[vedanth-jadhav]

Cursor stabilization allocates a new ID String for every rewritten move and every click, even when the existing string has enough capacity. The GPUI editor invokes this after reading retained/live cursor data, and MultipleSegment::cursor_events uses it during project cursor loading.

Use String::clone_from for the two event-ID assignments. This reuses each event's existing string storage. Segment construction, duration/fallback selection, pointer protection, click boundaries, and final-click duration confirmation stay the same. Two production lines change; no public API or cache is introduced.

Release measurements

Baseline 5c1e9d1be, Linux x86_64 container, AMD EPYC 9V74, two-CPU quota, CPU 0 affinity, Rust 1.88.0. Root release profile (opt-level=s, thin LTO, one codegen unit). Five alternating independent processes per version; 25 samples after three warmups. Medians of process medians:

Fixture	Before	After	Time reduction
Single-shape control, 3,600 moves	0.009013 ms	0.008202 ms	9.0%
Long-lived shapes, 36,000 moves	0.148470 ms	0.129361 ms	12.9%
Brief shapes, 36,000 moves	0.184352 ms	0.144754 ms	21.5%
Brief shapes, 216,000 moves	1.732789 ms	1.429030 ms	17.5%
No pointer metadata, 36,000 moves	0.182740 ms	0.140557 ms	23.1%
Growing numeric IDs, 36,000 moves	0.184172 ms	0.169390 ms	8.0%
Legacy word IDs, 36,000 moves	0.185554 ms	0.159735 ms	13.9%

Fixtures use the recorder's current short numeric ID format (next_cursor_id.to_string()), with growing-ID and legacy-name controls. Dense movement is sampled at 60 Hz: 36,000 events represent ten minutes and 216,000 represent one hour of continuous movement. Brief shapes last 250 ms per four-second cycle; long-shape controls last two seconds. Click events are spaced at half-second intervals. These are synthetic component workloads, not average recordings or an end-to-end editor startup/FPS claim.

Cloning inputs, serialization/checksums, and output destruction are outside timing. Every sample is checked, and all baseline/head fingerprints match, including the metadata-free case (its dominant shape is unique). The unchanged one-shape control is small and its timing should not be interpreted as a causal benefit from the rewritten assignments.

A separate System-allocator probe measured 27,013 allocation/reallocation events before versus 6,313 after for 216,000 numeric-ID moves and 7,200 clicks: 20,700 fewer calls. This counter excludes input construction and is separate from the timing binaries. Retained ID-string capacities sum to 223,200 bytes on both versions for this case.

Capacity is intentionally reused rather than shrunk. The growing-ID control retains 122,850 ID-capacity bytes versus 111,600 before; the legacy-name control retains 267,150 versus 260,400. Arbitrarily long old identifiers can likewise retain spare capacity after shortening. No process-RSS or universal retained-memory reduction is claimed.

Correctness and validation

A new normal test compares complete event encodings with independently constructed expected events for equal-length, growing, UTF-8, empty, and long-to-short IDs. It verifies untouched modifiers, click fields, timestamps, NaN coordinate payload bits and negative zero. Existing pointer preference, long-shape preservation, final-click confirmation, and metadata-free fallback tests pass.

• cargo fmt --all and cargo fmt --all --check
• cargo check -p cap-project --locked
• cargo test -p cap-project --locked: 179 library + 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 example on both revisions, save the two binaries, then alternate five runs each:

cargo build -p cap-project --release --locked --example cursor-stabilization-benchmark -j 2
taskset -c 0 target/release/examples/cursor-stabilization-benchmark

The example prints per-sample API timings and complete-event fingerprints. Raw samples and experimental allocator instrumentation remain outside Cap. No native GUI/capture or macOS/Windows run was performed. The allocation reuse and disclosed capacity retention are the only intended behavior differences.

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.

Summary

Reuses existing cursor_id storage with String::clone_from when updating moves and clicks.

• Cursor stabilization reuses ID strings for rewritten moves and clicks.

_{Reviews (1) · Last reviewed commit: "perf(project): reuse cursor ID strings d..." · 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.

[vedanth-jadhav]

CI triage for this head:

• Intel macOS GPUI job: GitHub reports “The hosted runner lost communication with the server.”
• macOS real-capture job: fails in “Blacksmith capture permissions and audio device” setup, before the recording tests run. The verification aggregate consequently fails too.
• The macOS sync log, now retrieved through the GitHub connector, identifies audio::tests::a_stale_render_is_released_once_the_next_render_finishes: the supposedly cached paused render is already absent.

I reproduced a cache-retention/publication race on main and submitted the focused fix in #3219. Its deterministic regression fails on baseline and passes with the fix; all 215 editor library tests pass locally. It registers retention before a fast worker can finish, while preserving publication after successful worker creation.

Linux/Windows Rust tests and the Rust Clippy gates passed for this PR. Rerunning the infrastructure jobs still requires upstream Actions permission: both the CLI and GitHub connector reject the targeted rerun with HTTP 403, “Resource not accessible by integration.”