Large documents stay interactive: the canvas re-renders only what a pan uncovers or an edit
touches, the layers panel lays out only the rows in view, and per-frame work no longer walks the
whole page.
- Benchmark: `cargo xtask bench [SIZES]` builds synthetic pages (`figcraft_engine::synthetic`:
screens with an auto layout header, a card group, button instances and shapes/text) at
1k/10k/50k layers and prints engine/render timings (`figcraft-cli perf --layers 1000,10000`) and
headless whole-UI frame timings (an ignored ui-egui test).
- Retained canvas rendering (`figcraft_render::Retained`, used by the render worker): views snap
to whole device pixels, so a pan shifts the previous image in place and renders only the
uncovered strips; an edit renders only where changed layers were and are
(`Document::changed_ids` diffs the two versions skipping shared structure; ancestors that paint
from their children's pixels - booleans, effects, masks - count whole; background blur reach
and noise are allowed for). Zoom, page, options, font or page-level changes, or large changes,
render the whole view as before. `ui.inspect` reports `canvas.renderWork`.
- The texture image is converted on the render thread, not the UI thread.
- Layers panel: only visible rows are laid out (stable ids while scrolling); rows are cached per
document version, page, expanded set and filter.
- Frame labels and layout grids of off-screen top-level frames are skipped; the Prototype tab's
connection sources are cached per document version.
- Text: the layout cache keys by a cheap fingerprint confirmed by `TextData` equality (no Debug
formatting per lookup, 32k entries) and hands out shared layouts; a layout generation marks
renders stale when fonts arrive. Overflowing text is culled by its glyphs, so it no longer
vanishes while its box is off screen.
Before -> after (release, Windows, 1440x900 view, ms; 10k / 50k layers):
- move one layer, re-render @100%: 7.65 / 6.49 -> 0.18 / 0.81
- move one layer, re-render zoomed out: 11.95 / 24.10 -> 0.16 / 0.77
- pan 40 px @100%: 7.53 / 6.82 -> 0.69 / 0.81
- pan 40 px zoomed out: 12.07 / 24.30 -> 1.62 / 4.17
- full render @100% / zoomed to fit: 7.78 / 44.22 -> 6.56 / 38.83 (10k @100%, 50k fit)
- UI frame, every screen expanded in layers: 6.16 / 39.07 -> 0.24 / 0.57
- UI frame, scrolling that panel: 6.11 / 40.18 -> 0.43 / 0.40
- UI frame, layers filtered by name: 1.03 / 6.92 -> 0.17 / 0.36
- UI frame, Prototype tab: 0.89 / 6.45 -> 0.18 / 0.80
- UI frame, idle: 0.31 / 1.43 -> 0.11 / 0.39
Unchanged and within budget: single-layer edits through the engine ~0.01 ms, normalize after a
one-node edit < 0.01 ms, hit test 26 / 93 us. Left: save/load at 50k (~0.7 s each), zoomed-out
renders at 50k (~40 ms).
Tests: retained renders match full renders after pans, moves, recolours, hide/remove/add, undo,
masks, overflowing text, edits under a background blur and an inner-shadow/noise layer, a drop-
shadowed group, and a blur panned in from the view's edge; zoom/page/options/page background
render full; junk views don't panic. Layer-row and prototype-source caches follow edits, undo,
page switches, expansion and the filter. `changed_ids`, the text cache, and the perf CLI.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Text runs get bidi embedding levels from the Unicode bidi algorithm
(unicode-bidi, MIT/Apache-2.0). Right-to-left runs are shaped right to
left, kept in logical order for line breaking and placed in visual order
per line (rule L2, trailing spaces at the paragraph level). Glyphs and
lines know their direction, so the caret, clicks and selection
highlights follow right-to-left runs. Text without right-to-left
characters skips the analysis.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
