test: unify container smoke tests and make the fixture exercise the Dockerfiles

Three drifted copies of the container smoke test, and a fixture exercising none of the non-trivial Dockerfile behaviour (2026-07 audit). v0.11.0 (#125) shipped most of the fix; items 1-3 remain. containers/quantecon is the full image, containers/quantecon-build the lean one (containers/quantecon-build/README.md:1).

1. The fixture sidesteps the theme and the FreeSerif path

containers/quantecon/tests/minimal-jupyter-book/_config.yml:26 (html_theme: sphinx_book_theme) never imports the quantecon-book-theme both images ship; :28 (fontpkg: "") suppresses the FreeSerif path the .otf-symlink workaround exists to fix.

fontpkg matters most: both Dockerfiles symlink .otf for the FreeFont TTFs because Sphinx's default xelatex fontpkg wants Extension=.otf while Ubuntu ships .ttf (containers/quantecon/Dockerfile:54-61, stated :55-56; containers/quantecon-build/Dockerfile:54-59); emptying it takes the TeX-defaults path, so a symlink regression stays green. test-xelatex.tex:5 sets \setmainfont{DejaVu Serif}, so nothing else loads a FreeFont face through fontspec.

Theme versions can diverge too: full pins quantecon-book-theme==0.21.0 (containers/quantecon/environment.yml:17), lean a floor >=0.21.0 (containers/quantecon-build/environment.yml:66).

Real regression risk: dropping fontpkg can legitimately turn one image's pdflatex build red — the point of the change, so validate before merge (below).

2. test-container.sh is dead code the README still points at

containers/quantecon/tests/test-container.sh: untouched since 2025-11-20 (48d6a92), in no workflow, set -e only (:5, no -W), full image hardcoded at :14, :21, :29, :37 (lean invisible) — superseded by smoke-test.sh in a real container job. But containers/quantecon/README.md:223-228 still gives it as "To run all tests:", :230-234 presents its four steps as the test suite, and :241 demotes it only to an "Older…" entry. Deleting it loses no local Docker path: run-local-tests.sh:31-34 spells out the docker run … smoke-test.sh invocation, caveat HOME=/root so it cannot reproduce #85.

3. run-local-tests.sh drifts from the shared script

run-local-tests.sh:86 (html) and :99 (pdflatex) omit -W --keep-going, which smoke-test.sh:55 and :84 both pass, so a warning-level problem that fails CI passes locally. Align them, or comment that the divergence is deliberate for the macOS host toolchain.

Validating a fixture change (read before item 1)

test-container.yml fires only on workflow_dispatch or a completed Build QuantEcon Containers run (:6-11) — no pull_request trigger, so a fixture PR gets no CI signal. Dispatch on the branch with gh workflow run test-container.yml --ref <your-branch>; checkout uses github.event.workflow_run.head_sha || github.sha (:45), so it tests your fixture against the published :latest images. Otherwise the first signal is post-merge: build-containers.yml rebuilds on a push to main touching containers/**, then Test Container follows.

Not #100 (no automated tests for publish/preview/cache action logic), which stays open for that gap; this issue is the container tests being unable to fail.

Checklist

• containers/tests/smoke-test.sh <image> — pull, marker, xelatex, html, pdflatex, size — done in v0.11.0 (#125) as containers/quantecon/tests/smoke-test.sh (marker :44-45, xelatex :48-50, html :55, pdflatex :84); keep it there, not the proposed containers/tests/; pull is now the container: job and size a separate job (test-container.yml:77-101, compressed manifest bytes, not the old docker images figure)
• Collapse the two test-container.yml jobs into a matrix calling it — done in v0.11.0 (#125): .github/workflows/test-container.yml:22-25 matrix over quantecon/quantecon-build, :32-37 real container: jobs, :52-53 asserts HOME=/github/home so a silent revert to docker run is caught
• Confirm the lean image now gets a pdflatex build — done in v0.11.0 (#125), shared script :84 on both legs
• Fixture: execution on — done in v0.11.0 (#125): minimal-jupyter-book/_config.yml:10 execute_notebooks: "force", :16 raise_on_error: true; intro.md:33-44 numpy/scipy, :48-56 pandas, :63-81 a real matplotlib PNG, :92-102 plotly to_image; it can still fail — smoke-test.sh:40-42 stages a raising cell, :59-78 requires red and the sentinel, wired at test-container.yml:61-63 with if: always()
• Verify the fixture would have caught the kaleido/Chrome regression from #85 — done in v0.11.0 (#125): the plotly to_image cell runs in a real container: job, the HOME=/github/home condition #85 needed
• set -euo pipefail in run-local-tests.sh — done in v0.11.0 (#125): :5, plus a python/kaleido preflight at :24-42
• Fixture: remove the fontpkg: "" override at _config.yml:28
• Fixture: switch _config.yml:26 to quantecon_book_theme
• Dispatch Test Container on the branch and confirm both images stay green with the fixture change
• Delete containers/quantecon/tests/test-container.sh
• Repoint containers/quantecon/README.md:223-234 at smoke-test.sh / the Test Container workflow and drop the :241 "Older…" entry
• Align run-local-tests.sh:86 and :99 with -W --keep-going, or document why the macOS host toolchain differs

Start with the remaining checklist in containers/quantecon/tests/minimal-jupyter-book/_config.yml, then compare containers/quantecon/tests/test-container.sh, smoke-test.sh, run-local-tests.sh, and containers/quantecon/README.md. Read .github/workflows/test-container.yml before changing the fixture, and dispatch it on the branch with gh workflow run test-container.yml --ref . Done means both images pass the fixture, the obsolete script is removed, the README points to the shared workflow, and local flags are aligned or their divergence is documented.