[pmrv]

Each stage method (run_averaging, run_integration, reversible_scaling, temperature_scaling, pressure_scaling, scan_temperature_range) creates its own LAMMPS session and closes it as its last statement. So if anything in the stage raises, the session stays open. In library mode that keeps the pylammpsmpi/executorlib workers alive until the process exits.

Changes

• BaseRunner gets __enter__/__exit__, which close the session on the way out, with or without an exception (calphy/runner.py).
• Phase gets the same: __exit__ closes the session opened by the stage it wraps. Stages now create their session with lmp = self._open_lammps() instead of ph.create_object(...), so the job remembers it. Stage bodies are otherwise unchanged.
• In routines.py (and the averaging call in dcci.py), each stage call becomes

  with job:
      job.run_averaging()

On success the stage has already closed the session itself, and the second close does nothing on both backends: LibraryRunner checks _closed, the executable runner's buffer is already empty, and the restart-file removal checks that the file exists.

__exit__ never suppresses the exception. If close() itself fails while an exception is already propagating, the close error is logged and the original exception is re-raised. A dead LAMMPS worker can easily make the close fail, but the error that matters is the one that got us there.

Setup failures are covered too. create_object emits the init commands after it has constructed the runner, and LibraryRunner.__init__ activates ML-IAP after the executor already exists. If either step raises, the runner closes itself (through __exit__) before the error propagates, because at that point nobody else holds a reference to it.

The DCCI driver's own two-cell sessions (dcci.py, lmps[phase]) are not touched here.

Testing

• Full suite: 444 passed, 26 skipped. This includes the golden command-stream tests, and the emitted LAMMPS command streams are unchanged.
• New tests: the runner context manager closes on success and on exception, a failing close() raises when no exception is in flight and does not mask one that is, with job: closes the session of a failed stage, and a failure in ML-IAP activation or in the init commands closes the executor. Both of those fail without the fix.
• No end-to-end run against a real LAMMPS binary.

🤖 Generated with Claude Code

[pmrv]

@srmnitc I had issues before with calphy/executorlib when lammps errors during a run, e.g. on exploding 'Domain too large for neighbor bins'. This should help.