CodeGraph CLI MCP Server

## Development Context - Feature: Lock-free concurrent data structures for CodeGraph - Technical Stack: Rust 1.75+, atomics (Acquire/Release), crossbeam (queue, utils, skiplist), arc-swap (RCU), tokio (tests), loom (memory ordering tests) - Constraints: Zero global locks on hot paths; bounded memory; no blocking ops in async; integrate cleanly with existing workspace; avoid breaking existing crates - Success Criteria: 10,000+ QPS throughput for queueing pipelines; lock-free graph adjacency operations; memory ordering validated with targeted loom tests; unit tests and examples included ## Library Insights (Phase 1) - crossbeam-queue::ArrayQueue - Bounded MPMC, lock-free; `push`, `pop` return errors on full/empty - Very fast for many-producer/consumer scenarios; backoff recommended under contention - crossbeam-skiplist::SkipMap - Lock-free concurrent map with ordered keys; fast reads/writes without global locks - Suited for per-node adjacency indices and node storage - arc-swap::ArcSwap - Atomic `Arc<T>` with lock-free reads; `rcu` for copy-on-write updates - Ensures correct Acquire/Release semantics for pointer publication - loom - Deterministic exploration of interleavings for atomic algorithms - Use feature-gated tests for SPSC memory ordering sanity checks ## Best Practices (Phase 2) - Prefer wait-free SPSC for hottest paths (producer-consumer pipelines). Use Acquire/Release for head/tail publishing. - For MPMC, use proven lock-free bounded queues (ArrayQueue) and prefer backoff/yield on contention to reduce cache-line ping-pong. - For lock-free graph ops, keep reads trivially lock-free (ArcSwap + clone-on-write); writers retry via RCU CAS. - Validate invariants with unit tests; use loom selectively to keep test times in check.