Summary

Reduces turbo run --dry wall-clock time by 5-12% on large monorepos (~1700 tasks) by cutting unnecessary filesystem operations and CPU work.

User time dropped 39% (3.7s → 2.25s) and system time dropped 30% (1.87s → 1.32s) on the largest benchmark repo. Wall-clock improvement is bounded by irreducible I/O (git subprocesses, remaining directory walks).

Changes

Commit 1: Fast-path invariant globs

Literal file globs like package.json and turbo.json — appended to every task's include list — now resolve with a single symlink_metadata() instead of a full directory traversal via wax::Glob::walk(). Also separates glob compilation from the walk so EMFILE retries don't recompile patterns, and removes per-call tracing from visit_file/glob_with_contextual_error (~6000 calls in large repos).

Commit 2: In-memory exclude matching + parallel hashing

Exclude globs are now matched against already-known paths in memory instead of spawning a separate filesystem globwalk. hash_objects uses rayon for parallel file hashing. git_ls_tree_repo_root_sorted builds a BTreeMap directly, skipping the HashMap→BTreeMap conversion.

Commit 3: Arc-share FileHashes + batch mutex

expanded_hashes stores Arc<FileHashes> so distributing results to ~1700 tasks is a refcount bump, not a HashMap clone. calculate_dependency_hashes acquires the tracker mutex once instead of per-dependency.

Benchmarks

hyperfine with --warmup 5, 10 runs each. faster = this PR, baseline = current main.

Large repo (~1700 tasks):

faster:   2.004 s ± 0.057 s
baseline: 2.165 s ± 0.033 s
→ 1.08 ± 0.03x faster (User: 2.26s vs 3.70s, System: 1.32s vs 1.87s)

faster:   1.996 s ± 0.054 s
baseline: 2.241 s ± 0.192 s
→ 1.12 ± 0.10x faster

Medium repo (~200 tasks):

faster:   1.135 s ± 0.074 s
baseline: 1.158 s ± 0.067 s
→ 1.02 ± 0.09x faster

Small repo (~5 tasks):

faster:   640.6 ms ± 27.0 ms
baseline: 655.9 ms ± 40.3 ms
→ 1.02 ± 0.08x faster