rdma_accept_handshake can panic via unreachable!() under simultaneous mutual handshake

[xiaguan]

Problem

PegaEngine::rdma_accept_handshake (pegaflow-core/src/lib.rs) does:

1. invalidate_connection(client_addr) — tear down any stale connection
2. get_or_prepare(client_addr) — expects Prepared, treats Existing as unreachable!("just invalidated connection")

Steps 1 and 2 take the backend lock separately. If two nodes handshake with each other simultaneously (exactly the deployment shape of the transfer-lock incident: peers booting together and cross-fetching), the local node's outbound complete_handshake_for can insert addr_connections[client_addr] in the window between the two calls. get_or_prepare then returns Existing and the unreachable! panics the gRPC handler.

Impact

Panic in the accept path; the peer's handshake fails and retries. Low probability per attempt but structurally reachable, and mutual-handshake bursts are the norm after a rolling restart.

Suggested fix

Replace the unreachable! with an error return (peer retries and then reuses the fresh connection), or make invalidate+prepare atomic under one lock.

Context

Flagged during the #399 review (pre-existing).

🤖 Generated with Claude Code