Add noopStreamGen and noopConsume moroutines + matching benches that pit 1-direction stream (worker→parent) against 1-direction channel (parent→worker) on equal footing — both are one cross-thread hop per item. Findings (on this branch, HEAD): stream (worker generates → parent consumes) ~746K items/s channel (parent generates → worker consumes) ~652K items/s stream round-trip (parent → worker → parent) ~464K items/s vs unoptimized (bb41d5d): stream 1-dir: ~66K (+11.3x) channel 1-dir: ~67K (+9.7x) Two takeaways: - The earlier "channel beats stream" observation was an artifact of comparing channel's 1-hop case to stream's 2-hop (pass-through) noopStream-with-input case. When measured 1-direction-vs-1-direction, stream and channel are tied pre-optimization. - Post-optimization, stream is ~15% faster than channel — that's the win from atomics-based backpressure over message+adaptive-yield. Applying atomics to the Distributor would likely close that gap. Also keeps channelFanout at varying worker counts as a separate section — shows the Distributor's single-threaded producer bottleneck (peak at ~2 consumers, regresses past 4). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>