Task 3.1.2: Add simnet-based in-process testbed

[srene]

Parent

Sub-issue of #10 (Task 3.1: Prepare testbed environment)

Description

Add an in-process discv5 testbed that runs N nodes inside a single Go program, wired through github.com/marcopolo/simnet as the simulated UDP transport. Pairs with github.com/marcopolo/ethereum-network-latencies for realistic per-pair latencies derived from real-world Ethereum mainnet measurements.

This is a parallel testbed to the Docker-based one (Task 3.1.1, #17). Both have a place:

	Docker testbed	simnet testbed
Real OS networking	✓	✗
Runs on macOS	✗ (needs Linux for tc)	✓
Runs in CI without root	✗	✓
Time to spin up 100 nodes	~30 s (Docker overhead)	<1 s
Realistic per-pair latencies	manual via tc filter rules	natively supported via LatencyFunc
Deterministic	✗	✓ (with simnet's synctest integration)
Single-process debugging	hard	✓

The simnet testbed is well-suited for protocol-correctness testing, scale tests beyond what Docker can boot, and reproducible CI runs. The Docker testbed remains useful for final integration smoke tests against real OS sockets.

Scope

• New Go module under testbed/simnet/ to keep its dependencies (simnet, ethereum-network-latencies) out of the main go.mod. Mirrors the cmd/keeper precedent.
• Adapter shim translating between simnet's net.PacketConn (using net.Addr) and discv5's UDPConn interface (using netip.AddrPort).
• CLI driving N node instances with configurable per-pair latency, bandwidth, and topology.
• Realistic latencies via the MarcoPolo/ethereum-network-latencies dataset (~7000 masked Ethereum-mainnet IPs with continent encoding and pairwise predicted RTTs from a gradient-boosted decision tree trained on RIPE Atlas + WonderNetworks data).
• Topic registration / search workload runner.
• Per-node structured JSON logging compatible with the existing testbed/analyse.py analysis pipeline.

Verified scale

Scaffolding so far is verified to bootstrap cleanly at:

• 1000 nodes — 324 MB memory, 1.3 s user CPU
• 2000 nodes — 818 MB memory, 3.5 s user CPU

Extrapolating: ~5K nodes comfortable on a 16 GB laptop, ~20K on a 64 GB Linux server. Materially more capacity than Docker, which tops out around 100-300 nodes per host.

Phase

Phase 3 — Weeks 16-24