[BUG] `ChemPropModel.predict()` creates spurious `lightning_logs/version_*` directories on every prediction call

[smcolby]

Every call to ChemPropModel.predict() silently creates a new lightning_logs/version_N/ directory in the current working directory. In a typical active learning run — where predict() is called once per committee member per AL iteration — this produces hundreds of stray directories.

In openadmet/models/architecture/chemprop.py, the ephemeral pl.Trainer constructed inside predict() is initialised with logger=None:

# openadmet/models/architecture/chemprop.py  ~line 583
trainer = pl.Trainer(
    logger=None,          # <-- bug: should be logger=False
    enable_progress_bar=False,
    accelerator=accelerator,
    devices=devices,
)

In PyTorch Lightning 2.x, logger=None is not equivalent to disabling logging. It causes Lightning to fall back to its default logger — TensorBoardLogger when TensorBoard is installed — which writes a version_N/events.out.tfevents.* directory tree rooted at {cwd}/lightning_logs/. The correct value to disable logging entirely is logger=False.

This can be confirmed by the contents of the stray directories:

lightning_logs/version_0/
├── events.out.tfevents.…
└── hparams.yaml