[EDToaster]

Stack 4/5. Based on #63. It makes each frame cheaper to build in toastty-render: shaping rows and building GPU instances. The rendered output is unchanged.

What changed

• Dense char table for the shaping fast path (text/glyph_rasterizer.rs). Characters below U+3000 cost one table load. Before, each one took up to three SipHash set lookups. The sets stay the source of truth, and a test checks every table entry against the old logic.
• Row shapes memoised by their exact text (shape_line_memo, new text::row_shaper::RowShaper).
  • There is one memo per mode-2027 setting, trimmed to about two screens of rows each frame.
  • It is cleared when an old shape could be stale: a char found to ligate or become uncacheable, a cache miss, an atlas eviction, or a font-size/DPI change.
  • A row with a glyph that got no atlas slot is not memoised, so it is retried, as before.
  • Repainting unchanged CJK text drops from 33 ms to 0.18 ms.
• No background quad for default-bg cells on full-clear frames. The clear already painted that colour, so this roughly halves the instances on a text screen. It is gated by may_skip_default_bg, which is off whenever edge bleed is active. It also replaces the old drop-default-bg pass for images drawn below cell backgrounds.
• sRGB → linear is a 256-entry table holding bit-identical values. Truecolor cells used to cost six powf each.
• Faint (SGR 2) colour is remembered across a run, instead of costing nine powf per cell.
• Blink detection is done in the row-text pass instead of a second scan.
• Bench.
  • bench/paced.py plus the paced_text, paced_tui, paced_cjk and *_big scenarios draw one full-screen synchronized frame every 20 ms, so the renderer is actually timed under load.
  • render_term gains TOASTTY_BENCH_CPU_ONLY=1 / TOASTTY_BENCH_MIN=1, which report the min and median of 400 CPU-only frames and stay usable on a loaded machine.

Review

An independent reviewer read the branch and found three problems. All are fixed here and confirmed by a re-review.

1. Dropped: cosmic-text's shape-run-cache feature. It was in an earlier version of this branch.
   • Its key leaves out text direction, so a word made only of neutral characters, shaped inside an RTL line, came back mirrored in a later LTR line (the cells rendered blank). The row memo then kept the bad shape.
   • Cargo feature unification also turned it on in toastty-markdown, where nothing trims it.
   • Regression test: rtl_neutral_word_does_not_leak_into_ltr_row. It fails with the feature forced back on.
   • Cost: new CJK text in every row is back to about the base cost (see the table).
2. The default-bg skip changed pixels under edge bleed. This happened with extend_background_when = always, scrolled back by part of a row: the bleed quads overlap neighbouring rows. It is now gated off whenever bleed is active.
   • tests/skip_default_bg.rs covers padding, bleed, a partial top row and a bottom sliver, in both Rgba8Unorm and the app's Bgra8UnormSrgb.
   • The test also asserts that forcing the skip under bleed does change pixels.
3. The row memo no longer freezes degraded shapes. These are atlas-slot failures and chars that later become cache misses.

Numbers

CPU per frame, shaping + instance building only (TOASTTY_BENCH_MIN=1, min of 400 frames):

scenario	before	after	speedup
repaint 200×60	497 µs	181 µs	×2.7
repaint 300×120	1498 µs	549 µs	×2.7
new text every frame, 256-colour, 200×60 / 300×120	951 / 2852 µs	329 / 1000 µs	×2.9
new text, truecolor, 200×60 / 300×120	1001 / 2997 µs	313 / 948 µs	×3.2
new text with CJK/emoji, 200×60 / 300×120	10.3 / 32.2 ms	9.4 / 29.8 ms	×1.1
repaint of faint text, 200×60	1796 µs	234 µs	×7.7
repaint of CJK text, 200×60	33.3 ms	0.18 ms	×190
CJK scroll, one new line per frame, 200×60	32.9 ms	0.36 ms	×92

The 256-colour rows overstate the base cost a little, because the bench passes no palette table.

End to end (bench/run.py, Xvfb, 3 reps back to back, compared with #63):

scenario	frame ms p50	frame ms p99	CPU s
paced_text (128×42)	0.90 → 0.74	1.28 → 1.08	0.57 → 0.53
paced_text_big (256×73)	2.10 → 1.65	2.71 → 2.21	1.53 → 1.41
paced_tui	0.84 → 0.67	1.13 → 0.91	0.54 → 0.48
paced_tui_big	1.74 → 1.30	2.49 → 1.67	1.37 → 1.19
paced_cjk	5.54 → 5.39	7.00 → 7.26	1.99 → 1.96
scroll_wheel_big	1.12 → 1.07	overlaps on re-runs	1.04 → 0.98

• Unchanged: idle CPU, idle memory, startup and key → frame p50.
• Key → frame p99: 0.1 ms higher in one set of runs (1.56 → 1.77 ms, 133 samples). An interleaved run against perf 5/5: startup on worker threads (first frame 82→49 ms), binary −1 MB #65 measured 1.70 vs 1.61 ms, so it is within noise. The reviewer timed the keystroke path in isolation: it costs a few µs at most.
• About 0.7 ms of each Xvfb frame is software presentation, which the test setup forces, so on a real compositor the gains take up a larger share of each frame.

Visual evidence

Built from #63 and from this branch, run on Xvfb with identical scenes, and compared with magick compare -metric AE. All 14 pairs differ by 0 pixels: 7 scenes, each in the default config and in an 8 px padding + extend_background_when = always config. The scenes are colours/attributes, CJK/emoji/box drawing/combining marks, cursor over a glyph, selection, scrolled back by 1 and 2 notches, and back at the bottom.

Text scenes, before (#63):

After (this branch), 0 differing pixels:

Scrolled back with padding and edge bleed on, the case the review caught. Before (#63):

After (this branch), 0 differing pixels:

For reference, the pre-review build (default-bg skip ungated) against #63 on that scene: the differing pixels are in red along the bottom partial row, 756 px in total. The shipped build differs nowhere.

Testing

• cargo test --workspace --no-fail-fast: 2426 passed, 0 failed.
• cargo clippy --workspace --all-targets: no new warnings compared with perf 3/5: parser + grid hot path — 2–7× parse, lazy scrollback (idle RSS 101→60 MB) #63.
• cargo fmt --check is clean for toastty-render.
• Every commit builds on its own.