[Jaafer Rahmani (bIackr0se)]

Description

memory.get_scores_async(score_category="refusal") (and the deprecated get_scores) never returns a score. Score.score_category is a list, and ScoreEntry stores it in a JSON column, but _execute_get_scores compared the whole column with the argument (ScoreEntry.score_category == score_category). A stored ["refusal"] is never equal to "refusal", so every category query by name comes back empty (reproduced on SQLite; Azure SQL builds the same whole-column comparison). Only an argument equal to the whole stored value matched, such as the complete list in its stored order.

This PR builds the condition with _add_list_conditions(..., exact=True), the helper behind the seed list filters (harm_categories, authors, groups), in its whole-element mode. A query now returns the scores whose category list contains the requested name as a whole element, compared case-insensitively as that mode does: "hate" matches ["hate"] and ["violence", "hate"], but not ["hate_speech"]. A list argument keeps working and returns the scores that list every given category, in any order. The parameter type is now str | Sequence[str] | None, and the docstrings say how it matches.

Real run: two SubStringScorers score two stored assistant replies in IN_MEMORY memory, then the scores are read back:

                                                             main                          this PR
get_scores_async(score_category="refusal")                   []                            [['refusal']]
get_scores_async(score_category="privacy")                   []                            [['credentials', 'privacy']]
get_scores_async(score_category="unused")                    []                            []
get_scores_async(score_category=["privacy", "credentials"])  []                            [['credentials', 'privacy']]
get_scores_async(score_category=["credentials", "privacy"])  [['credentials', 'privacy']]  [['credentials', 'privacy']]

Tests and Documentation

Added test_get_scores_by_score_category to tests/unit/memory/memory_interface/test_interface_scores.py. It stores scores with ["hate"], ["violence", "hate"], ["hate_speech"], ["HaTe"], None and [] and checks whole-element matching, case folding, "speech" matching nothing, stored empty and None categories not matching the queried category names, a list argument, and the intersection with score_type. It fails on main (the first query returns []).

pytest tests/unit/memory: 1923 passed, 1 skipped. ruff check, ruff format --check (0.17.1) and ty check on the changed files pass.