FIX: threshold scorer attributes result to first score, not the aggregate

[WatchTree-19]

Description

FloatScaleThresholdScorer computes its true/false decision from the aggregate of the wrapped scorer's outputs, but when that scorer returns more than one score it reuses scores[0] verbatim for everything except the value. The category, rationale and metadata therefore describe the first constituent score rather than the one that actually crossed the threshold.

AzureContentFilterScorer returns one float_scale score per harm category, ordered by category name, so this is reachable in normal use.

Reproduction

Wrapped scorer returns Hate 0.0, SelfHarm 0.0, Sexual 0.0, Violence 0.857. Threshold 0.5, default MAX aggregator. On main:

score_value       : True
score_category    : ['Hate']
score_metadata    : {'azure_severity': 0, 'original_float_value': 0.857}
rationale         : ... Rationale for scale score: Hate rationale, severity 0

A content-safety flag that fired on Violence, reported as Hate at severity 0, with an original_float_value from a different category sitting next to it. The value is correct and everything explaining it is wrong, so the score reads as a false positive on a category that scored zero. Anyone filtering or reporting by score_category gets the wrong attribution silently.

Fix

The else branch of the same method, taken when every piece is filtered out, already builds the score from aggregate_score.category and aggregate_score.metadata. This makes the populated branch consistent with it, so category, metadata and rationale all come from the aggregate.

For a single wrapped score the aggregator returns that score's own category, metadata and rationale, so that path is unchanged.

Tests

• test_float_scale_threshold_scorer_attributes_result_to_aggregate_not_first_score fails on main with assert 'Violence' in (['Hate']) and passes with the fix.
• test_float_scale_threshold_scorer_single_score_attribution_unchanged pins the single-score path so this doesn't quietly change common usage.

The existing test_float_scale_threshold_scorer_returns_single_score_with_multi_category_scorer covers this exact code path but only asserts the value and the number of scores returned, which is why it went unnoticed. Its own fixture data (Hate 0.2, Violence 0.0, Sexual 0.8) would have been mis-attributed to Hate.

All 1628 tests in tests/unit/score pass. ruff format --check and ruff check clean.

Note for reviewers

combine_metadata_and_categories merges constituent metadata with dict.update, so where several scores share a metadata key (azure_severity) the last one wins by iteration order rather than by which score drove the aggregate. That is pre-existing aggregator behaviour and I have not changed it here. If you would prefer the threshold score to attribute to the argmax constituent specifically, rather than to the aggregate as a whole, that is a slightly larger design change and I am happy to do it that way instead.