Integration surface for checkpoint/resume across hosts

## Problem

MXC provides a state-aware sandbox lifecycle:

`provision -> start -> exec -> stop -> deprovision`

I am integrating MXC with an external distributed execution/state system and am testing a failure case that crosses the current sandbox lifecycle boundary:

1. Workload starts in an MXC sandbox on host A.
2. Host A/process disappears unexpectedly.
3. External durable state survives the failure.
4. A new MXC sandbox is provisioned on host B.
5. The workload resumes from its last verified checkpoint rather than restarting from zero.

The execution/state layer is intentionally external to MXC. I am not proposing that MXC become a distributed scheduler or persistence system.

## Integration question

What is the intended/stable MXC integration boundary for associating external checkpoint/resume state with a sandbox execution across different hosts?

Specifically, it would help to know whether maintainers expect integrations to identify executions through:

- sandbox/session identity
- lifecycle metadata
- caller-managed execution IDs
- SDK hooks/events
- another abstraction

## Why I am testing this

The goal is a reproducible failure test:

`host A dies -> durable state remains -> MXC starts on host B -> execution continues`

I can build the integration against the current TypeScript SDK and Linux backend and contribute a small reproducible example/adapter if this fits the project's direction.

I am primarily looking for guidance on the correct architectural boundary before building against an internal detail that may change during the preview.

Start with the current TypeScript SDK and Linux backend, tracing the provision, start, exec, stop, and deprovision lifecycle. Compare the available sandbox/session identity and lifecycle metadata with the proposed host-A failure, host-B resume scenario. Done means documenting or validating a stable integration boundary, ideally with the reproducible example or adapter described in the issue.

Written by the indexing model from the issue text.