LLMJudgeScorer.__init__ wipes the class-level category when constructed directly

[knisar]

Surfaced while reviewing #27; the author flagged it there.

BaseScorer.__init__ uses None as the sentinel for "keep the class attribute":

if category is not None:
    self.category = category

LLMJudgeScorer.__init__ defaults category to "" and passes it through, so the sentinel never fires and every judge subclass loses its declared category when constructed directly:

>>> FactualityJudge(api_key="x").category
''
>>> GroundednessScorer(api_key="x").category
''

All ten judge scorers are affected. It doesn't bite via the registry because resolve_scorers passes category explicitly, but anything constructed by hand (additional_scorers, notebooks, tests) ends up grouped under the scorer name in reports, and category-based policy expectations don't match.

Fix is a one-liner in rai_toolkit/scorers/llm_judges.py: change the default to category: str | None = None. Add a test asserting FactualityJudge(api_key="x").category == "MIT-3.1" and one for a subclass constructed with an explicit override.

[YusefSyed]

I’d like to work on this. I’ll preserve each judge subclass’s class-level category when no override is supplied, add focused coverage for both the default and an explicit override, and disclose AI assistance in the PR as requested by the contribution guide.