Summary

• Parallelize git ls-tree and git status inside RepoGitIndex::new using std::thread::scope
• Pre-build the RepoGitIndex eagerly on a background thread during RunBuilder::build(), overlapping the git I/O with package graph construction
• Share the pre-built index across calculate_file_hashes and get_internal_deps_hash instead of each building their own

These three git subprocess calls (git ls-tree -r -z HEAD and git status -z) previously ran sequentially and only started after the package graph was fully constructed. They have no data dependency on the package graph — they only need the git binary path and repo root. By starting them earlier and running them in parallel, we hide the entire cost behind work that was already happening.

Benchmarks

Tested with --dry runs across three monorepos of varying sizes.

Large monorepo (~hundreds of packages)

Benchmark 1 (this branch):  6.893 s ± 0.075 s
Benchmark 2 (main):         7.056 s ± 0.059 s

Summary: 1.02 ± 0.01 times faster

Medium monorepo (~100 packages)

Benchmark 1 (this branch):  1.476 s ± 0.108 s
Benchmark 2 (main):         1.784 s ± 0.217 s

Summary: 1.21 ± 0.17 times faster

Small monorepo (~6 packages)

Benchmark 1 (this branch):  672.3 ms ± 25.7 ms
Benchmark 2 (main):         677.1 ms ± 55.4 ms

Summary: 1.01 ± 0.09 times faster (within noise)

The biggest win is in medium-sized repos where the git index build is a larger fraction of total runtime. Small repos fall below the index threshold and are unaffected. Large repos benefit in absolute terms (~163ms) but the percentage improvement is smaller since other work dominates.