fix(sandbox): report stopped when the agent container terminates so executions don't reclaim-loop

[cjustice]

Problem

When a sandbox's agent container is OOMKilled (or otherwise terminates) mid-turn, the execution gets stuck in an endless reclaim loop until the hard deadline and the user never gets a response.

Root cause

KubernetesExecutorBackend.status_by_id derives session status from the pod phase alone. The phase stays Running as long as any container runs — and each sandbox pod has a per-sandbox tool-server sidecar that survives the agent container's death. So after the sandbox container is OOMKilled (exit 137), status_by_id still returns running.

The execution stream consumer (runtime_control) then sees the dead container's stdout stream EOF, checks backend.status(...), gets running, and sets the execution back to retry_wait expecting a reattach to recover. But every reattach to the dead container's stdout EOFs immediately, so it loops — execute_claimed roughly every 60s — until the hard-deadline reaper terminalizes it. No failure is ever surfaced to the user.

Observed in a deployed GKE cluster: a memory-heavy build OOM-killed the sandbox container; the attach stream closed ({"status":"Success"}) at the exact instant of the OOM kill; podPhase=Running (sidecar still alive) → reclaim loop for the entire deadline window.

Fix

status_by_id now inspects the agent (sandbox) container's own state. If that container has terminated, it returns stopped even when the pod phase is Running, so the stream-EOF path finalizes the execution (failed) instead of reclaim-looping on a dead container's stdout.

Testing

• Added test_status_by_id_reports_stopped_when_agent_container_terminated — OOMKilled agent container + Running pod phase → stopped; live agent container → running. Passing (uv run pytest tests/test_sandbox_kubernetes_backend.py); ruff check clean.
• Root cause reproduced live (a real monorepo build OOM-killed the sandbox container at its memory limit); this change finalizes such executions cleanly instead of looping.

[Zygimantass]

We've merged the Rust rewrite in #344, meaning the old API and Slackbot services are deprecated. As such we're closing this PR - if you think this is a mistake, please reopen the PR and ping me.