smolvm v1.7.3

What's Changed

• Ship the disk templates zstd-compressed and expand them to sparse files on first use by @BinSquare in #771
• Copy the storage template by seeking between its data extents instead of scanning its whole logical size by @BinSquare in #770
• Harden fused rollout lifecycle by @BinSquare in #801
• Avoid copying CUDA module state for clone channels by @BinSquare in #804
• Load device-resident LoRA policies through managed executors by @BinSquare in #806
• Reject a registry-image machine that has no network at create instead of failing every start with a raw DNS error by @BinSquare in #807
• Perform machine file reads and writes inside the running workload container so uploads are visible to exec by @BinSquare in #810
• Reuse CUDA module handoffs across clone workers by @BinSquare in #808
• Cache pulled OCI images on the host so repeat ephemeral machine runs skip the registry pull by @BinSquare in #805
• Signal guest boot-readiness with an event-driven vsock doorbell as the primary ready signal, keeping the marker file and control-channel ping as fallbacks by @BinSquare in #811
• Map CUDA module handoffs across clone workers by @BinSquare in #814
• Preserve per-device GPU admission headroom by @BinSquare in #826
• Promote CUDA pool readiness and refill improvements by @BinSquare in #828

Full Changelog: v1.7.2...v1.7.3