Allow a retrain bank to be built without precomputed scores

Building a bank currently requires finished attribution scores, because the validate path calls `load_scores_loss_signed` unconditionally and writes a `score_sum` column for every row of `validation.csv` — but when `exclude_zero_scores` is false, scores are only used for the document count and that reporting column, both of which are avoidable. Making scores optional there would let a filter's random controls be trained during the scoring pass instead of after it, which is about 10 hours on a 1M-document row.

Start by tracing the validate path that unconditionally calls load_scores_loss_signed and writes score_sum to validation.csv. Check how exclude_zero_scores controls score use, then make the no-exclusion path work without precomputed scores while preserving score-dependent behavior when exclusion is enabled; done means a retrain bank can be built during the scoring pass without requiring finished attribution scores.

Written by the indexing model from the issue text.