discv4/discv5 remainders on develop: NEIGHBORS, NODES cap, node-map flood eviction

[pk910]

Overview

Targets develop. Adds the three discv4/discv5 fixes that develop was still missing after its independent rework — the remainders from the #44 audit set that develop did not already cover (BUG1 ENRRESPONSE, G9 ingest size, BUG5 bond, --serve-all are all already in develop).

Each fix is its own commit. All CI checks pass locally (go mod verify, go vet, gofmt, staticcheck, go test -race, make build), and each fix is verified against a full glamsterdam-devnet-7 devnet (all EL+CL clients, bootnodoor as sole bootnode) with no regression.

Commits

1. discv4: always send NEIGHBORS and advertise the real TCP port. An empty routing table produced a silent (zero-packet) response, so go-ethereum's querier waited out its full request timeout instead of returning early on the first reply. Always send ≥1 NEIGHBORS packet. Also advertise the node's real TCP port from its ENR instead of copying the UDP port.
2. discv5: cap NODES response to ≤15 nodes / ≤5 packets. go-ethereum honours only the first NODES packet's total and reads at most 5 packets; a dual-serve response could emit 6, silently dropping nodes. sigp/discv5 caps at 16.
3. discv4: evict a stale node instead of dropping new peers when the map is full. lookupOrCreateNode returned a non-retained node once at MaxNodes, so under a flood of distinct signed IDs a genuinely new peer's inbound PING marked bond state on a discarded object and could never bond (memory-growth DoS → bonding-lockout DoS). When full, evict one unbonded entry to admit the new node; bonded, endpoint-proven peers are never evicted this way.

Verification (glamsterdam-devnet-7, all clients, bootnodoor as sole bootnode)

No regression vs develop: all 12 nodes discovered, 11 discv5 sessions, 0 EL fork rejections, 0 bootnodoor errors/warnings, storm-free (≤17 packets to any single peer over the run), CL 5/5 mesh, EL 5/5 (besu/erigon 4/5, unchanged — that gap is the besu-v4-only vs erigon-v5-only artifact, not bootnodoor), 0 client-side bans.

Each fix addresses a real concern:

• BUG3 — shown live. geth harness against an empty-table bootnode: develop delivered 0 NEIGHBORS packets (geth waits out its timeout on every FINDNODE); this branch delivered 10 (immediate empty replies). A fresh bootnode now answers FINDNODE instead of going silent.
• BUG4 — TestSendNeighborsAdvertisesEnrTCPPort: a peer with ENR tcp=40404/udp=30303 is advertised as TCP=40404 (was the UDP port).
• BUG2 — capped so a dual-serve response can't exceed go-ethereum's 5-packet read limit.
• Bound discv4 handler node map to prevent unbounded growth #34 — TestFloodDoesNotEvictBondedPeers: under a flood past MaxNodes, a bonded peer is retained and a fresh peer is still admitted; the map stays bounded and no genuine peer is locked out.

Note

Opening against develop since that is where the parallel rework lives. If you'd rather I retarget once develop merges to master, happy to.