[CLI] predict(ckpt_path="best") raises ValueError even though ModelCheckpoint is configured under trainer.callbacks

[oooooio]

Bug description

When running trainer.predict via LightningCLI with --ckpt_path best, Lightning raises:

ValueError: `.predict(ckpt_path="best")` is set but `ModelCheckpoint` is not configured to save the best model.

However, ModelCheckpoint is configured under trainer.callbacks in the YAML, and training is executed via LightningCLI. This looks like either:
1. LightningCLI isn’t detecting the ModelCheckpoint from the CLI YAML when resolving best during predict, or
2. The checkpoint_connector’s detection logic for best doesn’t handle the CLI/YAML path or log directory resolution, or
3. A mismatch between monitor metric registration and best-checkpoint discovery is not surfaced clearly (i.e., the callback exists, but the condition for “best” was silently unmet).

What version are you seeing the problem on?

v2.5

Reproduced in studio

No response

How to reproduce the bug

trainer:
  callbacks:
    - class_path: lightning.pytorch.callbacks.ModelCheckpoint
      init_args:
        monitor: val_loss
        mode: min
        save_top_k: 1
        filename: "{val_loss:.4f}"
        save_last: false



uv run src/cli.py predict -c config.yaml --ckpt_path=best

Error messages and logs

# Error messages and logs here please

Environment

Current environment

#- PyTorch Lightning Version (e.g., 2.5.0):
#- PyTorch Version (e.g., 2.5):
#- Python version (e.g., 3.12):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):

More info

No response

cc @lantiga @mauvilsa

[oooooio]

I just found a very similar issue here: #17312.
It looks like the same root cause — using --ckpt_path best