[MegaRedHand]

Draft: pins ethrex to 8f847524, the head of ethrex's feat/discv5-predicate-queries. That branch has no ethrex PR yet and is built on the open lambdaclass/ethrex#7217. Not mergeable until both are in an ethrex release and the pin goes back to a tag.

Includes #605. This branch merges feat/ethrex-discovery-contact-table (#605), since predicate queries are built on the contact-table API that PR migrates to. #605 can be closed in favor of this one, or merged first; the diff here then shrinks to the last three commits.

Motivation

The dial loop ranks what discv5's crawl happens to find, which only reorders random draws. About one mainnet record in 32 advertises a given attestation subnet, so a subnet this node subscribes to can stay without peers for as long as the crawl takes to stumble on one. ethrex's new DiscoveryHandle::find_nodes and records_matching let the search be aimed instead, the way Lighthouse uses sigp/discv5's find_node_predicate.

What this does

1. ethrex pin bump

v25.0.0-based #7217 head (4462253e) to 8f847524 (v30.0.1 + the branch). Builds with no API changes: the contact-table API is the same. The lockfile moves only ethrex crates; make cooldown-check passes.

2. Subnet hunt (discovery/subnets.rs)

 dial tick ── wanted subnets ──► watch ──► hunt task (one hunt at a time)
     ▲                                       │ 1. records_matching: what discovery holds
     │                                       │ 2. find_nodes: walk the DHT for the rest
     └──── admitted peers ◄──── mpsc ◄───────┘

Piece	Behavior
What is short	A subscribed subnet (what our attnets advertises) with fewer than TARGET_SUBNET_PEERS = 3 connected peers advertising it (Lighthouse's number)
Predicate	Admission (LeanFilter) plus at least one wanted attnets bit; the bit test runs first so most records are turned away before anything is decoded
Cadence	One hunt at a time on its own task: cached matches first, then a walk of up to 10 s, at most 6 peers per hunt, 30 s between hunts
Gating	Nothing is wanted while the dial loop has no budget, so no walk runs for peers nobody would dial
Re-offers	A peer a hunt handed over is left out of hunts for 10 minutes: a mainnet peer at its inbound cap connects and says goodbye, and discovery keeps offering it
Where finds go	The existing candidate queue. rank_candidates gains a tier: custody gaps, then wanted subnets, then any uncovered subnet
Metric	lean_discovery_subnet_hunt_peers_total

3. ENR sync between discv5 and the API

Direction	Status
discv5 → API	Done. ethrex writes every record it re-signs (discv5 IP voting) into the SharedLocalNode it is spawned with. discovery::LocalEnr reads that mirror, and /eth/v1/node/identity reads it per request through ethlambda_rpc::EnrSource (a closure, since the RPC crate does not depend on ethlambda-p2p).
ethlambda → discv5	Not supported by ethrex yet. DiscoveryServerProtocol has no message that edits the local record. Its only consumer-driven edit, the fork_id watch, writes the EIP-2124 eth entry, not eth2. Writing into the mirror does nothing: discovery serves its own copy and overwrites the mirror on its next edit. So the served eth2 still keeps the startup digest across a fork boundary (update_served_enr still only warns; its doc now says this). What ethrex would need: a protocol message that applies an edit through NodeRecord::edit (one seq bump, re-signed) and mirrors the result, as refresh_fork_id already does for eth.

Merge resolution with #605

Main had made LeanFilter Clone with a shared fork-boundary Policy; #605 had made it non-Clone because discovery takes it by value. Both hold now: clones share the Policy and the DialTargets map, and SpawnedDiscovery carries a clone, so set_fork_id still reaches discovery's copy. #605's behavior change is inherited: --discovery.advertise-ip sets nat_extip_set, so IP voting no longer overwrites an address the operator named.

Assumptions

• "Subnets we're missing" means subscribed subnets below 3 connected peers, not only subnets with zero.
• Duty subnets from a validator client (beacon_committee_subscriptions) are not fed to the hunt; that stays a separate gap.
• Counts come from the attnets of peers dialed from discovery. Inbound peers never show us an ENR, so they count for nothing: the hunt errs toward hunting.

Testing

• make lint, cargo check --workspace --all-targets --locked, make cooldown-check
• cargo test --workspace --profile release-fast --lib --bins: all pass. leanSpec spec suites not run.
• New unit tests: the hunt predicate (wanted bit, other network, already offered), offer's limit and dedup, short_subnets, the no-budget gate, and the new ranking tier.
• Not covered: an end-to-end hunt across live discovery servers. ethrex's own discv5_server_tests cover find_nodes; this side tests what it feeds it and what it does with the answer.

Follow-ups

• Return the pin to an ethrex release tag once #7217 and the predicate-query branch land.
• Feed VC duty subnets to the hunt (with expiry), as Lighthouse and Grandine do.
• Ask ethrex for a local-record edit message, then refresh eth2 at fork boundaries.
• records_matching does not exclude ethrex's tried set, so a hunt can re-offer a peer the contact poll already handed out within the same 10 minutes; the dial loop absorbs it (AlreadyInProgress, skip if connected).