perf(relay): defer post-commit dispatch and avoid verify clone

[tlongwell-block]

Summary

Relay performance stack for the W1/W8 lane:

• defer post-commit event dispatch out of the EVENT OK ack path
• preserve audit-log backpressure by awaiting the bounded audit enqueue before detaching fan-out work
• avoid deep-cloning accepted events for signature verification by sharing through Arc

Stack:

1. 144f00bc — perf(relay): defer post-commit event dispatch
2. 605f7127 — fix(relay): preserve audit backpressure after W1
3. c61b4c14 — perf(relay): share event with sig-verify task via Arc instead of deep clone

Correctness notes

• Event acceptance/audit durability ordering is preserved: bounded audit enqueue remains awaited before detached post-commit work.
• Detached work is limited to Redis publish/local fan-out/workflow/delivery metrics after DB commit and accepted event response path separation.
• W8 changes allocation/ownership shape only; signature verification still runs before event persistence/dispatch.
• Delivery-before-OK was never a NIP/repo correctness guarantee; W1 intentionally prioritizes sender-perceived OK latency by moving post-commit dispatch after the ack path. The observed local receiver delivery cost is named and accepted below.
• Bounded post-commit worker/channel is intentionally left as a Tier-3 follow-up only if fresh-session/prod-like data recreates cross-pod p999 tail regression.

Validation

Code validation already run on this stack:

• cargo fmt --check
• cargo test -p buzz-relay handlers::event::tests
• full cargo test -p buzz-relay
• pre-push hooks passed before remote verification
• remote branch verified at c61b4c149fe30f26e060370fea645412986adf93

Benchmark guardrail / tradeoff:

• W1 raw ack p99 win across A/B runs: -11% to -28%; raw W1 cross-pod p999 tail regression observed and investigated.
• 605f7127 restored audit backpressure; clean cross-pod p999 recovered to 5.60ms vs raw W1 15.64ms and baseline 6.98ms.
• W8-on-fix c61b4c14: all runs accepted, 0 timeouts; ack p99 remains below baseline across all three protocols; W8 p99 deltas vs W1+fix are ≤0.12ms on 200-byte bodies; guardrail PASS.
• Formal §4.13 delivery-ordering addendum: local first-delivery p99 rises +0.3–0.5ms same-pod because fan-out moved from pre-OK inline to post-commit spawned. This is the architectural cost of the ack p99 win (-14 to -23%); the two are the same lever. No repo NIP, TLA+ invariant, or Tamarin lemma binds delivery ordering relative to OK — this is a soft-consistency change, not a fence break. Cross-pod first-delivery p99 is within measurement noise as expected (receivers already routed through Redis). Trade accepted per bench readout.
• Bench artifacts: RESEARCH/RELAY_PERF_BENCH_RUNS/w1fix-605f7127-* and RESEARCH/RELAY_PERF_BENCH_RUNS/w8-c61b4c14-*.