[RaslaTL19]

Summary

• Finished tasks are released. The vendored ldk-node runtime kept cancellable tasks in a JoinSet that was only drained at shutdown. Every finished inbound connection task therefore stayed tracked until the node stopped. Cancellable tasks are now tracked with tokio_util::task::TaskTracker and a CancellationToken, so a task is released when it ends. This follows upstream ldk-node's "Release completed cancellable tasks" (Release completed cancellable tasks lightningdevkit/ldk-node#997). That change landed on main after 0.7.1 and is not in a release yet.
• Shutdown. Stop closes the tracker, cancels every task in it and waits until each has been dropped. Dropping the runtime also cancels them.
• Late spawns are refused, as upstream does. A closed tracker refuses new cancellable tasks, so a task spawned while a stop is in progress or after it is dropped without being polled. Before, such a task went into a fresh set and could outlive the stop. Node::start opens a fresh set, so restart still works.
• What stop does not cover. It covers tasks spawned through the runtime only. A task that one of them hands to Tokio directly, such as the socket task inside lightning_net_tokio::setup_inbound, is not tracked, as before. That task ends when Node::stop disconnects all peers.
• Concurrent inbound connections are capped. The inbound accept loop moved to inbound.rs. At most MAX_CONCURRENT_INBOUND_CONNECTIONS (512) accepted connections are handled at a time across all listening addresses. A socket accepted beyond that is closed at once.
• What the cap does not touch. It applies only to newly accepted sockets. Established connections and outbound connections, including reconnects to persisted peers, are unaffected.
• Limitation of the cap. 512 is above LDK's own limit of 250 connected peers without a channel, but the extra slots are headroom, not reserved. The peer is not known when a socket is accepted. So while 512 accepted connections are open, a new inbound connection from a channel peer is refused too. A channel peer with a stored address is still reached by the outbound reconnect. One without a stored address has to retry until a slot frees.
• Accept errors no longer stop the listener. A failed accept is logged and retried after 100 ms. A failed socket conversion ends only that connection.
• Dependency. tokio-util (rt feature) is added to the vendored manifest. It is already in the root lock at 0.7.18, and the vendor lock pins the same version.
• Docs. BITSOV-PATCH.md documents the delta.

Tests

• inbound::tests::finished_inbound_connection_tasks_are_released: 200 connect/close cycles on an ephemeral loopback port. On the previous runtime, 201 tasks stayed tracked (200 finished connections plus the accept loop). Now only the accept loop remains.
• inbound::tests::connections_beyond_the_limit_are_closed_and_slots_are_reused: with 4 slots, a 5th socket is closed, the 4 open ones stay up, and a freed slot is reused.
• inbound::tests::stopping_cancels_open_inbound_connections
• runtime::tests::task_spawned_while_stopping_does_not_outlive_the_stop: a running task spawns a child after the stop has begun. The child is never polled and nothing stays tracked after the stop. With the previous fresh-set behaviour, one task stayed tracked.
• runtime::tests: finished tasks are released; stop and drop cancel live tasks; a task spawned after a stop is refused; spawning works again after a restart.
• The inbound test fixture stops its tasks before dropping the runtime, so --nocapture output is clean.

cargo test --locked --manifest-path vendor/ldk-node/Cargo.toml --lib
cargo test --locked -p konsensus-lightning
cargo test --locked -p konsensus-node --bins
cargo clippy --locked -p konsensus-lightning -p konsensus-node --all-targets -- -D warnings