fix(p2p): synchronize access to the inbound listener

[0xrlawrence]

Description

p.listener is written by ListenForInboundPeers from the goroutine Start() spawns, and read unsynchronized by Stop() and by several test helpers. The race detector reports it as the single largest source of races in this package.

The practical consequence is at shutdown: Stop() can observe a nil listener that has in fact already been created, skip the Close(), and leak the socket — while the accept loop keeps running against a listener nobody closed.

Changes Made

• Guard the field with a dedicated listenerMu.
• Publish the listener under the lock, then run the accept loop against a local reference so the hot accept path stays lock-free.
• Read it under the lock in Stop().
• Add ListenerAddress(), a synchronized accessor, and use it in the tests that previously reached into n.listener directly.

Testing

go test -race ./p2p/: data races in this package drop from 65 to 8, and no remaining race involves the listener.

The 8 that remain are pre-existing and unrelated — Stream/MultiConn state in conn.go and encrypt.go. Left for separate changes rather than widening this one.

go test ./p2p/ passes.

Note for reviewers

This touches ListenForInboundPeers, which #521 also edits. The two changes are independent but will need a trivial rebase depending on merge order.

🤖 Generated with Claude Code