[dylanuys]

Summary

Preserve prediction counts for every supported media class in the bookkeeping summaries emitted after a benchmark run. Video predictions labeled semisynthetic or rendered were absent from per_source_accuracy, which made downstream reporting show partial totals and misleading 0/0 rows.

Scope

This is a reporting/bookkeeping-only change. It does not change:

• model inference or preprocessing
• dataset discovery, sampling, or evaluation coverage
• ground-truth labels or predicted labels
• MCC, Brier, SN34, benchmark score, ranking, or pass/fail outcomes
• benchmark execution behavior or runtime characteristics

Those metrics are calculated from the complete result rows before per_source_accuracy is assembled. This patch only stops the final summary payload from discarding prediction-count buckets for labels 2 and 3.

Changes

• include semisynthetic and rendered counts in per-dataset prediction distributions
• carry all four counts into per_source_accuracy
• add a regression test covering the complete per-dataset-to-per-source reporting path

Existing runs

Existing scores and outcomes do not require correction. Their lossy per_source_accuracy bookkeeping can be rebuilt from stored parquet records without rerunning inference.

Validation

• pytest -q tests/unit/test_multiclass_metrics.py — 18 passed
• Ruff checks passed