feat: run benches in a libkrun microVM for predictable numbers

[enekos]

What

Opt-in --isolation microvm runs a bench inside an ephemeral libkrun microVM (via the krunvm CLI) instead of on the host — pinned vCPUs, fixed RAM, clean rootfs — so numbers don't drift with host load. One code path on macOS (Hypervisor.framework) and Linux (KVM). No libkrun FFI, no new build-time dependency; we just shell out to krunvm.

Default stays local — nothing changes unless you ask for the VM.

Design

• New sandbox module — Isolation::{Local, Microvm}, resolves VmOpts, boots an ephemeral guest, mounts the workdir 1:1 + a persistent build cache, execs through /bin/sh with a normalized PATH, and deletes the VM on drop (no leaks on error/panic).

• Wired into both bench entry points through their single choke points:

  • aatxe perf-vs — both A/B sides run in matching guests, so the comparison stays fair.
  • aatxe run — the shared run_runner all three language adapters (TS/Go/Rust) funnel through, so they all get it for free.

• Only the knobs that move numbers are surfaced (everything else = libkrun defaults):

  flag (env)	default
  --isolation	local
  --vm-cpus (AATXE_VM_CPUS)	2
  --vm-mem (AATXE_VM_MEM)	2048
  --vm-image (AATXE_VM_IMAGE)	rust:1 / golang:1 / node:22

Plug-and-play

make microvm-setup     # one-time: install krunvm (Homebrew on macOS)
make microvm-doctor    # readiness check
make perf-vs-vm        # = make perf-vs ISOLATION=microvm

Missing krunvm fails with an install hint, not a cryptic spawn error. Build caches (cargo/go/npm) persist under ~/.cache/aatxe/vm, isolated from the host's own target/ (different arch), so only the first run pays the image-pull + cold-build cost.

Testing

• cargo fmt --check, clippy -D warnings, full suite green (112 tests, incl. 9 new sandbox unit tests on argv construction / shell-quoting / config resolution).
• Verified end-to-end on macOS that the flag is exposed on both commands and the microVM path surfaces the actionable error through the full stack; the refactored local run path still captures/parses/writes and propagates env+cwd.

Caveat

The live VM boot path (krunvm create/start/delete) is unit-tested at the argv level but not exercised live — krunvm isn't installed on my box. All krunvm invocation is centralized in sandbox.rs (create_argv / guest_argv), so if the installed version names a flag differently it's a one-line fix. Validate with make microvm-setup → make perf-vs-vm.