[OptimistOtaku]

GPUI's Whisper and Parakeet caption paths pass owned word vectors to normalize_caption_words, which allocates a second vector and copies every retained word's text. Normalize the existing vector and strings in place, then attach punctuation before applying the existing duration cap.

This retains Unicode trimming, leading/consecutive punctuation, word order, start/end float bits, caption text, and chunk boundaries. Punctuation may still grow the preceding string. No dependencies, public API, project format, or platform-specific code change.

Measurements

Native Windows x86_64, Ryzen 5 5600H, Rust 1.88 GNU, repository release profile (opt-level=s, thin LTO). Median of nine alternating paired samples after warmups. Input batches are cloned before the timer; timings include normalization and dropping the complete output. The benchmark compares the exact upstream function against production code and checks full output equivalence before timing.

Input words	Whitespace padding	Before	After	Speedup
32	No	2.233 µs	1.623 µs	1.38×
32	Yes	2.422 µs	1.356 µs	1.79×
1,000	No	94.226 µs	66.918 µs	1.41×
1,000	Yes	86.839 µs	52.445 µs	1.66×
12,000	No	1.151 ms	0.715 ms	1.61×
12,000	Yes	1.021 ms	0.533 ms	1.92×

A second complete run also improved every case. These are normalization timings, excluding speech recognition, model loading, and UI work.

cargo test --locked -p cap-project --release --lib benchmark_caption_word_normalization -- --ignored --nocapture

Validation

• cargo fmt --all, cargo check --locked -p cap-project
• cargo test --locked -p cap-project: 196 library tests and four integration tests passed; five benchmarks ignored by default, new benchmark run separately.
• cargo clippy --locked -p cap-project --all-targets -- -D warnings
• git diff --check

Differential cases cover Unicode/ASCII whitespace, empty tokens, punctuation chains, multilingual text, signed zero, NaN/infinite times, duration clamping, rendered text, and grouping. A separate test verifies reuse of vector and string allocations. Native GPUI launch is unverified locally. Fork Actions dispatch is disabled by GitHub; upstream PR CI will provide additional platform validation.

Confidence Score: 5/5

The empty CI-retry commit appears safe to merge.

Summary

The latest commit is an empty CI retry. Source files are unchanged since the previous review.

• Caption normalization reuses its word storage without changing captions.

_{Reviews (2) · Last reviewed commit: "chore: retry CI after macOS runner disco..." · Reviewed by Greptile}

[vercel]

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

A member of the Team first needs to authorize it.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[richiemcilroy]

Thanks! Can you get CI green? Then we'll take a look.

[OptimistOtaku]

Thanks Richie. The failed macOS engine shard reported a runner disconnect, so I pushed a code-identical commit (c9acbc01b) to trigger fresh CI; rerunning the upstream job directly requires maintainer permissions.

The fresh engine tests, all four GPUI builds, Clippy, sync tests and recording checks have passed. The performance measurement job is still running, so I can't call CI fully green yet. The two Vercel checks also require authorization from a Cap team member. No source changes were needed for the runner failure.

[richiemcilroy]

Thanks! Can you get Greptile to 5/5 on the latest commit? Then we'll take a look.

[OptimistOtaku]

@greptileai please review the latest commit, c9acbc0, as requested above. It is an empty CI-retry commit; the source changes are identical to the previously reviewed commit.

[OptimistOtaku]

The fresh Greptile review is now 5/5 on the latest commit, c9acbc01b. All code checks have passed too, including the performance budget. The only remaining red checks are the two Vercel previews requiring Cap-team authorization. Thanks for taking another look.