[pablodeymo]

🗒️ Description / Motivation

spawned v0.6.0 was released on 2026-11-06. It adds actor links, monitors, exit reasons and an opt-in metrics feature, and fixes several shutdown races (idle stop/dispatch ordering, reserved publications at shutdown, monitor waiter retention). This PR moves ethlambda's own actors onto it and turns the metrics feature on, so /metrics reports how the BlockChain and P2P mailboxes behave.

What Changed

Bump (Cargo.toml): spawned-concurrency and spawned-rt 0.5.0 → 0.6.0. Handler and protocol signatures are unchanged, so no actor code moves.

Lockfile: the three spawned crates; windows-* 0.57 → 0.100 (Windows-only targets, inactive here); loom, generator, scoped-tls enter as lock-only entries behind spawned-rt's cfg(spawned_loom) target dependency and never build (cargo tree -i loom finds nothing).

ethrex-p2p v25.0.0 keeps its own spawned 0.5.0. No spawned type crosses that boundary, so both versions coexist.

Metrics (crates/net/rpc): spawned-concurrency gets features = ["metrics"] on the workspace dependency. spawned's collector is separate from the prometheus registry and ships no exporter, so get_metrics appends spawned_concurrency::metrics::render_prometheus() after the registry text. New series on /metrics: spawned_mailbox_depth, spawned_queue_wait_seconds, spawned_handler_duration_seconds, spawned_request_latency_seconds, plus message/handler/request/actor lifecycle counters. Process-wide aggregates, no per-actor labels. docs/metrics.md has the table and how to read them. The rpc crate had spawned as a dev-dependency only; it is now a dependency.

Correctness / Behavior Guarantees

• Per-message cost without the feature is unchanged: the enqueue path and the receive loop in tasks/actor.rs differ from 0.5.0 only by the metrics hooks, plus one is_cancelled load per receive. send_after is identical. With the feature on, each message carries an inline observation and one Instant::now() per enqueue and per handler start/end (every observation is timed; timing_sample_every stays at its default of 1, since a deterministic 1-in-N sample would alias with the per-slot periodicity of lean traffic).
• Shutdown (wait_for_shutdown: context().stop() then join()) keeps its semantics; join() is now a wrapper over wait_exit().
• spawned's own audit of ethlambda at 78359e2c (research/backend-experiments/2026-11-06-runtime-isolation/CLIENT-WORKLOADS.md) notes the chain actor already runs on Backend::Thread, so none of 0.6.0's experimental async schedulers reach it; the metrics are the measurement that audit recommends taking first.

Tests Added / Run

• crates/net/rpc/tests/http_servers.rs: /metrics body must contain spawned_mailbox_depth.

• make fmt, make lint: clean on both commits

• cargo test -p ethlambda-rpc: 361 passed, 0 failed (metrics commit)

• The same two commits rebased onto nightly (branch bump-spawned-0.6.0-nightly): clean rebase, cargo check --locked consistent, lint clean, rpc tests 361 passed

• make test-node: 1544 passed, 0 failed (bump commit)

• make test-consensus: 1324 passed, 0 failed (bump commit)

• 3-node local devnet, 0.5.0 image vs 0.6.0 image (same build flags, same roster/keys/ports, 65 slots each, run back to back with nothing else building). Same input is not possible on a live devnet, so this is a regression gate, not a perf verdict:

  Metric (3 nodes)	0.5.0	0.6.0
  head / justified / finalized slot, per node	65 / 63 / 62	65 / 63 / 62
  finalizations, blocks received	248, 130	248, 130
  tick interval mean	800.003 ms	799.996 ms
  attestation gossip arrival delay mean	37.4 ms	37.7 ms
  aggregation gossip arrival delay mean	8.1 ms	7.6 ms
  block gossip arrival delay mean	52.7 ms	27.7 ms
  fork-choice block processing mean	17.4 ms	17.7 ms
  block building mean	676 ms	584 ms
  process CPU seconds, summed	492	453
  RSS per node (process metric)	566 MiB	567 MiB
  ERROR log lines	0	0

  Consensus progress is identical and every actor-sensitive timing (tick interval, attestation/aggregation arrival, fork choice) is within noise. The two larger deltas (block arrival, block building) are live-network and XMSS-proving quantities with n=130 and n=65; not attributed to spawned.

• 3-node local devnet on this branch's image (nightly + these two commits), 7 min / 97 slots, finalized 94, 0 ERROR lines, Prometheus + Grafana up via lean-quickstart --metrics. Per node, 2-minute rate windows at the end of the run:

  spawned_* series	ethlambda_0 (aggregator)	ethlambda_1	ethlambda_2
  mailbox depth (instant)	0	0	2
  queue wait p50 / p99	87 µs / 53 ms	58 µs / 30 ms	61 µs / 21 ms
  handler duration p99	785 ms	831 ms	834 ms
  messages handled / s	4.3	3.6	3.5
  enqueued = handled over the run	1722 = 1722
  rejected / discarded / panicked	0	0	0
  actors active	2	2	2

  Reading: the ~800 ms handler p99 is the block-building tick (XMSS proving); the tens-of-ms queue-wait p99 is the messages that landed behind it, and the mailbox drains to zero in between. spawned_request_latency_seconds stays empty because ethlambda's actors only cast, never call.

Note: make cooldown-check will fail until 2026-11-20 since 0.6.0 was published today.