Trail table row permalinks after cell text

[tony]

On a prose table such as the "Stable identity, not name or index" table on the Rust server, session, window and pane page, each row's ¶ permalink sat at the right edge of the first cell, a long way from the label it links. The first column's text also touched the table edge.

Rows now show the ¶ immediately after the first cell's last line of text, with a small gap, hidden until the row is hovered, focused or targeted, as the section ¶ behaves. Markdown table cells share one 0.75rem inline inset, first and last columns included.

What changed

• site/src/styles/global.css: .row-anchor-link keeps its static position at the end of the cell's last line and leaves the flow, so it never widens the column or wraps onto its own line (a --expand-environment flag wrapped at 390px in an in-flow version). It uses the section permalink's theme colors and the row's :hover, :target and :focus-within reveal that .row-label-link already uses.
• Typography zeroes the outer padding of the first and last column from the utilities layer. A rule on .prose table:not([class]) in that layer, the way the heading overrides do it, replaces it. Component tables carry a class and keep their own padding.
• site/scripts/check-dev-sample.mjs: checkRowPermalinks, run by checkDevSample (--sample) and by the full check-dev.mjs flow. It loads the Rust, tmux and Ruby workspace CLI tables in light and dark at 1440, 768 and 390px and checks every anchored row:
  • the permalink starts 0.1em to 0.5em after the end of the text, on the last line box;
  • the first column's text has the same left inset as the other columns, and as its header;
  • the permalink is invisible at rest and appears on hover, focus and :target with at least 3:1 contrast against what is behind it.
• CONTRIBUTING.md lists row permalinks among the browser checks.

Verification

• The gate fails on origin/main CSS with the named failure on all three pages, for example row "Session" permalink starts 39.8px after its text (Rust and tmux) and row "--json" permalink starts 100.7px after its text (Ruby). With only the ¶ rule applied and the cell inset removed, it fails on first-column text sits 0.0px from the cell edge. It passes with this change, in Chromium, Firefox and WebKit.
• The gate also passed on four additional pages (Rust transports, TypeScript queries, Python options-and-hooks, Rust batching-commands): 46 rows on seven pages in all.
• pnpm run lint is clean. site/test/row-anchors.test.ts, prose-markdown.test.ts, declaration-body-spacing.test.ts and header-pickers.test.ts pass (100 passed, 4 skipped); two prose-markdown link tests hit vitest's 5s timeout while the host's load average was above 90 and pass with --testTimeout=120000.

Not run

• node site/scripts/check-dev.mjs --sample did not get past its first step. checkPickerFilters times out on its 5s page.goto of /demo/widgets/pickers/ on this host (load average 50 to 110) before reaching the new check. That step does not touch this change, and it failed the same way with the origin/main stylesheet. The wiring is checked by node --check and by importing the module; the check itself ran against a static page that renders the same Markdown through rehypeRowAnchors with this branch's compiled stylesheet, and against the Rust page in a live dev server before it was stopped.
• The runtime of the sample loop with the new check is unmeasured; the check adds two cold page compiles (Rust, Ruby) to the sample.
• Style-parity and visual baselines need an assembled _site, which was not built. The visual baseline pages (API reference pages) contain no .prose tables.

[tony]

Folded into #157, which now carries every fix from the 2026-10-18 review on one branch so they share one CI run, preview and deploy. This commit is on that branch unchanged. Discussion and evidence stay here.

[tony]

Correction: one line differs on #157. The check-dev.mjs import of check-native-layout.mjs there also brings in #162's alignment checks, so both changes import cleanly.