[vedanth-jadhav]

group_key_events clones every completed group's key vector, including each key's owned String, then clears the working vector. The GPUI editor calls this through generate_project_keyboard_segments when building keyboard overlays; legacy keyboard synchronization uses it too.

Move the key entries into an exactly sized output vector with Vec::append, retaining the working vector's allocation for the next group. This removes the deep key-text copies without repeatedly growing a fresh working vector. Group text handling, grouping decisions, modifier state, timestamps, IDs, and APIs are unchanged. The production change is three call-site substitutions and one small private helper.

Measurements

Baseline f510acfa1 against this branch. Linux x86_64 container, AMD EPYC 9V74, two-CPU quota, process pinned to CPU 0, Rust 1.88.0; repository release profile (opt-level=s, thin LTO, one codegen unit).

Five alternating independent processes per version, 25 samples per fixture after three warmups; table shows medians of process p50s. Positive reduction means faster. Fixtures model ten keys/second, typing hello separated by spaces; the shortcut variant presses/releases Meta periodically. Ten fixture sizes/modes were measured, from 100 to 36,000 keys.

Fixture	Grouping before → after	Reduction	Track generation before → after	Reduction
typing, 1,000 keys	0.111245 → 0.082202 ms	+26.1%	0.574179 → 0.583371 ms	-1.6%
typing, 36,000 keys	3.871336 → 3.287110 ms	+15.1%	25.169635 → 23.217586 ms	+7.8%
shortcuts, 1,000 keys	0.136932 → 0.121168 ms	+11.5%	0.670519 → 0.643980 ms	+4.0%
shortcuts, 36,000 keys	5.311675 → 4.919214 ms	+7.4%	27.031934 → 26.233404 ms	+3.0%

The 36,000-key fixture represents about one hour and 72,000 captured events (75,600 with shortcuts). The complete generate_project_keyboard_segments API measurement includes warm event-file reads, binary decoding, projection, and grouping. It excludes recording metadata loading, app dispatch, rendering, and output destruction. No native UI/FPS claim is made. The 1,000-key plain-typing generation control is 1.6% slower; small whole-path changes should not be overstated.

Every process serialized all grouped/generated tracks. All ten fixtures and both APIs match the baseline byte-for-byte across all runs (aggregate SHA256 df803d338e6e50f1ccabe2f095dfe96160de20f9657cfacb9c751cd407cd0ebd). In the one-hour plain-typing fixture, 6,000 completed words each contain five key strings, so this avoids 30,000 deep string copies while preserving output sizes and the scratch vector's capacity.

Correctness and checks

A new normal test covers group completion on a nearby space, a distant space, a new key after a pause, and a modifier. It pins the completed group's full serialized contents, key offsets and times after subsequent reuse, and verifies the source events are unchanged. Existing modifier, special-key, backspace, projection, binary compatibility, and timing tests pass.

• cargo fmt --all and cargo fmt --all --check
• cargo check -p cap-project --locked
• cargo test -p cap-project --locked: 179 library tests and 4 integration tests passed; 4 existing benchmarks ignored.
• cargo clippy -p cap-project --all-targets --locked -- -D warnings
• git diff --check

Reproduction

Use this branch's benchmark source on both revisions. It increases warmups/samples and selects the appropriate p95 sample for the larger sample count; the comparison above uses p50.

cargo build -p cap-project --release --locked --example keyboard-grouping-benchmark -j 2
CAP_KEYBOARD_BENCH_OUTPUT=/tmp/keyboard-output.json   taskset -c 0 target/release/examples/keyboard-grouping-benchmark

Save baseline/head binaries separately, alternate five processes each, compare the output files, and take medians of the p50 columns. The existing --check-scaling option remains available. Raw samples/outputs are kept outside the production repository.

Risk is limited to ownership during group completion. Vec::append empties the accumulator and transfers the original owned entries; later clearing/reuse cannot modify completed groups. No cache, concurrency, or public data layout changes. This is separate from the merged modifier-scan/indexing optimizations and the pending key-symbol/keycap UI changes.

Publication validation: rebased onto ec15b9840; 183 scoped tests passed, plus formatting, check, and strict all-target Clippy. Benchmarks above were recorded before this rebase; the production paths changed by this PR remain as measured.

Confidence Score: 5/5

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

Summary

Moves completed groups’ keys instead of cloning their strings, while keeping the working vector’s allocation.

• Completed keyboard groups move their key entries instead of copying them.

_{Reviews (1) · Last reviewed commit: "perf(project): move completed keyboard k..." · 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.