turborepo

e145bc6b - perf: Optimize engine builder, task visitor, and untracked file discovery (#11956)

perf: Optimize engine builder, task visitor, and untracked file discovery (#11956) ## Summary Three targeted optimizations to the `turbo run` hot path, focused on areas identified via `--profile` on repos of varying sizes. ## Benchmarks Hyperfine (30 runs, 10 warmup) comparing mainline to this branch across three internal monorepos: ### Large repo (~1000 packages) | | Mean | Range | |---|---|---| | **mainline** | 1.265s ± 0.176s | 1.091s – 1.756s | | **this PR** | 1.514s ± 0.115s | 1.370s – 1.801s | | | **1.20–1.26× faster** | | `build_engine` self-time: 283ms → 74ms (−74%). `queue_task` self-time: 330ms → 181ms (−45%). ### Medium repo (~120 packages) | | Mean | Range | |---|---|---| | **mainline** | 807ms ± 82ms | 744ms – 1149ms | | **this PR** | 823ms ± 71ms | 765ms – 1055ms | | | **~1.02× (within noise)** | | ### Small repo (~5 packages) | | Mean | Range | |---|---|---| | **mainline** | 581ms ± 40ms | 523ms – 714ms | | **this PR** | 579ms ± 50ms | 512ms – 708ms | | | **~1.00× (within noise)** | | The optimizations scale with `packages × tasks`, so the large repo sees the most benefit. Medium and small repos are dominated by filesystem I/O (`find_untracked_files`) and lockfile parsing, which are unaffected. ## Changes - **Engine builder**: Cache the turbo.json extends chain per package name and move the `visited` set check before `task_definition()`. The chain resolution only depends on the package, not the task, so all tasks in the same package share the cached chain. The early `visited` check avoids recomputing task definitions for duplicate BFS entries. - **Task visitor**: Defer the `env()` call to the non-dry-run branch. The execution environment map is unused during dry runs, so this skips per-task RwLock acquisition, `DetailedMap` cloning, and wildcard regex matching in that path. - **`find_untracked_files`**: Replace `Mutex<Vec>` with per-thread local buffers that flush via `mpsc` channel on drop. This eliminates per-file mutex contention in the parallel directory walker. ## Testing All existing tests pass across the three modified crates: - `turborepo-engine`: 62 tests (covers extends chains, cycles, diamond inheritance, `extends: false`, BFS graph construction) - `turborepo-scm`: 110 tests (covers untracked file detection, git index equivalence, package boundary isolation)