Make a table survive its own source

[luca-chen198]

One real note surfaced three defects in the same subsystem: it rendered as a completely blank page.

The blank page

The note's only content was a 14-line table, so the whole document collapsed into one hidden source run. A collapsed block hides its raw text by shrinking it to a near-zero font and kerning the rest of its width away — but .kern applies to every character in its range, so a run of n characters ends up width + n * kern wide. Feeding it the run's FULL width overshot by (n − 1) widths.

AppKit clamps most such lines to zero width, which is why this never showed. One line in that note held z̄ — a z plus U+0304 — and that line did not clamp: it laid out at x = +1783 with width −1783, dragging usageBoundsForTextContainer to −1778…695 on a 700pt container.

Measured, one variable at a time:

document	usage bounds
the note	(-1778.2, 0, 2473.2, 655.5)
same note, combining mark removed	(5.0, 0, 690.0, 655.5)
same note, after this fix	(5.0, 0, 690.0, 673.5)

Two more the same note found

\| truncated its row. GFM escapes the delimiter, and the escape beats every inline context because rows are split before inline parsing runs — so `Var[x(1) \| z(1:3)]` is one cell. Splitting on the raw character cut the row short at the header's column count and silently dropped the rest of the line.

<br> had no meaning. It is the only in-cell line break the format has (a row is one source line), and it drew literally. Return inside a row could therefore only tear the table in half.

Return in a table

Return — and ⇧↵, which AppKit sends as the same insertNewline: — now inserts <br> instead of splitting the row. The table's outer edges deliberately stay a plain newline: that is the only way out of a table sitting at the end of a document.

Tests

7 new tests across TableHiddenSourceGeometryTests and TableCellSourceTests; each was confirmed red against the unfixed sources. Full suite: 454 passing.

🤖 Generated with Claude Code