fix(sandbox): prevent Daytona heartbeat sleep leak

[bingran-you]

Summary

• replace signal-based heartbeat teardown with a stop-file and bounded polling
• prevent the race that leaves sleep 60 orphaned with captured pipes open
• pin the regression guard to the heartbeat path introduced by PR Fix Toolathlon Daytona blocker retries #894

Root cause

The fast-command path could send TERM after the heartbeat shell spawned sleep 60 but before it stored the child PID. The trap then exited without killing the sleep process. That orphan retained stdout/stderr, so subprocess.run(..., timeout=2) timed out even though the parent shell had already returned 0.

Validation

• uv run pytest -q tests/test_daytona_dind_compose_up.py — 12 passed
• 100 fast-command iterations on macOS — passed
• 100 fast-command iterations in Debian/Linux — passed
• forced 2-second heartbeat emission — passed
• uv run ruff format --check src tests tools
• uv run ruff check src tests tools
• uv run ty check
• uv lock --check
• uv run python -m pytest tests/ -q — 4835 passed, 8 skipped, 7 deselected
• pip-audit — no known vulnerabilities, one documented ignore

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a560034f73

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Stop waiting a full poll tick after the child exits

On Daytona DinD execs with timeout_sec >= 600, this sleep 1 keeps the wrapper process alive until the next poll even after wait "$bf_pid" has observed the real command exit. _compose_exec passes the same timeout_sec down to _sandbox_exec, which wraps the whole compose command in GNU timeout {timeout_sec}; timeout --help says it kills a command that is still running after the duration and exits 124. As a result, a command that succeeds near the deadline, e.g. at 599.5s for a 600s verifier/agent timeout, can be killed before cat "$bf_out" and exit "$bf_rc", turning a successful run into a timeout and dropping its output.

Useful? React with 👍 / 👎.

[bingran-you]

@kywch @Yiminnn @xdotli — this is a narrow release-blocking process-cleanup fix. All required CI and the real rollout smoke are green; a human approval is the remaining merge gate.