[Bug] Fix possible race condition in HCCL handshake

[marcobarlo]

Summary

This PR fixes issue #263.

• Fix a race in Ascend HCCL PD peer handshake: HcclMemRegRequest can run
  before the background accept() thread publishes conn_handles_dict[peer_id].
• Decoder then returns HcclErrorResponse; prefiller crashes with
  AttributeError: 'HcclErrorResponse' object has no attribute 'server_mem_handle_bytes'. PD transfer fails; proxy hangs on KV-ready.
• Scope: HcclChannel._handle_init_msg MemReg path only
  (lmcache_ascend/v1/transfer_channel/hccl_channel.py).

Why

• On HcclInitRequest, the server starts complete_handshake() (background
  accept()), waits only for accept_started_event, then replies Init.
• Client may finish connect() and send MemReg before T3 writes the handle.
• HcclReadyRequest already waits on _peer_ready_events; MemReg did not.
• Fix: wait on the same ready_event (timeout 120s) before looking up
  conn_handles_dict, and raise clearly if the handle is still missing.

Test plan

• Apply only this change.
• Launch Ascend PD disagg (HCCL, prefiller + decoder).
• Send a request through the disagg proxy.
• Inject delay right after accept().
• Decoder no longer logs MemReg KeyError / init-loop failure before
  “Connection established”.
• Prefiller no longer raises on server_mem_handle_bytes; KV transfer
  completes and proxy returns successfully.

Notes

• Threads: T1 lazy_init_peer_connection, T2 _init_loop →
  _handle_init_msg, T3 nested complete_handshake.

[gemini-code-assist]

Code Review

This pull request addresses a race condition in the HCCL handshake timeout (Github issue #263) by waiting for the handshake ready event before retrieving the connection handle. The feedback suggests handling the case where the handshake event is None explicitly and raising a more idiomatic ConnectionError instead of a KeyError when the connection handle is missing.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[matthewygf]

maybe 120s is a bit too long tho, wdyt @chloroethylene

[matthewygf]

Left a cmt.

[matthewygf]

lgtm, and tested