[JacobStephens2]

Summary

The macOS shards' remaining time is git processes. Most come from Worktree ownership inspection and the launch-side recheck (#858). Homebrew's git also costs about 13 ms per process on the runner, where Xcode's costs 9 ms. This PR leaves the checks alone and makes each git process cheaper in macOS CI:

 test-macos-gate / test-macos-shard (macos-26)
   checkout, rustup, rust-cache
+  sudo mdutil -a -i off                    # Spotlight stops indexing the tests' repos
+  link "$(dirname "$(xcrun -f git)")"/git* into $RUNNER_TEMP/xcode-git
+  put $RUNNER_TEMP/xcode-git first on PATH # Xcode's git 2.50.1, not Homebrew's 2.56.0
   cargo nextest run …

• docs/ci-performance.md adds "After Inspect Worktree ownership with fewer git processes to shorten macOS CI #857: the macOS breakdown": five diagnostic passes, with method, runs, and a per-test breakdown by cause for the 26 slowest tests.
• tests/security_run.rs: two tests run thirdshift with only bin on PATH. A shared link_git now links git-upload-pack/git-receive-pack beside git when present. Xcode's git looks for them only beside itself. Their assertions are unchanged.
• Decide whether Worktree ownership inspection may recheck the launch repository less often #858 got the macOS numbers: Decide whether Worktree ownership inspection may recheck the launch repository less often #858 (comment)

Where the time goes (26 tests alone on macos-26, dtrace; caller split from Linux strace of the same tests):

wall time of a slow test
└── git processes alive: 77–99% (88% summed)
    ├── 41%  rev-parse --git-common-dir    launch recheck, #858, untouched
    ├── 43%  ownership inspection          untouched
    └── 17%  everything else
fakes 1–3% · shells <2% · fsync <0.1 s · sleeps overlap git work

Evidence

macOS test-step seconds (the cargo nextest run step, build included, as in #894's table). Run-to-run noise on main is ±15% (#857):

Run	Shards 1 / 2 / 3 / 4	Sum
main after #857, 39424072554	959 / 1189 / 1063 / 1037	4,252
main after #857, 39424576921	1076 / 1019 / 986 / 935	4,020
main after #857, 39430050727	1099 / 769 / 1066 / 982	3,920
This PR, 39559241797	622 / 619 / 607 / 575	2,423
This PR, final run, 39559241308 (re-run)	406 / 435 / 522 / 379	1,742
First main run after merge	not yet run; add when it lands

Both of this PR's runs, on the same commit, are well outside the noise: the first is 38–43% below the three post-#857 main runs and the final one 56–59% below. Build excluded, their nextest time was 2,198 s and 1,590 s. The gap between the two is runner variance, larger than #857 measured.

Opening the PR with test:macos already applied started an opened and a labeled run. The first was cancelled, leaving cancelled checks on the head, which docs/ci-performance.md warns about. Re-running it replaced them and gave the second measurement. In that re-run, Linux test failed once in tests/console_install.rs:269, unit_preserves_literal_paths_with_shell_and_systemd_metacharacters: Text file busy (ETXTBSY) executing a freshly copied binary. That file is untouched here, and test passed in the first run and on re-running the job. It is a pre-existing race, left for the Day shift.

Diagnostic passes (nextest time, build excluded, from docs/ci-performance.md): five main runs summed 3,327–3,595 s. With Spotlight off alone they summed 3,209 s, inside the noise. With Xcode's git alone, 2,132 s.

Before: in pass 4, with Xcode's git, security_run an_unavailable_github_cli_has_a_safe_pre_audit_dependency_cause and missing_node_fails_before_any_session_and_names_the_prerequisite failed: git-upload-pack: command not found.
After: both pass on Linux. On macOS they pass in this PR's full-suite run (39559241797: all four shards passed).

Local x86_64 Linux: cargo nextest run 2,168 passed; cargo test 2,171 passed; cargo clippy --all-targets -- -D warnings and cargo fmt --check clean.

Merge Danger

Door: two-way. Revert the two CI steps to get Homebrew's git and Spotlight back.

Blast Radius: CI-only.

macOS CI now tests against Xcode's git 2.50.1 instead of Homebrew's 2.56.0. A bug that shows only with Homebrew's newer git would not show on macOS. Linux CI keeps Ubuntu's git. No src/ change, and Worktree ownership checks are unchanged.

Unaddressed findings

Standards

• Duplicated Code: the two new steps repeat in test-macos-gate and test-macos-shard. Declined. ci.yml already repeats each job's toolchain and cache steps per job, for example "Install the Rust version…" in test, test-macos-gate and test-macos-shard. The gate's comment points to the shards' explanation.
• which git lookup duplicated with tests/pickup_run.rs::install_git_wrapper. Left as is. That duplication predates this PR, and the reviewer marked it optional.

Spec

• Scope creep: the gate gets the same two steps. Declined. The gate runs the macOS-specific tests as the full suite runs them. Giving it a different git from the main and nightly shards would test those tests on a git the full suite no longer uses. macos-plan is untouched, so no event's macOS work changes (Run the full macOS suite on main and nightly, with a one-job macOS gate on PRs into main #893 stays out of scope).
• Homebrew's git no longer runs anywhere in CI. No change. This is the trade-off under Merge Danger, recorded at the end of the doc section.

Review coverage: both reviewers read all three changed files. No generated output was excluded.

The diagnostic workflow lives on the issue-894-profile branch, which the doc's run links point to. Delete it whenever its logs are no longer needed.

Closes #894

🤖 Generated with Claude Code

Built with claude · claude-opus-5-5 · medium

Security review outcome

No unaddressed introduced findings.