[foxnne]

What changes

A markdown table whose cells wrap over many lines no longer changes height as the reader scrolls into it. The scrollbar no longer twitches by a line when paragraphs scroll back into view, either. The table-heavy test sample is now a generated document whose tables actually wrap. Both integration tests that guard this failed on it before the fix, and the old frozen plan copy hid the bug.

Why it moved. Everything traces to one dvui behaviour: a text layout on its first frame wraps at an assumed 1000pt, not at the column, so it measures short. The preview's bookkeeping believed those measurements in four places:

1. Table cells settled short. A row measured only on budgeted frames, with skipped frames between, loses its text layouts in every gap, so every draw is a first frame. Two such draws agree perfectly, and the cell settled at the short height. A table re-entering the draw set also overwrote settled cells with short first-frame sizes. On the synthetic sample's last table, the final row settled 59pt short: 1789 measured off screen, 1730 on screen. Now a first-frame cell measurement never counts as a draw and never replaces a real size. It is kept (provisional) only when there is nothing better.
2. Rows lagged a frame behind their cells. The grid sizes a row from what its cells reported last frame. With (1) fixed, the cache stayed right but the grid did not, and a settled table measured short, cleanly, on the frame after a first-frame draw. Now a drawn cell with a settled size at this width is held to at least that height, as a culled cell already was.
3. A pinned block was measured by its pin. The wrapper emits exactly the pin, so a .deferred table came out of deferral with its untrusted height filed as a clean measurement. Now any pinned block is measured by its unconstrained inner box, which only pinned-above-the-fold blocks used to be. That also makes it safe to pin every block with a believed height, not only tables. A paragraph scrolled back into view emitted a line short for a frame. That was the remaining 19.6pt in "rapid scrolling up and down".
4. .measured meant two things. It meant both "clean at this width" and "laid out at a width that is gone". record must throw the second away for a contaminated measurement, so it threw the first away too. Stale heights are now their own state, .stale. invalidateForWidth also no longer turns a never-measured .estimated slot into a "measurement" of its zero placeholder.

The task this came from guessed at the table estimate for wrapped cells. Instrumenting the failing block showed the estimate was never involved: the table settled at 1789.1 from a clean off-screen measurement, and the on-screen layout was the one that was wrong.

The sample. tests/data/markdown_sample_tables.md is now generated by tests/data/markdown_sample_tables.py (seeded, so it reproduces exactly). It has one 48KB three-column table whose cells wrap over many lines, three smaller tables, and prose. After the tables come blocks that repeat word for word, so source hashes are not unique (the old sample had seven, which the traversal test relies on), and one long paragraph for bench-markdown's "prose" sample. Those come after the tables on purpose: content inserted above them shifted which scroll frames see the big table, and that alone hid the block-heights failure. The frozen plan copy at that path is gone. bench-markdown's "prose" sample now uses this document as well.

SDK impact

• None

Verified

• macOS: all five gates (zig build, zig build test, zig build test-integration, zig build check-web, zig build test-sdk-version), re-run after rebasing onto main past docs: plans move to the private fizzyedit/plans #296.
  • On the parent, the new sample fails both tests: block 29 changed height while scrolling: 1789.1 -> 1730.4, and a rapid-scroll total of 18967.6 against 19476.0. With the fix, all 17 markdown preview integration tests pass.
  • With the fix, the 4-column **Row N** variant of the sample also passes all 17 (reported failing with 2610.4 -> 2297.5 before; I did not re-run that on the parent). So does the old frozen plan copy.
  • New unit tests in block_heights.zig: a contaminated measurement does not replace a clean one at the same width, and a width change leaves a never-measured block on its estimate.
• Windows:
• Linux: CI's test-integration
• Web: check-web compiles. Not run in a browser.

Not checked by eye in the running app. The evidence is the headless layout tests above.

Follow-ups

• Every block with a settled or deferred height is now pinned while drawn. A real change still gets through, one frame late: the inner box disagrees, the entry drops to .measured, and the pin comes off. That is the lag pinned-above-the-fold blocks already had.

🤖 Generated with Claude Code