Escape a paragraph line that would reparse as a table delimiter row - #19877

Fixes #19847.

Description

A GFM delimiter row indented by four or more spaces is not a table, but after formatting it becomes one, so the second run produces different output:

<!-- Input -->
| x | y |
    |---|---|

<!-- Prettier stable -->
| x | y |
|---|---|

<!-- Prettier stable, second run -->
| x   | y   |
| --- | --- |

<!-- Prettier main -->
| x | y |
\|---|---|

Cause

remark only accepts a delimiter row indented up to three spaces, so at four the two lines are a plain paragraph and arrive as a single text node with the indentation already stripped. Printing a paragraph drops leading whitespace on continuation lines, which is insignificant everywhere else, but here it turns the text into a valid table.

Change

Escaping the delimiter row so it cannot be re-parsed, which is what the same file already does for the analogous case one line up:

// escape indented pseudo setext header, e.g. `Previous line↵␣␣␣␣===`
return `\\${text}`;

The issue asks for the table to be printed on the first run instead. That would change what the document renders as — the input is a paragraph, and a table is not the same output — so this keeps it a paragraph, consistent with the setext case.

The line is reconstructed from the sentence's sibling nodes rather than from options.originalText. A delimiter row is only |, -, : and spaces, so it is always contained in one sentence, and reading it from the AST means blockquote and list markers never enter the picture.

One thing I left out deliberately: GFM only forms a table when the header row and the delimiter row have the same number of cells, and this does not check that, so a delimiter row under a header with a different cell count is escaped although no table would have formed. Counting the header's cells means reconstructing a line that can span several nodes once it contains emphasis or code spans. No existing snapshot changed without it — happy to add it if you would rather have the exact rule.

Tests

tests/format/markdown/paragraph/issue-19847.md, which runs under all three proseWrap modes and covers the compact, spaced, aligned and no-outer-pipe delimiter shapes, a single-cell row, blockquote and list item nesting, and a trailing paragraph line. Two cases in it should not change: a three-space indent is a real table and still prints as one, and a delimiter row alone after a blank line is an indented code block.

tests/format/{markdown,mdx}: 1503 snapshots, none of the existing ones changed. Full suite is unchanged against a stashed tree — same four failing suites either way, patterns-dirs.js and three jsx suites that fail on main here too.

I also checked idempotency and meaning over 576 generated combinations of header shape, delimiter shape, indent width and container, comparing the parsed tree before and after formatting: no instability and no case where the escape changed what the document parses to.

Checklist

• I’ve added tests to confirm my change works.
• (If changing the API or CLI) I’ve documented the changes I’ve made (in the docs/ directory).
• (If the change is user-facing) I’ve added my changes to changelog_unreleased/*/XXXX.md file following changelog_unreleased/TEMPLATE.md.
• I’ve read the contributing guidelines.
• I did not use AI to generate this PR.
• (If the above is not checked) I have reviewed the AI-generated content before submitting.

This PR was written with AI assistance, and reviewed before submitting.