node, chain: parallel aggregate submission for #907

[ch4r10t33r]

Summary

• Merge #916 parallel aggregate work: replace serial std.Io.Threaded (concurrent_limit=1) with bounded non-blocking submits onto the shared ThreadPool (aggregate_max_inflight=4).
• Check the in-flight cap before the expensive state cloneAndRelease, so saturated workers do not waste deep clones.
• Add zeam_aggregate_skip_total{reason="spawn_failed"} when ThreadPool enqueue fails (never fall back to inline ~10s FFI on the libxev thread).
• Clarify Prometheus HELP text: lean_pq_sig_aggregated_signatures_building_time_seconds measures per-att_data wrap time inside computeAggregatedSignatures, not the XMSS STARK worker; operators should use zeam_aggregate_worker_duration_seconds for end-to-end FFI cost (#907).

Context

On the ansible devnet, zeam aggregators were dropping ~50% of slot triggers as in_flight because the XMSS FFI takes ~11s while the slot interval is ~6s and only one aggregate worker could run at a time (#907). Live logs on zeam_8 with a #916-era image showed the skip rate improved to ~10%, but publish lag and timely=none merge timing remain separate follow-ups.

This branch is based on current main (includes #908 trivial-input pre-filter) and supersedes the open #916 branch for review/merge.

Test plan

• zig build test --summary all
• Deploy to devnet aggregator (zeam_8) and confirm zeam_aggregate_skip_total{reason="in_flight"} rate drops vs pre-parallelize aggregation as an aggregator #916 baseline
• Confirm zeam_aggregate_worker_duration_seconds p50 stays ~10s for multi-sig inputs (parallelism helps throughput, not per-run FFI cost)
• Confirm zeam_aggregator_publish_aggregations_total continues incrementing on duty subnet under load

[zclawz]

Reviewed PR #919.

No blocking findings from my pass. The change correctly moves aggregate submission off the single dedicated Io.Threaded worker onto the shared bounded thread pool, does the in-flight cap check before the expensive state clone, releases the in-flight counter on all early/error paths I checked, joins outstanding aggregate workers before chain teardown, and updates the Prometheus help text/skip reasons consistently.

Validation:

• GitHub checks on 6fa31fe5dc220e84f692a1e86999e14e47cf6600 are green (test, build, lint, docker/prover jobs).
• I attempted local zig build test --summary all, but this workspace’s Zig/dependency combo fails before repo tests with dependency build-script API errors (Build.Graph.io / b.modules.put), so I’m treating CI as the authoritative validation here.

LGTM from me.

[zclawz]

Approved. I re-checked the current head after the force-push; the tree matches the version I reviewed, with no new diff from my prior pass.