feat: add phys-bits (MAXPHYADDR) as a second live-migration safety axis

[maxgashkov]

Add phys-bits (MAXPHYADDR) as a second live-migration safety axis

Problem

ProxCLMC computes the lowest common ISA baseline (x86-64-vN). That's necessary but not sufficient: the guest's advertised physical-address width (phys-bits / MAXPHYADDR) must also be <= every host's. The two axes are independent — a cluster cleanly standardized on x86-64-v2-AES can still corrupt guests on live migration.

An unpinned guest on a named CPU model advertises QEMU's default 40 phys-bits. On a mixed cluster (e.g. Intel nodes at 39-bit, AMD at 48-bit) a VM booted on a node at least 40 bits wide can place its virtio MMIO BARs above a narrower node's limit. The aperture is fixed at boot and travels unchanged across a live migration; when it lands on the narrow node, the next virtqueue kick faults (#PF, reserved bit) and the guest kernel-panics. Offline migration, HA failover and reboots cold-boot on the destination and are unaffected.

Change

• Parse the address sizes line from /proc/cpuinfo into a per-node phys-bits.
• Compute the cluster's lowest common phys-bits (undetected nodes are skipped).
• Show phys-bits as a fourth column in the node listing.
• When a node sits below the guest default — min < min(GUEST_DEFAULT_PHYS_BITS, max) — pin the reported CPU type to the cluster minimum, e.g. Cluster CPU type: x86-64-v3,phys-bits=39. --list-only carries the same suffix.
• Output is unchanged for uniform clusters and clusters all at or above the guest default.

GUEST_DEFAULT_PHYS_BITS is hardcoded to 40 (QEMU's default for named models, verified on pve-qemu 11.0.0 via qom-get/query-cpu-model-expansion). If QEMU ever raises that default this can under-report until the constant is bumped.

Testing

• cargo test — unit tests for cpuinfo parsing, cluster min/max, and the pin predicate.
• cargo clippy clean.
• Run live on a 3-node cluster (2x Intel 39-bit, 1x AMD 48-bit): correctly detects the mismatch and recommends phys-bits=39:

./proxclmc
Detected nodes:
prox1 | 10.0.1.10 | x86-64-v3 | 39
prox2 | 10.0.1.11 | x86-64-v3 | 39
prox3 | 10.0.1.12 | x86-64-v3 | 48

Cluster CPU type: x86-64-v3,phys-bits=39

./proxclmc --list-only
x86-64-v3,phys-bits=39

Notes

• No new CLI flags or dependencies.
• LLM used (Opus 4.8) but it was heavily steered and the result edited by a human.