Reduce release build time and binary size

[John-David Dalton (jdalton)]

LLM Description written by Codex: GPT-6

User impact

Release builds can complete faster and produce a smaller CLI binary. The repeated hosted measurement on Rust 1.93.1 reported median build times of 350.780s with the default settings and 333.036s with one codegen unit, a 5.06% ratio-of-medians reduction. The measured binary was 16.27% smaller.

Change

Set [profile.release] to one codegen unit. Set ci-release, perf, and addon explicitly to 16 to preserve their current inherited configuration.

Evidence and limits

The repeated hosted run is here. It used Ubuntu 24.04 x86_64 and Rust 1.93.1. Its source revision predates the current main source tree; current main has changed Rust code since the measurement, so exact build-time and size results may differ. This measurement does not claim runtime improvement or results for other release platforms.

Exact-head validation

On macOS 27 arm64 with Rust 1.93.1 and clean, separate targets, the candidate release build completed in 210.836s versus 220.474s for the same commit with a codegen-units=16 override. Compiler invocations confirmed codegen-units 1 and 16 respectively. This is one pair (4.37% faster in that pair), not repeated performance evidence. Binary sizes were 18,978,624 versus 22,864,448 bytes (17.0% smaller). --version and --help output matched. The release-mode golden tests passed: 1 redirect fixture test and 2 SBT golden tests.

CI

The CI run 38307581268 has ci-ok failing because one Cargo fixture test could not connect to static.crates.io to download cfg-if; 38 tests in that job passed and one failed on the download. The failure log is here. The change does not modify Cargo.lock or dependency configuration.

---

Note

Low Risk
Build-profile-only change with no runtime logic; main risk is longer release compile times or different optimization on non-overridden profiles if inheritance is misconfigured.

Overview
Tunes Rust release builds by setting codegen-units = 1 on [profile.release], aiming for a smaller shipped CLI and modestly faster release links (per hosted benchmarks in the PR description).

Because ci-release, perf, and addon inherit release, the PR pins codegen-units = 16 on those profiles so CI test shards, perf/bench builds, and the Node addon keep faster parallel codegen and do not pick up the single-unit release setting.

^{Reviewed by Cursor Bugbot for commit a51923e. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final reviewer: I turned auto-merge off on a51923e. ci-ok is red because the cargo 1.93.1 lock-v4 (ubuntu-latest) job failed (job log). The PR also has no approval and no Ready for review label. Auto-merge will be turned back on once CI is green and the PR qualifies.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Ready for review (burn-down agent)

• Head: a51923e414f05789da6c7570c7169656368b06ea
• CI: 35/35 green, 19 path-skipped. The cargo 1.93.1 lock-v4 (ubuntu-latest) failure the final reviewer flagged passed on re-run (run 38307581268, attempt 2).
• Bugbot: reviewed this head, no findings.
• Reviewer: one-line Cargo.toml profile change. Shipped release builds use codegen-units = 1, and the ci-release/perf/addon profiles stay at 16 so CI link times don't grow. The timing evidence in the description is a single pair of runs.
• Slack announcement not sent: the Slack connector in this session has no send tool.

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

Final review brief

What it does: Sets codegen-units = 1 on [profile.release], the profile that builds the shipped CLI binaries. The three profiles that inherit from release (ci-release, perf, addon) get an explicit codegen-units = 16, so CI test shards, benchmarks and the Node addon build exactly as before. The PR measured a ~16-17% smaller binary and a ~4-5% faster release build.

Risk: low. One file, four added lines, build config only. The only thing that changes is how the shipped binary is compiled (release.yml cargo build --release / cross build --release); the code is the same. One codegen unit with full LTO is a common, well-tested rustc setup.

Look here:

• Cargo.toml:78: the actual change to release
• Cargo.toml:84, :91, :99: pins that keep ci-release / perf / addon at the rustc default of 16
• release.yml:171: where release is used (cross-compiled targets at :175)

Verified: read the full diff. Checked that no other profile inherits release and that no workflow or .cargo config overrides CARGO_PROFILE_*_CODEGEN_UNITS. Checked that only release.yml builds with --release, and its build job has no tight timeout. Checked that CHANGELOG.md is untouched. CI: ci-ok and clippy green on a51923e; the earlier cargo 1.93.1 lock-v4 failure was a static.crates.io download error and passed on re-run. Bugbot: reviewed this head, no findings.

Changes I made: none.

Open questions: none blocking. The PR description's "CI" section still says ci-ok is failing; that is out of date (it went green on re-run). Release-build timing was measured on Linux x86_64 and macOS arm64 only; Windows and cross targets weren't measured, but nothing suggests they would regress meaningfully.

Auto-merge (squash) is armed, so approving sends this straight to the merge queue.

---

Generated by Claude Code