feat(sandbox): keep mount endpoints stable across sandbox restarts

Problem

The native sandbox_agent mount currently receives a concrete Sandbox Agent baseUrl when the VM is created. That endpoint is not stable across sandbox replacement: Docker allocates a new host port, and hosted providers may return a different URL after recreation or resume. A lazy sandbox that shuts down while the VM remains alive therefore cannot safely reuse the original mount descriptor.

Desired behavior

• Give each AgentOS VM a stable sandbox endpoint for the lifetime of the VM.
• Route that endpoint through a host relay, mutable resolver, or equivalent indirection to the currently active Sandbox Agent client.
• Let both the native filesystem mount and process bindings share the same current client and lazy lifecycle.
• Update authentication headers/tokens when the backing sandbox changes without remounting or rebooting the VM.
• Fail clearly while startup is in progress or has failed; concurrent calls must remain single-flight.
• Add native coverage that accesses the same mounted path before and after replacing the backing sandbox without restarting the AgentOS VM.

Provider sandbox IDs may remain stable for reconnect-capable backends, but their network address must not be treated as a cross-provider invariant.

Start by tracing the native sandbox_agent mount, AgentOS VM lifecycle, and process bindings described in the issue; no file or test path is provided. Add native coverage that uses the same mounted path before and after replacing the backing sandbox without restarting the VM, and verify stable routing, updated authentication, clear startup failures, and single-flight concurrent calls.