No pool needed — goroutines already fit the shape.

Goroutines are cheap enough (a few KB of stack, M:N-scheduled onto OS threads by the Go runtime) that pooling them would fight the language rather than match it — this port's biggest structural departure from the shared design's “one shared worker pool.” A bus-wide buffered channel used as a counting semaphore still bounds how many drain goroutines run concurrently, playing the role a sized pool plays elsewhere, layered on top of each stage's own per-channel semaphore — Go's standard idiom for a semaphore, since the language ships no counting-semaphore type in its standard library. Go has no default/named parameters, so the envelope's optional slip/sender/headers become the functional-options pattern (MakeEnvelope(to, payload, WithSlip(...), WithSender(...))) instead.

sync.Mutex + sync.Cond back the per-stage queue, itself a container/list.List rather than a slice — repeatedly re-slicing a FIFO from the front never shrinks the backing array, a real memory-growth gotcha for a long-running bus, while a linked list gives true O(1) push/pop at both ends. sync.Cond has no built-in timeout, so a time.AfterFunc timer broadcasts on expiry and the waiter re-checks its own deadline on wake — the same pattern other ports implement with wait_until/Monitor.Wait(timeout). And since nothing here is a pool Shutdown can join, an atomic in-flight counter (the same technique C# needs for the same reason) lets awaitDrain wait for zero in-flight and zero depth, not depth alone.

The one real bug found and fixed: a lost-wakeup race in schedule(). It silently gave up when a bus-wide permit acquisition failed, with no guarantee anything would ever retry the channel — if a channel's own producer had already finished and its last drain goroutine lost that race, its remaining backlog sat queued until Shutdown's timeout expired. This showed up intermittently in this very benchmark's chan config across two separate sessions. The fix: releaseBusPermit now sweeps every registered channel with a non-empty queue and re-offers it to schedule() whenever a permit frees, not just the channel that happened to release it. Verified: go test -race clean, 10 consecutive full benchmark runs (30 chan trials) with zero drain failures, versus 2-of-3 failing before the fix under the same conditions.