fix: harden V3 isolation and add reproducible VM regression

[loverustfs]

V3 CPU contracts could discard live replay records, accept unrelated TCP batch replies, and release queue resources before an unknown completion was drained. This change rejects those cases, prevents reads from binding to future slots, and preserves tenant-scoped reclamation and monotonic identifiers.

The VM runner now reproduces the CPU checks with locked dependencies, bounded command timeouts, descendant cleanup, versioned JSON, JUnit and raw logs. It checks exact output contracts and dependency boundaries, disables live RustFS probes even through Cargo configuration, and refuses unsupported hardware profiles. English roadmap, replacement criteria, PR plan and VM instructions define the next integration gates.

Validation

• python3 scripts/ci/vm_regression.py --profile cpu --offline --output artifacts/vm-regression-final: all 18 gates passed on macOS arm64 with Rust 1.98.0 and Python 3.9.6.
• The default workspace suite passed 181 Rust tests; runner tests passed 30 Python tests. Default/all-feature Clippy, doc tests, optional/combined features and all CPU demos passed.
• GitHub CI passed on Linux x86_64, native aarch64 and aarch64 QEMU for both push and pull-request runs.
• Regression coverage includes replay capacity/clock/domain checks, credential retirement and redaction, identity validation, restart-epoch exhaustion, reply correlation, quarantine/drain, future-slot aliasing, fault accounting and Cargo endpoint injection.

Compatibility and limits

Wire/key formats remain unchanged. No dependency or unsafe code was added. Existing V1/V2 contracts remain in place; wrong or incomplete data never becomes a hit.

This is CPU evidence, including simulated DPU contracts. Native RDMA, GPU engine export/import, GDS, DOCA/SPDK and BF3 qualification remain NOT_RUN_HARDWARE. The V3 ACL/replay/rate hooks are not a completed TCP security rollout. There are no hot-tier erasure-code or durability replicas. CI uses the same runner and retains its artifacts. The separate Grok VM backtest remains pending.