[richiemcilroy]

Word-highlighted captions (active_word_highlight with the colour style) rebuilt a Buffer, ran full Shaping::Advanced shaping and dropped the layout cache on every frame. Plain captions (4577e9d) and the text overlay (164bfef) already reuse the last frame's layout when their inputs are unchanged.

Change

• The highlight path now goes through the same cache as plain captions. The cache key is everything that reaches set_text / set_rich_text: text, font size, wrap width, family, weight and the colours. For highlighted captions the colours are the exact rich-text spans (byte range plus the u8 RGBA colour of each word and gap), so the active word, the transition blend and the caption fade are all in the key.
• On a hit the previous Buffer is reused. On a miss it is built with the same calls, in the same order, with the same arguments as before. Output is unchanged by construction.
• Shaping moved into shape_caption_text and a small CaptionText cache so it can be tested without a GPU. The per-word needle no longer clones when uppercase is off.

Why not shape once and recolour at draw time: glyphon 0.9 reads each glyph's colour from LayoutGlyph::color_opt, which cosmic-text 0.14 sets during shaping and exposes only through & accessors. Recolouring would need a fork of either crate. Keying on the u8 colours keeps output identical. Shaping is skipped on every frame where no word's colour or alpha changes: between transitions, once a fade-in has settled to its final u8 value, and on held words.

Tests

• shaping_tests::cached_word_highlight_layout_matches_an_uncached_shape_every_frame (CPU, runs in cargo test): steps a highlighted caption at 60 fps through fade in, every word change and fade out, switching wrap width every 45 frames, for three variants (default, uppercase with Unicode and a gold highlight with a short transition, zero transition). Every frame it compares each laid out glyph (position, font, glyph id, colour, cluster range) from the cached buffer with a buffer built by a verbatim copy of the old per-frame code. It also asserts the cache reshapes on fewer frames than it draws.
• shaping_tests::word_highlight_cache_reshapes_when_any_layout_input_changes: text, uppercase, font size, wrap width, family, weight, highlight colour, transition duration, active word and caption fade each force a reshape, the result matches the uncached build, and identical inputs do not reshape. Also covers switching between plain and highlighted text.
• GPU tests (ignored, need an adapter): every_caption_frame_renders_like_a_fresh_layer gains two highlighted variants (gold, outline, uppercase, pop; zero transition with a long fade). New highlighted_frames_follow_output_size_changes_like_a_fresh_layer changes the output size every 7 frames. reused_layer_follows_style_and_output_size_changes gains nine highlighted cases. All compare rendered pixels with a fresh layer.
• New ignored bench benchmark_word_highlight_prepare times CaptionsLayer::prepare over 600 frames at 1080p and prints an FNV digest of every rendered frame, so base and head output can be compared directly.

Measurements

benchmark_word_highlight_prepare, 1920x1080, 600 frames at 60 fps (2204 visible samples over 4 timed passes), base (main plus the bench commit) and head built with the same release profile on the same runner and run interleaved, 6 rounds each. Numbers are the median of the per-round total CaptionsLayer::prepare time for 600 frames. Scratch run: https://github.com/CapSoftware/Cap/actions/runs/39538547724

Case	Linux base	Linux head	Change	macOS base	macOS head	Change
plain (no highlight)	5.47 ms	5.54 ms	+1.1%	15.62 ms	17.90 ms	noise
colour highlight (default white)	38.70 ms	32.42 ms	-16.2%	49.88 ms	38.50 ms	-22.8%
colour highlight, gold, outline	56.39 ms	53.48 ms	-5.2%	62.48 ms	58.76 ms	-6.0%
colour highlight, back-to-back words	38.50 ms	27.46 ms	-28.7%	50.51 ms	36.14 ms	-28.4%

Linux is blacksmith-8vcpu-ubuntu-2404 (lavapipe), macOS is blacksmith-6vcpu-macos-26. The macOS VM drifts a lot between rounds (plain ranged 7.7 to 20.3 ms on both sides), so read its plain row as unchanged; Linux is stable and shows the plain path is unchanged.

The gain is bounded by how often no colour changes between frames. With the default 0.25 s word transition, a word's fade in keeps changing its u8 colour for most of the word, and a gold highlight changes the blue channel by 255, so it settles even later. Back-to-back words (the usual transcript shape, no fade out between words) hit the cache most.

Pixel output: the bench hashes every rendered frame. Base and head digests match for every case on both platforms (Linux: plain 46a57c3a96ae770f, highlight 07e8a12b482c5927, gold outline d0669b00b3f7ff1a, back-to-back 0070e1ef26b7663b; macOS: ab266f7b1fd93bfb, 5185a2325fe0539d, 741779ddcc4df37c, c76e43bfd068fbeb). The ignored GPU tests pass on both runners in the same run.

CI

• Engine verify FULL dispatched for captions, export and playback on Linux, macOS and Windows: passed (https://github.com/CapSoftware/Cap/actions/runs/39539519286).
• PR Engine verify: only the known macOS crash/recorder-kill/*/static-screen fragment-span flake fails.
• sync-tests fails on main too: crates/rendering/tests/display_layout.rs still passes &ProjectConfiguration to ProjectUniforms::new (test added in 19de9e5). Not touched here.

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

Highlighted captions now reuse their shaped text when the text, font settings, width, and exact span colours stay unchanged.

• Word-highlighted captions reuse their layout while its inputs stay the same.

_{Reviews (1) · Last reviewed commit: "perf(rendering): reuse word-highlighted ..." · Reviewed by Greptile}

[blacksmith-sh]

Found 2 test failures on Blacksmith runners:

Failures

Test	View Logs
cap-verify::crash/crash_safety	View Logs
cap-verify::crash/crash_safety	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[richiemcilroy]

@greptileai review

[richiemcilroy]

@greptileai review

[capy-ai]

Capy found no issues.

View 1 other finding in Capy.