Low-alloc dispatch: dual behavior model, opt-in arena scheduler, allocation-free park primitive

[thnak]

Drives the activation-scoped grain-call dispatch path toward true-zero steady-state allocation, ranked by measured bytes. Three orthogonal features that ship together; the branch tip builds clean (0 warnings) and is green across scheduling 45/45, Fault 9/9, Integration 38/38, unit 549–550/550 (the 1–2 residual failures are documented wall-clock timing flakes that shift across runs and pass in isolation).

Commits

1. IActivationBehavior dual behavior model (Part A)

An opt-in per-activation behavior lifetime alongside the existing per-call IGrainBehavior, killing the ~792 B/call DI-scope + resolution bucket for grains that opt in.

• One activation-lifetime IServiceScope + one cached behavior instance in GrainActivation (constructed once, reused per call, disposed on deactivation).
• LocalGrainCallInvoker fast path: no scope, no construct, no locked ResolveService. Reentrant + IActivationBehavior throws NotSupportedException.
• BehaviorStateAnalyzer: suppress QRK0020/QRK0021 for IActivationBehavior (per-activation fields are legitimate state); QRK0022 still fires; new QRK0023 (Warning) flags [Reentrant] + IActivationBehavior.
• PostCoreAsync variants use PoolingAsyncValueTaskMethodBuilder so dispatch async boxes are pooled.
• Measured (DispatchPipelineBenchmarks): 2039 → 1253 B/call for opted-in grains (−786 B); mean latency 14.2 → 6.8 µs.

2. Opt-in ArenaScheduler (next-gen scheduler, ArenaV2)

A fresh IActivationScheduler on dedicated worker threads, per-worker work-stealing deques, and a sharded injection queue — the P1 skeleton of the next-gen scheduler design. Selected only via SiloRuntimeOptions.SchedulerKind == ArenaV2; the legacy scheduler stays the default until the arena scheduler clears the full benchmark suite.

• Spill-to-ThreadPool on await removes both blocking-drain latency and the bounded-worker reentrancy deadlock.
• Per-worker in-flight backpressure (Dekker-fenced); same lost-wakeup-free park ordering as the legacy scheduler.
• Coverage: async-resume, concurrency stress, in-flight bound, stealing, message-priority-lane tests. Perf runners gain --v2.

3. Low-alloc dispatch — allocation-free park primitive + shutdown fix

• WorkerParkSignal replaces the per-worker SemaphoreSlim(0, int.MaxValue), whose cancelable WaitAsync(ct) slow path allocated ~200 B/park in the park regime. A single-consumer/multi-producer async auto-reset event backed by a reusable ManualResetValueTaskSourceCore<bool> — a park now allocates nothing. Token-free; DisposeAsync Set()s every worker on shutdown. A boolean auto-reset event is sound because the wake is edge-triggered (one wake re-sweeps all shards). Verified against 20M multi-producer/consumer cycles of the exact wake protocol, zero lost wakes.
• Folded-in earlier reductions from the same arc: spin-before-park, lock-free long[] bitmask idle registry (replacing ConcurrentStack<int>), StatelessWorkerRouter delegate caching, state-passing GrainActivationTable.GetOrCreateAsync<TState> overload.
• Shutdown fix (pre-existing latent bug the primitive uncovered): since the Channel → ConcurrentQueue ready-queue rewrite, ScheduleAsync stopped rejecting work after shutdown, so an activation deactivating after the scheduler is disposed (the DI dispose order) posted its OnDeactivate turn onto dead workers and hung forever. The old Channel threw ChannelClosedException here, driving GrainActivation's inline-drain fallback; the SemaphoreSlim primitive was masking the gap by accidentally throwing ObjectDisposedException from a disposed semaphore. ScheduleAsync now throws ObjectDisposedException when _cts.IsCancellationRequested (one hot-path volatile read), restoring the reject-on-shutdown contract explicitly. Locked in by ActivationSchedulerShutdownRejectsScheduleTests (both tests fail — one by hanging — with the guard removed).
• Measured (AllocByTypeProfiler, park regime): ~200 → 1.3–1.9 B/call, residual is just the pooled MailboxWorkItem<T>.

Notes for reviewers

• Commits are split by feature. Two shared runtime files (GrainActivation.cs, LocalGrainCallInvoker.cs) carry both Part A and Part B edits and land in commit 1 as the runtime enablement for IActivationBehavior; their pooling-builder micro-opts are described in commit 3's body and the design spec.
• Design specs under docs/superpowers/specs/ (2026-07-12-next-gen-scheduler-design.md, 2026-07-13-low-alloc-dispatch-design.md, 2026-07-13-scheduler-v2-pingpong-cost-profile.md).
• Follow-up tracked in the spec: ArenaScheduler.ScheduleAsync has the same latent reject-on-shutdown gap (opt-in, currently accidentally-masked); needs ValueTask.FromException since its ScheduleAsync is synchronous.

🤖 Generated with Claude Code