and has been red for months.
• Separate pre-existing breakage (not addressed here): the container tests are currently failing on a ChromeNotFoundError from kaleido v1 (unpinned) on Plotly static-image export (BCG_complete_mkts.md, BCG_incomplete_mkts.md, knowing_forecasts_of_others.md) — a distinct "unpinned dep bumped a major" problem worth its own issue/PR. The container test signal won't go fully green until that's fixed too.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes reproducibility and downstream notebook execution failures in the lean quantecon-build container by pinning its core scientific Python stack to the Anaconda 2025.12 baseline, aligning it with the full quantecon image and with lecture repositories that pin anaconda=2025.12.

Changes:

• Pin the lean container’s core scientific stack (numpy/scipy/pandas/matplotlib/etc.) to the Anaconda 2025.12 versions to prevent “latest at build time” drift.
• Update the lean container README to document that the science stack is pinned to the 2025.12 baseline.
• Add a changelog entry describing the fix and the downstream failure it addresses.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
containers/quantecon-build/environment.yml	Pins the core scientific stack to Anaconda 2025.12 versions to prevent dependency drift in the lean image.
containers/quantecon-build/README.md	Documents that the lean image’s science stack is pinned to the Anaconda 2025.12 baseline.
CHANGELOG.md	Records the fix and the rationale (preventing numpy drift causing downstream notebook execution errors).