fix(server): bound graceful shutdown so SIGINT/SIGTERM cannot hang on open streams

[wtfsayo]

Summary

Fixes #6 - Ctrl+C did not shut down the server when long-lived streams were open.

• axum::serve(...).with_graceful_shutdown(...) waits for in-flight requests to finish. SSE streams (/v1/processes/{id}/logs?follow=true, /v1/acp/{id}, opencode compat event feeds) never complete on their own, so the drain blocked forever and the process needed kill -9.
• Also handle SIGTERM: sandbox-agent daemon stop already sends it, and it previously fell through to the default action - an instant terminate that skipped shutdown_servers cleanup (opencode sidecar, ACP runtimes) entirely.
• Bound the drain: after the first signal and cleanup, a spawned watchdog exits the process after a 3s grace window, or immediately on a second SIGINT/SIGTERM.

Verification (live, darwin/arm64)

Reproduced on unpatched build: server + sleep 300 process + open logs?follow=true SSE → SIGINT → still running at 10s; exits only once the SSE connection closed.

Patched build:

• SIGINT, no streams: exits immediately
• SIGINT with open SSE stream: drains, force-exits inside the 3s bound
• SIGTERM: graceful cleanup then exit (previously skipped cleanup)
• second signal during drain: immediate exit

cargo test -p sandbox-agent --lib: 22 passed. cargo fmt --check clean, cargo clippy clean (one pre-existing warning unrelated to this change).

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.