fix(net): classify Invalid-block penalty + peer lifecycle metrics

[constwz]

Summary

Two coupled changes targeting #320 (peers drop to zero after sync):

1. Categorized Invalid handling in block_import/service.rs — replace fix: do not penalize peers for Invalid blocks from new_payload #297's blanket skip with error-string classification. Race-shaped failures (state root, parent block, unknown parent, missing trie, receipts root, insert_canonical, reorg) skip the penalty path; structural failures keep the BadBlock mechanism so real offenders are still caught. Aligns with geth's fetcher semantics (drop on header verify failure, not execution failure).
2. Peer lifecycle summary metrics (node/network/peer_metrics.rs) — subscribes to NetworkEventListenerProvider's PeerEvent stream and emits one INFO line per 60 s window with opens/closes bucketed by DisconnectReason. Idle windows stay silent. Purpose: in prod we had no visibility into why peers were disconnecting.

Why re-tighten after #297?

#297 suppressed the BadBlock penalty for every PayloadStatusEnum::Invalid. That fixes the drain caused by concurrent-reorg races (which is real — 4 race-Invalids bans a peer, recovery +1/sec ≈ 4.5 h to re-earn), but also lets genuinely malformed blocks from malicious or broken peers escape the reputation system. Classifying by error string lets us keep the race-race immunity without surrendering the structural safety net.

Example metrics output

INFO bsc::peer_metrics window_secs=60 opened=3 closed=15 \
     closed_by=disconnect_requested:8,too_many_peers:4,ping_timeout:3 \
     "peer lifecycle summary"

Test plan

• cargo check — pass
• cargo clippy --tests -- -D warnings — pass
• cargo test --lib node::network::block_import::service::tests::can_handle_invalid_new_payload — pass (structural branch still emits Err, BadBlock path preserved)
• cargo test --lib node::network::block_import::service::tests::reorg_race_invalid_does_not_penalize — new test, pass (race branch emits no Err outcome)
• Observe bsc::peer_metrics INFO line on a running testnet/mainnet node; confirm one line per minute only when there is activity, silent during steady state.

Refs #320, #297.

🤖 Generated with Claude Code

[zhk101]

reth is already logging the peer disconnect reason as metrics

reth_network_useless_peer
reth_network_subprotocol_specific
reth_network_already_connected
reth_network_client_quitting
reth_network_unexpected_identity
reth_network_disconnect_requested
reth_network_null_node_identity
reth_network_tcp_subsystem_error
reth_network_incompatible
reth_network_protocol_breach
reth_network_too_many_peers

[constwz]

@zhk101 thanks for catching this — you're right that reth's DisconnectMetrics already covers per-DisconnectReason counters as Prometheus metrics, so a separate log aggregator on top of the same data is redundant for any node that scrapes /metrics.

I've reworked the PR (commit `23d7248`):

Removed: src/node/network/peer_metrics.rs and its main.rs / mod.rs wiring.

Added instead: a BscBlockImportMetrics struct in src/metrics.rs (same #[derive(Metrics)] pattern as the other BSC metric scopes) that exposes per-outcome counters for engine.new_payload results:

```
bsc_block_import_invalid_race # race-shaped Invalid → no penalty
bsc_block_import_invalid_structural # structural Invalid → BadBlock applied
bsc_block_import_valid # accepted by engine
bsc_block_import_syncing # engine syncing, FCU forced
bsc_block_import_hash_mismatch # rejected before reaching engine
```

This fills the actual observability gap (the block-import path had zero metrics before this), and makes the race-vs-structural ratio — which is the central tuning knob for the classifier in this PR — directly visible in Prometheus instead of grep'd from logs.

Also rebased onto current `develop` to clear the merge conflict (the `fork_recover` refactor in #334 reshaped `ImportService` significantly; resolved by keeping the new `recovering_heads` / `failed_heads` / `announce_interval` fields and adding the `metrics` field alongside them). All 14 `block_import::service::tests` pass locally including the two new classification tests.