reflow

Reflow

Seven studies after Cheng Lou's Pretext demos, on exact2: one Contract, one Rust data crate, Linux, macOS, iOS and web. Original prose, set in Exposure Sans Original 03 (three faces copied from apps/typetour/assets; MIT in the Expose repository).

Pretext has two halves and this app shows both without the library.

• Text around shapes is the kernel's (wrap-flow / shape-outside, LLP 1043.000): the balls, the dragon, the drop caps, the pull quote and the picture are exclusions every host flows text around.
• Heights before layout are the data crate's. data/build.rs reads every glyph advance and kern pair the prose can use out of the app's own font files at build time; data/src/typeset.rs breaks lines by arithmetic over those advances with the same exact-textflow walker the hosts flow text with. Cards, columns and ASCII rows are measured there, and the kernel receives text already cut to the lines the arithmetic found. No font is parsed at runtime; the wasm carries a table, not a font.

The wall retains prepared title/body advances for its current seed. Resizing recounts lines and balances the columns from those advances; widths producing the same column count and integral column width reuse the placement too. A new seed replaces the preparations, so the cache holds one 1,200-card deck rather than growing with resize history. Geometry still follows each new column width.

A release-mode data-query probe on this Mac (2026-09-20, baseline/candidate/ candidate/baseline order, seven trials per group) measured median queries over 240 one-pixel resize steps: 3.89/4.04 ms before and 0.056/0.055 ms after. Cold queries remained about 4 ms; retained allocator-requested storage grew from 442,487 to 3,117,367 bytes for seed 1. These are data-query costs, not rendered frame times. Answers matched the old implementation across 208 width/seed/scroll combinations. The data regression also checks all 1,200 card heights and positions against fresh measurements across column boundaries and seed changes; the Linux tests verify painted heights and windowed scrolling.

The spread measures its pull quote in the italic face at the current font size and padded width before cutting the article columns. Its painted box and exclusion share that height, and the picture moves down if the quote needs room.

Tab / stable id	What to try
01 Balls — scene-balls (opens first)	Three spheres (ball-1..3) cross balls-prose; the words part on both sides. Positions are arithmetic on the clock, committed every 16 ms.
02 Dragon — scene-dragon	Drag dragon anywhere on the manuscript page (pan). One polygon is both the ink and the exclusion.
03 Masonry — scene-masonry	36 cards measured in the bold and regular faces, each placed in the shortest column. The footer of every card says the height the arithmetic predicted; shuffle reseeds.
04 Dynamic layout — scene-magazine	Drag magazine-handle at the page's right edge. Columns are re-cut at line boundaries on every sample.
05 ASCII art — scene-ascii	A lit sphere in a ring, in a proportional face: each row advances by the kerned width of the glyph it chose. ascii-naive samples on a uniform grid instead and the picture bends.
06 Occlusion — scene-wall	1,200 cards, every position known up front; only the cards near the viewport (wall-N) exist as views. Scroll wall-scroll.
07 Spread — scene-spread	An editorial spread whose drop cap, pull quote and picture are exclusions; each column's text (spread-0..2) was cut where the same walker, run in the data crate around the same shapes, said the flow would stop. One column on a phone.

export EXACT_UPDATE_TRUST=development

# Web (open the URL printed by the dev server)
bun host/web/dev.mjs --app reflow

# macOS; append --ios for the simulator
bun host/apple/build.mjs --app reflow --run
bun host/apple/build.mjs --app reflow --ios --run

# Linux's CPU painter, headless on macOS too
cargo build --profile host-dev -p reflow-linux
bun scripts/agent.mjs linux --app reflow --size 1000x900 tree

# The agreement tests: cosmic-text's layout of every card, column and row
# against what the data crate measured.
cargo test -p reflow-data -p reflow-linux --release

Drive it (the same lines with macos, web or ios in the host position):

bun scripts/agent.mjs linux --app reflow --size 1000x900 \
  "clock +1200" "layout balls-prose" "screenshot balls.png" \
  "tap scene-dragon" "tap dragon down" "tap move by -160 120 over 200" "tap up" \
  "layout dragon-prose" "screenshot dragon.png" \
  "tap scene-masonry" "screenshot masonry.png" \
  "tap scene-magazine" "tap magazine-handle down" "tap move by -300 0 over 300" "tap up" \
  "screenshot magazine.png" \
  "tap scene-ascii" "clock +2000" "screenshot ascii.png" "tap ascii-naive" "screenshot naive.png" \
  "tap scene-wall" state "screenshot wall.png" \
  "tap scene-spread" "layout spread-1" "screenshot spread.png"

bun scripts/agent.mjs linux --app reflow --test apps/reflow/app.test.contract

Where the arithmetic can disagree with a host: advances are summed per glyph with kern pairs, so a ligature or a shaping rule a face applies across a segment boundary is a fraction of a pixel the table does not know. Text is therefore set 2 px wider than it was measured (the card padding) so a disagreement can only leave air at the bottom of a card or column, never clip a word. Verified 2026-09-20 on this Mac: the Linux tests above pass with exact equality (36 card heights, three column line counts, two flowed columns with no clipped band); on macOS the same scenes match the screenshots' predictions and the spread's columns end inside their cut; on the web (headless Chrome) flowed text goes through the shared walker over canvas advances and the spread's two columns end at bands 45 and 44 of 46, while ordinary paragraphs are broken by the browser and a magazine column can end one line early. The macOS binary was ad-hoc signed (EXACT_IDENTITY=-) because the keychain's team identity refused codesign from a non-interactive shell (errSecInternalComponent). iOS ran on the iPhone 17 simulator (balls, spread at one column, ASCII), driven by the same agent lines.