Decentralized snapshot distribution: Statistical trust via threshold consensus

## Summary

Add statistical trust to the decentralized snapshot distribution flow by requiring threshold agreement from multiple peers before accepting an info-hash.

## Background

The POC (step 1) works on an honest-network assumption — any single peer's ENR entry is trusted. This is sufficient for proving the flow but not for production, where a malicious node could advertise bad info-hashes creating a reliability issue.

Statistical trust adds a lightweight defense: collect `{FrozenTx, InfoHash}` from multiple peers and require majority agreement before downloading.

## Design

### Receive Logic

When discovering a new `preverified.toml` version:

1. Collect `{FrozenTx, InfoHash}` from N discovered peers
2. Group by `(FrozenTx, InfoHash)` pair
3. Require threshold agreement (e.g., 3-of-5 peers report same info-hash for a given FrozenTx)
4. Only download and apply if threshold is met
5. Optional: weight by peer diversity (different /24 subnets) to resist localized Sybil attacks

### Properties

- Defeats isolated bad actors advertising false info-hashes
- Does **not** defeat coordinated Sybil attacks (that requires identity trust — step 3)
- Production-viable under an honest-majority assumption
- Cheap to implement — accumulate responses and count, no crypto needed

### Configuration

- `--snap.p2p.trust-threshold` — minimum agreeing peers (default: 3)
- `--snap.p2p.sample-size` — peers to query before deciding (default: 5)

## Acceptance Criteria

- [ ] Discovery reader collects responses from multiple peers
- [ ] Threshold consensus logic implemented
- [ ] Info-hash only accepted when threshold is met
- [ ] Configurable threshold and sample size
- [ ] Graceful fallback to centralized path if threshold cannot be reached