compute_metrics keeps non-finite scores that calculate_metrics filters, and the CV objective can crash on an empty result

[Maarmapa]

BaseModel.calculate_metrics skips metrics that return a non-finite value, with a warning that names the usual cause:

# DashAI/back/models/base_model.py:305-312
score = metric.score(y_transformed, y_pred)
if not math.isfinite(score):
    logger.warning("Metric %s returned a non-finite value (%s) for split %s "
                   "(e.g. only one class present in the split). Skipping.", ...)
    continue

compute_metrics (line 339) copies that scoring loop without the guard:

# DashAI/back/models/base_model.py:392-393
score = metric.score(y_transformed, y_pred)
results[metric.__name__] = score

It also keeps the same early returns as calculate_metrics — no metrics configured, or no data for the split — returning {} at lines 372, 377 and 383.

CrossValidationEvaluationStrategy.evaluate is the caller, once per fold:

# DashAI/back/evaluation/cv.py:203-206
train_scores = model.compute_metrics(split=SplitEnum.TRAIN)
validation_scores = model.compute_metrics(split=SplitEnum.VALIDATION)
folds_results.append(validation_scores[metric.__name__])

Two consequences:

1. A NaN from one fold poisons the whole trial. folds_results feeds np.mean(folds_results) at cv.py:244, which is the objective value the optimizer sees. One fold containing a single class makes the trial's score NaN, indistinguishable from a genuinely bad trial. calculate_metrics was written to avoid exactly this.

2. An empty dict raises KeyError. If compute_metrics takes any of its three early returns, validation_scores[metric.__name__] fails and the fold loop dies mid-trial, with a traceback pointing at a subscript rather than at the cause.

Suggested fix: share the scoring loop between the two methods, or add the same isfinite skip to compute_metrics. Either way it is worth deciding explicitly what a fold without the goal metric should do — skip that fold, or fail with a message that names the reason — rather than letting a KeyError decide.

Checked against develop @ b3b7296. Happy to send a PR if you tell me which of the two behaviours you want for the missing-metric case.