smolvm v1.7.2

What's Changed

• Fix CUDA fork compatibility for vLLM workloads by @BinSquare in #764
• Read older .smolmachine pack formats and point to a rebuild when a legacy pack cannot run by @BinSquare in #768
• Remove agent helpers whose only remaining callers are their own unit tests by @NickyHeC in #760
• Reclaim case-sensitive pack volumes on macOS that were left mounted by runs killed before releasing their lease by @BinSquare in #769
• ci: widen Linux clippy/tests and guard the CUDA guest build by @NickyHeC in #761
• fix: stream progress during detached local-archive start by @NickyHeC in #699
• cudart: stub Runtime exports required by conda libtorch_cuda by @NickyHeC in #638
• Construct the archive worker panic error with Error::other so the widened Linux clippy run passes by @BinSquare in #772
• Improve CUDA fork-pool capacity and worker lifecycle by @BinSquare in #774
• Pass the executable stub path rather than the sidecar name when exporting a machine by @BinSquare in #780
• Promote CUDA golden eviction, auto-graph, and clone restore fixes by @BinSquare in #778
• Probe the guest agent socket from the start of the readiness wait so a marker that cannot be written no longer costs five seconds per boot by @BinSquare in #773
• Add transactional batch forks by @BinSquare in #783
• Fix virtio-net guest→host connections via the gateway IP blackholing after the handshake by @BinSquare in #784
• tiny nit: Stop reporting a skipped test suite as all-passed by @Bnjoroge1 in #790
• Restrict the node's plain-HTTP loopback port to liveness routes by @BinSquare in #791
• Harden the registry pull path against SSRF: reject repository traversal and private pull hosts by @BinSquare in #792
• Confirm VM exit before deleting storage by @Bnjoroge1 in #788
• Add one-shot held fork slots for hot pool reuse by @BinSquare in #786
• Rebuild the macOS libkrun.dylib with a static ELF guest init so a source checkout can boot VMs by @BinSquare in #793
• Add automatic held-fork pool leases by @BinSquare in #794
• Batch automatic pool refills by @BinSquare in #795
• Stage lease payloads before worker release by @BinSquare in #796
• Bump libkrunfw to the connected datagram socket network-namespace fix and rebuild the bundled libraries by @BinSquare in #797
• Add automatic CUDA admission and fused rollouts by @BinSquare in #798
• Bump the workspace to 1.7.2 by @BinSquare in #799

Full Changelog: v1.7.1...v1.7.2