perf(rendering): reshape captions only when their text or style changes (up to 7.6x faster)

[notkainoa]

Captions are up to 7.6x cheaper per frame. The captions layer rebuilt and reshaped the caption on every frame, even though in most styles the text and color stay the same for the whole segment. Now it keeps the shaped text and reshapes only when it changes.

Caption style	Before (p50)	After (p50)	Speedup
default	56.7 µs	7.5 µs	7.6x
pill highlight	60.8 µs	11.6 µs	5.2x
outline	84.2 µs	34.8 µs	2.4x
color highlight	65.3 µs	65.5 µs	same (unchanged path)

Means: default 55.4 → 11.9 µs, pill 60.1 → 16.0 µs, outline 80.5 → 36.8 µs, color highlight 63.9 → 64.3 µs.

Runs on every frame of GPUI editor playback and export of any recording with captions.

What changed

• The shaped buffer is kept, and rebuilt only when the caption text, size, wrap width, font, weight or color changes.
• Per-word color highlighting still reshapes every frame, because its word colors really do change. That path is untouched.
• Fade frames still reshape. The fade alpha is baked into the shaped glyphs, and the outline passes draw from the same buffer, so reshaping is what keeps the pixels identical to main. Those frames are the remaining p95.

Output is unchanged

• 29,615 rendered frames hash identical to main with perf(rendering): reuse shaped text overlay layouts between frames (6.7-8.2x faster) #2524, perf(rendering): reshape the keyboard overlay only when its text changes (2.8x faster) #2525 and perf(rendering): reshape captions only when their text or style changes (up to 7.6x faster) #2526 applied together. Covers all 14 text animations, text styles, mid-playback edits, keyboard and caption styles, playback, backward scrubbing, random seeks, pauses and resizes. A planted cache-key bug makes thousands of frames differ, so the check catches mistakes.
• New GPU test: for default, outline + uppercase + pop, pill highlight and color highlight captions, every frame renders byte-identical pixels to a fresh layer.
• A second GPU test changes the font size, family, weight, text color and output size on a warmed layer, and switches between color highlighting and plain text on the same layer. Removing font size, family, weight or color from the cache key, or the reset after color highlighting, makes it fail. (Caption wrapping is off, so the wrap width never changes the shaped text.)
• cargo test -p cap-rendering --release --lib: 305 passed on current main. The GPU tests pass with --include-ignored.
• cargo clippy -p cap-rendering --all-targets -- -D warnings and cargo fmt are clean.

Benchmark setup

CaptionsLayer::prepare at 1920×1080: 4 spoken sentences at 60 fps. One warmup pass, then 3 measured passes pooled (824 visible frames per pass, 2,472 total). Each number is the median of 3 runs. Apple M1, release build, opt-level 3. "Before" is the same benchmark on main (3d19c5c).

CARGO_PROFILE_RELEASE_OPT_LEVEL=3 cargo test -p cap-rendering --release --lib benchmark_captions_layer_prepare -- --ignored --nocapture

Confidence Score: 5/5

The PR appears safe to merge; no actionable defects remain.

Summary

Keeps shaped caption text between frames and rebuilds it when the text or style changes.

• Caption rendering reuses shaped text until its text or style changes.

The previous and current commits have different merge bases with the target commit, so this review used the full PR diff.

_{Reviews (2) · Last reviewed commit: "test(rendering): check the caption layou..." · Reviewed by Greptile}

[vercel]

Someone is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[richiemcilroy]

@greptileai review