[danbugs]

Cuts v0.21.0, and builds against the hyperlight 0.19.0 release instead of the hyperlight main git pin, so the crate can go to crates.io again (closes #233). The release carries snapshot resizing (AppSandbox::snapshot_with_scratch_mb, #162), AppSandbox::dirty_bytes (#259), SandboxBuilder::on_interrupt_handle (#264), the kernel's scratch fixes, name lookups with AI_ADDRCONFIG under a restricted network policy (#248), the hluk build command checks (#227, #235, #236, #237), the Python CA bundle in the agent and python-shell images (#245) and /proc/mounts for PowerShell (#246).

0.21.0 rather than 0.20.2: the library gains pub items (snapshot_with_scratch_mb, dirty_bytes, on_interrupt_handle), hyperlight-host moves from 0.18 to 0.19, and snapshots from 0.20.1 are refused. No hluk-defined pub item is removed or renamed; those that carry hyperlight_host types now carry 0.19's: HostFunctions, Snapshot (and what takes or returns one), Error::Hyperlight, and the interrupt handle AppSandbox::interrupt_handle returns and SandboxBuilder::on_interrupt_handle hands over. The crate also re-exports hyperlight_host itself, now 0.19, so code that uses it may need changes: its SandboxConfiguration::set_heap_size, SandboxConfiguration::DEFAULT_HEAP_SIZE, SandboxBuilder::heap_size and SandboxBuilder::get_heap_size (hyperlight's builder, not hluk's) are gone. For Hyperlight 0.19's other breaking changes (among them PtRootFinder gets a SnapshotMemoryReader and returns Result<Vec<u64>>, Sandbox::snapshot removes map_region and map_file_cow regions, snapshot ABI 9), see its CHANGELOG. A NetworkFilter is now asked at each send on a connected UDP socket rather than at its connect(), as it already was at an unconnected sendto() (#248).

Hyperlight 0.19 gives the guest no heap region (hyperlight-dev/hyperlight#1968), and the kernel took its 1 MiB boot stack from it, so on 0.19 every guest crashed at boot ("Boot stack alloc failed"). The kernel bump here (unikraft plat-hyperlight-v2 9f414729 -> 3b7cba88) keeps the boot stack in the kernel's .bss, which a snapshot keeps as it kept the heap region, with a canary in its lowest word, and makes the kernel's PEB match 0.19's. The halt that ends boot checks the canary, so an overflow during boot fails the boot. Each later entry from the host, every guest function call and the checkpoint before a snapshot, checks it too, catching later use of the stack by the init thread (a native main() without LIBUKBOOT_MAINTHREAD that outlives the boot halt) or a stray write, so such a guest fails a snapshot rather than being captured (a snapshot with no call since the last snapshot or restore enters no guest code, and captures memory the boot halt or a checkpoint has already checked). The end of a call's entry, before its result, checks it for the call's own work, so a call that ran the stack down fails rather than returning. A shutdown checks it before reporting the exit, so a call in which the stack was run down and the guest then shut down fails rather than ending cleanly. At each of the four points the guest crashes, assertions or not, and without UK_CRASH, whose reboot crash action would re-enter the shutdown's check. At the boot halt, in a shutdown and at the start and end of an ordinary call's entry the report ("Boot stack canary overwritten (stack overflow?)") goes to the output hluk collects (take_output, drain_output, the echo); for any other entry, which may be a checkpoint, it goes out through DebugPrint, since the host refuses a host call (the console's HostWriteErr and HostWrite included) during a checkpoint and would otherwise end the run with HostCallDuringCheckpoint before the report reached it. plat/hyperlight now selects paging, the only way it has to allocate, so the defconfigs no longer set it. The kernel boots on hyperlight 0.19.0 and on the hyperlight main this tree built against before (6c17f7cd).

Five commits:

• kernel: bump the kernel and rebuild both kernels and the native-kernel fixture (just build-kernel, just build-native-kernel; just verify-kernel and just verify-native-kernel match). defconfig-elfloader and the native fixture's defconfig drop LIBUKPAGING and LIBUKFALLOC: plat/hyperlight now selects paging, which selects LIBUKFALLOC; the resolved .config and the kernels are unchanged by it. plat/hyperlight also refuses a build without a boot allocator, and drops pagetable64.S, whose only symbol (bpt_unmap_mrd) nothing reads with paging on. The CHANGELOG notes these for native-kernel builders. hluk stops sizing the heap: HEAP_SIZE and the two set_heap_size calls go, as nothing uses the region now, and the SNAPSHOT_CONTRACT doc in build.rs no longer lists a heap (the hyperlight main the tree builds against at this commit, 6c17f7cd, still maps one, at its default size, unused).
• hyperlight 0.19: hyperlight-host and hyperlight-common move to 0.19, from crates.io: the [patch.crates-io] git pin goes from Cargo.toml, integrations/clickhouse/Cargo.toml and demos/pptx-gen/Cargo.toml, and the three lockfiles take 0.19.0 (in re-resolving them cargo also moves some crates' windows-sys dependency to versions already in the locks, mostly 0.61.2). The CHANGELOG notes the move, as earlier bumps were noted, what the re-exported hyperlight_host 0.19 changes, that a native kernel built against an earlier plat/hyperlight must be rebuilt, and, folding in the kernel commit's note, that plat/hyperlight now selects paging, refuses a build without a boot allocator, and crashes the guest when it finds the boot stack's canary overwritten, so the transport arena entry drops its own rebuild sentence. SNAPSHOT_CONTRACT stays 8: it was bumped for the transport arena after v0.20.1 and has not been released. The dev builds since then carry it, and their snapshots are refused all the same: the kernel change rolls the key's other half (k<hash>, the embedded kernel's SHA-256), and Hyperlight's own snapshot ABI 9 refuses them too.
• CHANGELOG backfill: the entries Fix restricted DNS address-family probing #248 lacked (the AI_ADDRCONFIG fix under Fixed, the NetworkFilter being asked at each connected UDP send under Changed), and ci: stop a wedged HTTP demo from hanging the whole job #260 (a wedged HTTP demo fails once its five-minute wait for an answer runs out, instead of holding the arm64 nightly's demos job to its 90-minute timeout) and ci(arm64): soft-fail the known nested-virt failures on linux #263 (the known nested-virt failures on the Linux arm64 runner no longer fail the run) under Changed, as v0.19.1 noted its CI change. Give the nested-virt arm64 runner room on the timeouts it still blows #244 only gives tests and examples longer timeouts on the slow runner, so it gets no entry.
• justfile: just changelog-notes, which release.yml runs for the release notes, took the newest other v* tag as the previous release for GitHub's PR list, and verl-hyperlight-v0.1.1 sorts above v0.20.1. It now takes the release just before the given one among the vX.Y.Z tags (v0.21.0 compares against v0.20.1, v0.20.1 against v0.20.0), sorted field by field with sort -t. -k1,1n -k2,2n -k3,3n rather than sort -V, which older macOS lacks, and matched as fixed strings; it also passes GitHub the v-prefixed tag when the version is given without its v. A release-tooling fix, which the CHANGELOG has never recorded, so it gets no entry.
• Release v0.21.0, as for v0.20.1 and v0.20.0: bump version to 0.21.0 in Cargo.toml and the three lockfiles (hyperlight-unikraft only); rename the ## [Prerelease] CHANGELOG heading to ## [v0.21.0], with a fresh empty Prerelease above, and note under Changed that snapshots from v0.20.1 and earlier are refused (the transport arena entry no longer says so as well); bump the illustrative v0.20.1 image/version examples in docs/manifest.md, docs/templates.md and the install scripts to v0.21.0.

No Python package needs a new floor: verl-hyperlight and openai-agents-hyperlight need hluk 0.19.0, smolagents-hyperlight (and so mcp-server-hyperlight) hluk 0.18.0.

After merge: Create release on main, then Publish to crates.io with dry_run = false, now that the crate has no git dependency.

[Copilot]

🔵 Needs a closer look

The Hyperlight 0.19 boot-stack ABI change requires rebuilt-kernel correspondence checks and cross-architecture boot, snapshot, and restore validation beyond the supplied textual evidence.

No open issues 🎉

What changed in this PR

Prepares v0.21.0 for crates.io publication by replacing the Hyperlight git pin with released 0.19 dependencies, addressing #233. Aligns host configuration with the updated kernel and documents release compatibility changes.

Changes:

• Bumps the crate to 0.21.0, adopts Hyperlight 0.19, and removes git dependency overrides.
• Removes obsolete heap sizing and redundant paging settings.
• Updates release notes and version examples, and fixes previous-release tag selection for generated notes.

File	Description
tests/​fixtures/​native-kernel/​README.md	Clarifies that the platform selects paging.
tests/​fixtures/​native-kernel/​defconfig	Removes redundant paging and allocator selections.
src/​lib.rs	Removes obsolete guest heap sizing.
justfile	Selects the preceding release tag and normalizes its prefix.
integrations/​clickhouse/​Cargo.toml	Removes Hyperlight git overrides.
integrations/​clickhouse/​Cargo.lock	Records released Hyperlight dependencies and crate version.
install.sh	Updates the release-version example comment.
install.ps1	Updates the release-version example comment.
docs/​templates.md	Updates template substitution examples to 0.21.0.
docs/​manifest.md	Updates image, cache, and version examples.
demos/​pptx-gen/​Cargo.toml	Removes Hyperlight git overrides.
demos/​pptx-gen/​Cargo.lock	Records released Hyperlight dependencies and crate version.
defconfig-elfloader	Relies on platform-selected paging and allocator support.
CHANGELOG.md	Adds the release heading, compatibility notes, and missing entries.
Cargo.toml	Sets version 0.21.0 and adopts Hyperlight 0.19 without git overrides.
Cargo.lock	Locks released Hyperlight packages and version 0.21.0.
build.rs	Removes heap layout from the snapshot-contract documentation.

_{Review effort: Balanced}