perf: Resolve literal input paths via stat instead of glob walk (#11955)
## Summary
When tasks use `$TURBO_DEFAULT$` with additional literal inputs (like
`$TURBO_ROOT$/tsconfig.json`),
`get_package_file_hashes_from_inputs_and_index` was running every
include pattern through `compile_globs` (wax regex compilation) +
`walk_glob` (full directory traversal) — even for literal file paths
that resolve to exactly one file.
This separates literal paths from actual glob patterns. Literals get a
single `symlink_metadata` syscall. Only patterns with metacharacters
(`*`, `?`, `[`, `{`) go through the expensive compile + walk path.
## Impact
On a 630-package monorepo where the `typecheck` task includes
`$TURBO_ROOT$/tsconfig.json`, this eliminates ~990 out of 995 glob
compilations and directory walks.
**Profile data** (`turbo run typecheck --skip-infer --dry --profile`):
| Function | Before | After |
|----------|--------|-------|
| `walk_glob` | 51.5% (721ms) | 5.2% (68ms) |
| `compile_globs` | 28.5% (399ms) | <1% |
| Total spans | 18,148 | 15,164 |
**Hyperfine** (30 runs, `--warmup 10`, `typecheck --skip-infer --dry`):
| Repo | Optimized | Mainline | Ratio |
|------|-----------|----------|-------|
| Large (630 pkgs) | 1.623s ± 0.159s | 1.654s ± 0.168s | 1.02x |
| Medium (120 pkgs) | 821ms ± 66ms | 878ms ± 95ms | 1.07x |
| Small (5 pkgs) | 584ms ± 44ms | 589ms ± 108ms | 1.01x |
**Caveat**: Wall-clock improvement is within system noise on most runs.
The CPU time reduction is real and consistent (User time on the large
repo drops from ~1.96s to ~1.48s, a 25% reduction), but I/O and
system-time variance prevent it from reliably surfacing in hyperfine for
my machine.
turborepo
b0048b03 - perf: Resolve literal input paths via stat instead of glob walk (#11955)
