ADR-0009: Own work-stealing scheduler instead of WalkParallel or rayon

• Status: Accepted
• Date: 2026-08-18

Context

The Palamedes benchmark puts parallel ignore + pruning at 15.74x and zlob at 16.84x over the serial baseline — the scheduler is where the remaining gap and the API freedom live. ignore::WalkParallel's visitor model cannot cleanly carry our streaming, sharded-collect, cancellation, and error-policy semantics; wrapping it would let its design leak into our public API and cap performance at ignore's ceiling. rayon's global pool conflicts with per-walk thread limits and cancellation.

Decision

Build our own scheduler (Phase 3): work stealing via crossbeam-deque, lazy worker spawn (start on the caller thread, add helpers only when parallel work exists), per-worker path/read/matcher scratch and result shards, a single cancellation state, and a lossless error channel. The ignore crate contributes only its gitignore rule matcher (ADR-0006).

Consequences

• This is the project's performance differentiator — and its largest concurrency-testing obligation (loom models, stress tests, panic propagation).
• Phase 2 ships a single-threaded portable walker first; parallelism is layered on, and single- vs multi-thread result equality is a corpus invariant.