PDF/LaTeX render only runs post-merge, so a PR can break publish.yml with every check green

[d-morrison]

Important

This issue was originally filed on a false premise, twice over. Both
versions are retracted; what follows is the corrected, much narrower ask.
The retraction history is in the comments below.

What was wrong

The first version claimed \sb is a silent no-op because LaTeX defines
\let\sb=_, and asked for a 110-site sweep. That is false: macros.qmd is
included as document content, and Pandoc's latex_macros extension expands
\providecommand macros in the reader, so LaTeX never receives \sb.
Verified with quarto render --to latex, which emits
\mathopen{}\left[(1-\pi) + \pi\right]\mathclose{}; pdflatex then exits 0.
Upstream macros#85 is
closed as not planned.

The second version claimed "no check anywhere compiles this book's math
through LaTeX". Also false. .github/workflows/publish.yml:110 runs
quarto render --to pdf under tinytex: true on every push to main, and
its report-failure job opens an issue when that fails. It has been green
continuously, including on 873b4a09 --- the tip that carries all 110 \sb
uses. That green history was standing evidence against the original claim
before it was ever written.

The actual, remaining gap

The PDF build runs post-merge only. publish.yml triggers on
push: branches: main, so a PR that breaks the LaTeX output is not caught
until after it lands, at which point the failure arrives as an auto-filed
issue rather than as a red check on the PR that caused it.

The PR preview workflow renders HTML. MathJax and LaTeX accept different
things, so a PR can be green on every check and still break publish.yml on
merge.

Suggested work

Add a PDF (or LaTeX-only) render to the PR checks, or a scheduled pre-merge
job, so the failure lands on the PR rather than on main.

Two constraints worth respecting:

• Render through Quarto, never around it. Any check that extracts
  $$...$$ blocks and feeds them to pdflatex directly tests an artifact
  this book never builds, because it skips the Pandoc macro expansion above.
  That shortcut is exactly what manufactured the retracted claim.
• Cost. A full --to pdf needs renv and tinytex and is slow. If that is
  too heavy for every PR, --to latex on the changed chapter plus a
  pdflatex pass on the emitted .tex is much cheaper and still exercises
  the real expansion path.

Found while driving #1138.

Filed by Claude Code (AI agent).

[d-morrison]

Retracting this issue --- upstream macros#85 was based on a false premise and is closed.

Pandoc's latex_macros extension expands \sb before LaTeX sees it, so the 103 uses in this repo render correctly in every profile and no sweep is needed. Verified with quarto render --to latex: \sb{(1-\pi) + \pi} reaches the .tex as \mathopen{}\left[(1-\pi) + \pi\right]\mathclose{}, and pdflatex exits 0.

#1138 has reverted the \sb → \left[...\right] conversions it had made; the packaged macro was correct, and hand-expanding it drops the \mathopen{}/\mathclose{} spacing.

One part of this issue survives on its own merits and is worth keeping, so I have retitled it: the repo's pre-commit checklist renders --to html only, so no local or CI check compiles the book's math through LaTeX. That is a real gap regardless of \sb. A cheap smoke check would render one chapter --to latex and run pdflatex on the result --- through Pandoc, which is the mistake to avoid: a harness that extracts math and skips Pandoc tests an artifact the book never builds, and that is exactly how the \sb false alarm was produced.

Posted by Claude Code (AI agent) --- not written by a human.

[d-morrison]

Second correction. My previous retraction comment fixed the \sb premise but replaced it with another false claim --- that no check compiles this book's math through LaTeX. publish.yml does exactly that on every push to main, with tinytex: true, and has been green throughout.

I have rewritten the body rather than only the title this time. The earlier version also still recommended building CI around a $$-extraction pdflatex harness, which is the very shortcut that produced the original false alarm; that recommendation is gone, and the body now says explicitly not to do it.

The surviving ask is narrow: the PDF build is post-merge only, so a PR can break it with every check green.

Posted by Claude Code (AI agent) --- not written by a human.