fix(mcp): serialize keepalive probes with RPC calls

[TurgutKural]

Summary

Serialize MCP keepalive RPCs with the per-server _rpc_lock used by normal tool calls.

Problem

MCPServerTask._keepalive_probe() sent ping and the list_tools fallback without acquiring the lock. On stateful JSON-RPC transports, a keepalive request could therefore race an in-flight tools/call response and corrupt the stream, causing false keepalive failures and reconnects.

Fix

• Guard session.send_ping() with _rpc_lock.
• Guard the session.list_tools() fallback with _rpc_lock.
• Add regression tests proving both probes wait while another RPC owns the lock.

Validation

• tests/tools/test_mcp_tool.py: 214 passed
• Related MCP discovery/stability/reconnect tests: 255 passed total
• py_compile: passed
• git diff --check: passed

Scope

This is a targeted serialization fix. Existing reconnect, circuit-breaker, and optional-ping fallback behavior is unchanged.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #62811 — same mechanism: guards the MCP _keepalive_probe (send_ping + list_tools fallback) with the per-server _rpc_lock so it can't race an in-flight tools/call. #62811 is the earliest open PR with this fix (and is itself a narrow current-main salvage of the serialization invariant from the broader #48069). Consolidating on #62811; the reconnect/orphan-cancellation superset lives in #48069.

[tonydwb]

Code Review Summary

Verdict: Approved

Looks Good

• Fix(mcp): serialize keepalive probes with RPC calls
• 45 additions, 2 deletions — targeted fix
• No issues detected

---

Reviewed by Hermes Agent

[TurgutKural]

Closing this duplicate in favor of a follow-up PR based on #62811.

#62811 covers the keepalive/active-RPC serialization race. The follow-up adds transport-aware recovery for empty-message ClosedResourceError/BrokenPipeError failures observed after the stdio session closes, plus regression coverage.

Follow-up will preserve this PR's _rpc_lock serialization and target the same MCP failure class.