[soodoku]

Part of #135 (step 4).

What problem do you want to solve?

The compile tests run on one plain table, so they can't tell whether text with markup characters reaches the PDF intact. Text that changes silently still compiles: in LaTeX, set {a} prints as set a, and p < 0.07 prints as p ¡ 0.07.

What this adds

tests/api/test_output_formats.py gets a round-trip matrix:

• Inputs: 12 strings, each with one markup character (% & _ # $ { ^ ~ < @ * [).
• Placement: each string goes into the caption, the notes, a coefficient label and a model head.
• Formats: LaTeX and Typst, which makes 96 cases.
• The check: each case compiles, then checks with pdfplumber that the PDF text contains the string, with whitespace removed.
• Avoiding false matches: each string carries a word that appears nowhere else in the table, and the other parts get a neutral note. My first version matched the default note's p < 0.1, which made three LaTeX cases pass for the wrong reason.

Results on main: 46 pass, and 50 fail. The 50 are marked xfail(strict=True) with their cause:

Format	Failing on main	Cause
LaTeX	% & _ # $ { ^ ~ < in all four parts (36)	no escaping; notes-percent is #99
Typst	$ ~ @ in all parts, * in caption and model head (14)	fixed by #122

I ran the matrix against the #99 and #122 branches. #99 turns exactly notes-percent into an unexpected pass, #122 does the same for exactly its 14 Typst cases, and neither causes a new failure. Because the markers are strict, whichever of those PRs merges after this one has to drop its markers. I'll push that change.

Tests

• Full suite: 544 passed, 50 xfailed.
• pre-commit and ty are clean.
• No snapshots change.
• The suite runs about 30 s longer locally (96 compiles). The LaTeX cases skip where pdflatex isn't installed, as the existing compile tests do.

🤖 Generated with Claude Code