v0.3.1

Private-network installs use fewer backing reads, allocations, and socket writes.

• Small dd blocks are grouped into about 32 KiB transfers while preserving count, skip, record accounting, and foreground interruption.
• QEMU private networking batches up to 32 frames already queued and flushes immediately when the queue empties. Receive scratch is reused after packets are copied into gVisor-owned storage.
• IRIX SIGINT requests interrupt supported foreground file transfers, allowing the installer to seek without first receiving the rest of each file. Background sessions and unsupported pipelines keep their previous signal behavior.
• A complete QEMU frame header with a missing body is reported as a truncated frame.

In one fresh-disk QEMU/IRIX comparison on macOS, the grouped-transfer and framing
changes reduced the package phase from 27m45s to 22m51s. Median fixed-transfer
throughput increased from 25.3 to 31.9 MB/s. All six guest checksums matched, and
both installations rebooted to serial login.

This comparison predates the SIGINT fix and does not measure its additional
effect. Both installers logged the same package diagnostics. Details and limits
are in docs/qemu-network-performance.md.

Binaries: Linux and macOS (amd64/arm64), Windows (amd64).