fix(agent): use is_peer_closed_error for RPC EOF detection

handle_connection treated any read_message error whose string contained
"failed to read message header" as a clean EOF. That substring comes
from anyhow::Context and prefixes *every* header-read failure — clean
EOF, ConnectionReset, BrokenPipe, parse errors, you name it — so real
errors were silently swallowed and never logged.

Switch to the existing `vsock::is_peer_closed_error` helper, which
inspects the underlying io::Error kind for BrokenPipe / ConnectionReset
/ ConnectionAborted / UnexpectedEof. Anything else now propagates and
gets logged at error level by Agent::run.

Pre-existing in master; surfaced by Copilot review on PR #276.

Author: AprilNEA

guest/arcbox-agent/src/agent/linux/rpc.rs

@@ -24,6 +24,7 @@ use super::kubernetes::{
 use super::runtime::{handle_ensure_runtime, handle_runtime_status};
 use super::sandbox::handle_sandbox_message;
 use super::system_info::handle_get_system_info;
+use super::vsock::is_peer_closed_error;
 
 /// Result from handling a request.
 enum RequestResult {
@@ -43,8 +44,10 @@ where
         let (msg_type, trace_id, payload) = match read_message(&mut stream).await {
             Ok(msg) => msg,
             Err(e) => {
-                // Check if it's an EOF (clean disconnect)
-                if e.to_string().contains("failed to read message header") {
+                // Treat peer-closed errors as a clean disconnect; surface
+                // anything else (parse failure, unexpected I/O) to the
+                // caller so it's logged.
+                if is_peer_closed_error(&e) {
                     tracing::debug!("Client disconnected");
                     return Ok(());
                 }