[Code-Hex]

Why

vmnetsock.Datagram currently calls WritePacket for every datagram received from the peer. This change collects up to 32 queued datagrams and passes them to WritePackets together. It returns as soon as the socket has no more data, so a single packet does not wait for a full batch.

Both directions use public recvmsg and sendmsg calls inside RawConn callbacks. Each socket syscall still handles one datagram. Batching received packets reduces calls into vmnet. Each callback also holds Go's I/O lock across several packets.

In three local QEMU/HVF Linux guest runs on macOS 27.0.1 with MTU 1500, median guest-to-host TCP throughput rose from 3.19 to 5.76 Gbps. Host-to-guest throughput was 3.41 Gbps before and 3.43 Gbps after. These measurements cover one host and guest configuration.

Behavior

• Preserve packet boundaries and forward valid packets before reporting an invalid trailing packet.
• Keep the send position across EAGAIN waits to avoid duplicate packets.
• On ENOBUFS, discard the unsent remainder of the batch and continue with the next batch.
• Keep the existing Dup and Close ownership rules and shutdown path. The adapter adds no goroutines and never stops its borrowed interface.

Real-socket tests cover truncation, partial sends, read deadlines, and close. Native race tests cover cancellation, backpressure, and concurrent Dup and Close.

Related to #205.