Build a Markdown.Table instead of hand-rolling the markdown writer

[andreasnoack]

Follow-up to #1007, rebased on it now that it has merged.

#1007 fixed the escaping bug by hand, as a minimal short-term fix. This removes the reason the bug was possible: it stops reimplementing the markdown table format altogether. The rows still go through Base's print_matrix_row — that is what produces the decimal alignment — and the padded strings are handed to Markdown.Table, so escaping and the alignment rules come from the stdlib rather than from escape_markdown_pipes.

The justification for hand-rolling does not hold

The method carries this comment, from #664:

not using Markdown stdlib here because that won't give us nice decimal alignment (even if that is lost when rendering to HTML, it's still nice when looking at the markdown itself)

The stdlib does not produce decimal alignment, but it preserves it: padcells! adds the same width to every cell in a column, so pre-padded equal-width cells keep their decimal points lined up. Since print_matrix_row already emits exactly those strings, the alignment survives:

|     |      Estimate | Comments |  df |      t | p      |
|:--- | -------------:| --------:| ---:| ------:|:------ |
| x1  |   1.45666     |    Good  |   1 | -12.56 | 0.1200 |
| x2  | -23.14        |    Great |  56 |   0.13 | 0.3467 |
| x3  |   1.56734e-13 |    Bad   |   2 |   0.00 | <1e-15 |

Decimal points at the same offset in every row, pipes escaped, and the result parses as a Markdown.Table.

Costs

• Markdown moves from the test-only dependency Escape pipes when writing a CoefTable as markdown #1007 added to a direct dependency (still a stdlib, but a new entry in [deps]/[compat], and it drops out of [extras]/[targets]).
• The emitted spacing changes (|:--- | -------------:| rather than |:---|--------------:|), so both golden markdown tests get new expected output. Cell content is identical, and the rendered HTML is unaffected, as is the plain-text show method.

One thing to decide separately

The hand-rolled writer selects the alignment marker with j-1 in [ct.teststatcol; ct.pvalcol] while j indexes colnms, so the test-statistic column never receives its left-align marker and the p-value column receives it via the shift. That looks like an off-by-one, but changing it would change output beyond this refactor, so it is carried over verbatim with a comment. Happy to fix it here or in a follow-up, whichever you prefer.

#1007's regression test (a pipe in a header, a row name and a string cell; asserts equal cell counts per row and that the output parses as a Markdown.Table) is kept as-is and still passes. Full test suite passes locally on the rebased branch.

🤖 Generated with Claude Code