Replace systemd and start-up.sh with process-compose

[mishushakov]

Jupyter and the code-interpreter server are now supervised by process-compose in both the E2B sandbox and the Docker image, so the two paths can no longer diverge; restart policies, ordering and the ExecStartPre health check gate map over directly, while PartOf=jupyter.service becomes a readiness probe comparing Jupyter's reported start time against the one recorded at gate time (a sticky signal, unlike a liveness check which loses the race — Jupyter recovers in ~1s). Jupyter writes its log file directly rather than through process-compose, because kernels inherit its stdout/stderr and outlive it, and process-compose waits for those descriptors to close before acting on the exit — capturing them left a killed Jupyter unrestarted while any kernel ran, silently breaking the OOM recovery the systemd setup existed for. The restart tests never exercised any of this: pgrep -f 'jupyter server' matched the shell running it, so the kill only ever hit itself and the catch swallowed the error; they now use a bracketed pkill that fails on no match and judge recovery by running code, since the old server answers /health for several seconds while holding dead kernel handles. The debug=True systemd drop-in is gone (logs are always captured now) and debug_logs.py/README point at the new log files.

Verified against a clean Docker build: boot to working execution in 3s, 20s steady state with zero restarts or probe failures, kill Jupyter with kernels alive → both restart and execution works in 7s, kill code-interpreter → only it restarts in 3s. One trade-off: /var/log/jupyter.log is no longer rotated, since process-compose can't capture it; Jupyter can rotate its own log via Application.logging_config, but that sends startup stderr to /dev/null, losing exactly what make debug-template exists to show.

🤖 Generated with Claude Code

[cursor]

PR Summary

High Risk
Changes how every sandbox boots, supervises, and recovers from Jupyter/OOM kills—core infrastructure with subtle coupling between processes; misconfiguration could break production templates or leave sandboxes unable to execute code.

Overview
Replaces separate systemd (sandbox) and start-up.sh (Docker) supervision with a single process-compose.yaml that boots both environments the same way via process-compose up in template.py.

Jupyter logs go directly to /var/log/jupyter.log so kernels inheriting stdout/stderr do not block restarts. Code-interpreter still gates on jupyter-healthcheck.sh (now also recording Jupyter’s .started in /run/jupyter-instance) and is recycled when Jupyter is replaced via the new jupyter-instance-check.sh readiness probe—replacing systemd’s PartOf=jupyter.service.

Removes systemd units, start-up.sh, and the debug=True journal drop-in; debug_logs.py and the README now point at process-compose and file logs. Restart tests move from systemd.test.ts / test_*_systemd.py to supervision suites using bracketed pkill and waitForWorkingSandbox (runCode) instead of /health alone.

^{Reviewed by Cursor Bugbot for commit 469b439. Bugbot is set up for automated code reviews on this repo. Configure here.}