Critical: _find_checkpoint() glob patterns never match any trained checkpoint

[bradsmithmba]

Summary

EnhancedStrategyRecommender._find_checkpoint() searches for files matching checkpoints/regime_detector*.pt and checkpoints/regime_detector*.pth. No file produced by the training pipeline matches either pattern, so the ML path never activates even when a trained model exists.

Root cause

_find_checkpoint() (added in PR #42):

patterns = ['checkpoints/regime_detector*.pt', 'checkpoints/regime_detector*.pth']
matches = [f for p in patterns for f in glob.glob(p)]

Two problems:

Wrong directory. TrainingConfig.checkpoint_dir defaults to Path("models/checkpoints"), not checkpoints/. The training script overrides this to training_runs/<timestamp>/checkpoints/ when --output-dir is not set.

Wrong filename pattern. RegimeTrainer.save_checkpoint() writes best_model.pth and checkpoint_epoch_N.pth. The export_model_for_inference() path writes inference_model.pth. None of these contain regime_detector in the filename.

Impact

_find_checkpoint() always returns None. _detect_regime() always falls back to the SMA-crossover heuristic, silently. The entire ML pipeline installed in PR #42 is dead code until this is fixed.

Fix

_find_checkpoint() needs to:

1. Search the actual checkpoint directories the trainer writes to (models/checkpoints/, any training_runs/*/checkpoints/)
2. Match the actual filenames the trainer produces (best_model.pth, checkpoint_epoch_*.pth, inference_model.pth)

Example:

def _find_checkpoint(self) -> Optional[str]:
    patterns = [
        'models/checkpoints/best_model.pth',
        'models/checkpoints/checkpoint_epoch_*.pth',
        'training_runs/*/checkpoints/best_model.pth',
        'training_runs/*/checkpoints/checkpoint_epoch_*.pth',
        'training_runs/*/inference_model.pth',
    ]
    matches = [f for p in patterns for f in glob.glob(p)]
    return max(matches, key=os.path.getmtime) if matches else None

Related

Companion to issue: _detect_regime() calls load_state_dict() on a full checkpoint dict (should unpack model_state_dict key).