DirectDispatcher: resync coverage tracer after each request

A handler-side exception (even one caught and converted to a result, as
MCPServer's tool wrapper does) desyncs CPython 3.11's CTracer when the
DirectDispatcher request path resumes the awaiting coroutine chain via
throw — the same python/cpython#106749 the other transport seams already
work around. The resync_tracer() checkpoint at the end of
_dispatch_request restores tracing for the caller's subsequent lines.

Author: maxisbey

src/mcp/shared/direct_dispatcher.py

@@ -26,6 +26,7 @@
 import anyio.abc
 from pydantic import ValidationError
 
+from mcp.shared._compat import resync_tracer
 from mcp.shared.dispatcher import CallOptions, OnNotify, OnRequest, ProgressFnT
 from mcp.shared.exceptions import MCPError, NoBackChannelError
 from mcp.shared.message import MessageMetadata
@@ -251,6 +252,8 @@ async def _dispatch_request(
                 code=REQUEST_TIMEOUT,
                 message=f"Timed out after {opts.get('timeout')}s waiting for {method!r}",
             ) from None
+        finally:
+            await resync_tracer()
 
     async def _dispatch_notify(self, method: str, params: Mapping[str, Any] | None) -> None:
         try: