[reiase]

Build pvisor.exe on Windows for direct Linux VM execution, using the same vm-run arguments, guest configuration and authenticated exit handling as Linux and Apple Silicon macOS.

Code

• Add pvisor vm-run --rootfs ROOT --cpus N --memory MiB [--readonly DIR] [--firmware LIBRARY] -- /guest-command ... through pvisor_vm::api. Rootfs writes persist directly.
• Exclude Unix runtime dependencies from the Windows build. Build with cargo build --locked -p pvisor-cli --bin pvisor; reject unported commands.
• Fix Windows directory rename leaving cached descendants bound to obsolete paths. Preserve inode IDs and lookup counts, with native and guest-visible regressions.
• Fix OverlayNet reactor starvation under continuous peer traffic. Stabilize cache lifecycle tests by excluding error-backtrace symbolization from their unchanged deadlines, and check the reported exit cause.

CI

• Windows: native MSVC CLI build/parser tests and existing VM contracts; build repository firmware on Linux, then run the actual CLI and WHP API smoke on Windows.
• Linux: run the same CLI/guest workload on KVM. Store execution evidence under RUNNER_TEMP, outside restored Cargo caches; existing evidence directories remain protected against overwrite.
• macOS: retain product builds/signing and conventional tests. Remove the real HVF workload from hosted CI at the user's request: GitHub's arm64 runners do not expose nested virtualization. Keep the common driver available for manual execution on a capable Apple Silicon Mac. Hosted macOS CI does not establish real HVF acceptance.
• Each VM suite runs 12 boots by default, with 1/2 vCPU, exit codes 0/7/255, concurrent file operations, memory touches, timers, read-only rejection and cached directory rename. Host oracles check exact file results as well as exit status. Failed runs retain logs and reports.

Validation

During development:

• Local HVF: 12 default boots plus 120 additional boots passed.
• CLI: 129 tests; VM API/boundary: 14 tests; driver: 5 Python regressions passed.
• OverlayNet: all 98 tests passed; the unchanged flood regression passed 50 consecutive runs.
• Cache service: all five tests passed for ten rounds with outer backtraces enabled.
• Windows MSVC cross-checks, workflow lint, formatting and strict bilingual docs builds passed.
• Remote CI at cda63d4b: Windows native builds/WHP, Linux/KVM and macOS conventional tests passed. The Linux Go contract shard initially failed when log observation saw KVM_EXIT_FAIL_ENTRY (reason 18446744073709551615). Re-running that shard on the same commit passed the unchanged concurrent-start regression and the complete shard; the CI aggregate passed. The intermittent failure is not yet explained or fixed. Initial failure, passing rerun.

Support scope

Windows support covers vm-run, caller-owned rootfs and an optional read-only share. Job/container workflows, stage/review/apply, networking, snapshots, TUI/replay/Daemon and Windows packaging remain unported. Linux/macOS installations contain four host programs; Windows still needs the firmware DLL and a Linux rootfs/workload.

The platform document explains component dependencies, implementation differences and coverage. Hours-long stability, resource-leak baselines and concurrent VMs remain unverified. Strict local macOS Clippy for the full CLI was blocked by existing unused variables in pvisor-overlay-core/src/apply.rs:1698–1699.