VM dispose stalls after guest network use

Disposing a VM after guest code used the network is slow, and very slow when a response body was left unread.

Repro

const vm = await AgentOs.create({ permissions: { network: "allow" } });
await vm.javascript.evaluate(`(await fetch("https://example.com")).status`);
await vm.dispose(); // ~100 s

Actual

• Dispose logs ERR_AGENTOS_VM_TEARDOWN_DEADLINE: … active_tasks=1 … deadline_ms=5000 and then takes about 100 s to return.
• Reading the body first (await response.text()) brings the whole call back to ~400 ms, but any guest network use still adds roughly the 5 s teardown deadline.

Expected
Dispose closes guest sockets and returns within the configured shutdown deadline. Untrusted code should not be able to hold a VM open by leaving a connection unread.

Start by reproducing the AgentOs.create, vm.javascript.evaluate, and vm.dispose sequence in the issue, comparing unread and fully read response bodies. Trace the VM disposal and guest-network teardown entry points to identify why active_tasks remains; done means guest sockets close and dispose returns within the configured shutdown deadline, including for unread responses.