PR CI renders only the website profile, so LaTeX-only breakage lands on `main` before anything catches it

[d-morrison]

Problem

PR CI renders only the website profile. .github/workflows/preview.yml's render job
calls Morrison-Lab/gha/.github/workflows/preview.yml@v2, which its own header comment
documents as fixing render-profile website. The PDF, docx, and revealjs renders happen
only in publish.yml, on main:

quarto render --to pdf
quarto render --to docx --no-clean
quarto render --to revealjs --no-clean

So a change that renders fine as HTML/MathJax but breaks LaTeX passes every PR check and
only turns red after it has merged, on the publish run — where it also blocks the site
deploy for everything else that has landed since.

Evidence

#1134 introduced \v0 in
chapters/parametric-survival-models.qmd. Every check on that PR was green. It merged, and
the next publish run died in LuaLaTeX:

Use of \v doesn't match its definition.
l.1610 ...tackrel{\text{def}}{=}{\lambda}(t \mid \v0

(run 34375808884,
tracked as #1155, worked around in
#1159, root cause upstream in
d-morrison/macros#86.)

MathJax and LaTeX disagree about macro semantics often enough for this to recur —
\providecommand is a MathJax no-op that macros-table.qmd already has to rewrite, and
delimited-parameter \defs of the kind that caused this failure are a LaTeX-only concept.
The HTML profile cannot see either class of defect.

Possible directions

Not obviously one right answer, so listing the trade-offs rather than prescribing:

1. Render the changed chapters to PDF on PRs. Highest fidelity, and the narrowest form
   (only chapters whose files the PR touched) may be affordable. Costs a TeX install and
   LuaLaTeX time in the preview job.
2. A LaTeX-only smoke check. Run quarto pandoc over latex-macros/macros.qmd plus the
   changed .qmd files with -t latex, then compile the emitted math with lualatex. Much
   cheaper than a full render, and it catches exactly this class (unexpanded or
   mis-expanding macros). Would not catch layout or float problems.
3. A lint for known-broken macro spellings. Cheapest, and only ever catches defects
   somebody has already been bitten by.

Option 2 reproduced this specific failure locally in seconds, so it looks like the best
cost/coverage point, but option 1 is the only one that actually tests what publish.yml
runs.