fix(metric): track none_count on Stat to surface denominator-drop (unscoreable-as-pass)

[AUTHENSOR]

Summary

Stat.add(None) silently skipped values without counting them, so the reported mean excluded unscoreable samples from the denominator. A model-under-test that makes the judge unscoreable on its hardest samples (e.g. output that causes a judge parse failure or refusal) can erase those samples from its reported score — the "unscoreable-as-pass" class.

This is the same metric-integrity issue fixed in garak (NVIDIA/garak#1954) and ragas (vibrantlabsai/ragas#2832), now in HELM's Stat aggregator.

The drop

src/helm/benchmark/metrics/statistic.py:33-36:

def add(self, x) -> "Stat":
    if x is None:
        return self    # <-- count NOT incremented, sample vanishes from denominator

Several critique metrics return None on judge parse failure (e.g. gpt4_audio_critique_metrics.py:64-73 _extract_score_from_gpt_output → None on regex miss → stats[answer_name].add(None)), which then silently drops from the mean.

Fix

Add a none_count field that tracks skipped values, and surface it in bare_str():

• min=0.9, mean=0.9, max=0.9, sum=1.8 (2 ⚠3 skipped) when none_count > 0
• Unchanged output when none_count == 0

This makes the denominator-drop visible in metric output without changing aggregation semantics. Callers can check stat.none_count to detect the gap between the reported mean and the true worst-case.

Note

HELM's safety_metrics.py:76-77 already defends against the all-None case (raise SafetyScoreMetricException), and wildbench_metrics.py:27-28 raises on all-failed. This fix covers the partial-drop case (some but not all samples are None) that those guards miss.

Checklist

• Code follows project style (ruff line-length 120)
• ruff check + ruff format pass
• Additive — no change to existing aggregation semantics
• Self-reviewed