feat(cubesandbox): lifecycle + ephemeral + snapshot capability (Phase 2, stacked on #137)

[JackWeiw]

What

Phase 2 of the cubesandbox provider, stacked on #137 (Phase 1 exec-only). This lifts the provider from exec-only to the full replay capability surface, mirroring aenv's lifecycle shape over the native CubeSandbox pause/resume (the Cloud Hypervisor fork writes a memory snapshot the resume restores from -- the same memory-reuse oversubscription shape as aenv).

Scope (Phase 2 — full capability)

CubesandboxProvider gains the three replay Protocols:

• LifecycleCapable — pause = native state.cube_sandbox.pause(wait=True) (synchronous: the snapshot must be stable before the runner accounts the pause); resume = state.cube_sandbox = Sandbox.connect(inst.id) (cube's connect auto-resumes a paused sandbox -- the deprecated standalone resume() is replaced by connect -- and returns a fresh handle, swapped in like the manager's _attach).
• EphemeralCapable — create_one / kill_one (trajectory mode). Mirrors aenv verbatim: the path is provider-agnostic, routing through the base manager seams (_new_state / _create_single / _ready_checker / _apply_ready / _handle_of / _kill_one / _slot_templates) the cube manager already supplies. create_one forwards metadata + template; stamps _slot_templates so _to_instance stamps the resolved template. kill_one is finally-safe (missing state returns).
• SnapshotSizeCapable — snapshot_sizes returns None for now. The cube SDK's SnapshotInfo exposes only snapshot_id + names (no size fields), so per-pause size collection can't be satisfied from the control plane yet. None keeps the provider SnapshotSizeCapable (probed right after pause) while signalling "no data" -- the snapshot_size series event is skipped, not crashed. A follow-on can stat the cube snapshot dir on the host like aenv's scan_snapshot_sizes.

default_replay_mode flips exec_only → lifecycle (cube has native pause/resume WITH a memory snapshot). vmm_type stays None (vm_monitor deferred).

A try/except Sandbox import is added to the provider module (resume uses Sandbox.connect); the mock mirrors aenv's, keeping it importable without the SDK.

bench.py

The lifecycle/trajectory validation error messages now guide users to --provider aenv or --provider cubesandbox (cube is now LifecycleCapable + EphemeralCapable, so the aenv-only guidance was stale).

Tests

The 3 Phase 1 identity assertions (exec_only / not-Lifecycle / not-Ephemeral) are replaced with Phase 2 versions (lifecycle / IS-Lifecycle / IS-Ephemeral / IS-Snapshot) plus 11 behavioral tests:

• pause forwards wait=True + missing-handle raises
• resume swaps the handle via connect + missing-state raises
• snapshot_sizes returns None
• create_one runs the full manager seams (metadata/template forwarding, ready probe, _slot_templates stamping, sdk-failure raises)
• kill_one calls kill + marks dead + missing-state noop

All 47 cube tests + the full kernel/provider suite pass.

Verification

python -m pytest src/bench_core/tests src/env_provider/tests   # full suite green
python -m ruff check src/env_provider/cubesandbox src/bench_core/bench.py
python -m ruff format --check src/env_provider/cubesandbox
python -m pre_commit run --files <phase-2 files>              # all hooks pass

Stacking

• Base: feat(cubesandbox): exec-only provider (Phase 1, stacked on RFC #136) #137 (Phase 1 exec-only, feat/cubesandbox-provider-exec) — this PR targets that branch.
• Root: RFC 0002: Add cubesandbox provider (lifecycle-capable, Cloud Hypervisor microVM) #136 (RFC, feat/cubesandbox-provider-rfc) — base of feat(cubesandbox): exec-only provider (Phase 1, stacked on RFC #136) #137.
• Deferred: vm_monitor integration (cube-hypervisor vs cloud-hypervisor VMM name) and real snapshot-size host-stat collection.

When #137 merges, this PR retargets to the next base up.