FIX: keep the response handler's metadata in SelfAskLikertScorer

[Chen Yufeiyang (feiiiiii5)]

Description

SelfAskLikertScorer drops the metadata that the response handler parsed out of the judge's reply.

UnvalidatedScore.to_score installs the parsed metadata on the Score (pyrit/models/score/score.py:406), and then _convert_score replaces that dict:

# pyrit/score/float_scale/self_ask_likert_scorer.py:472
score = unvalidated.to_score(...)                     # installs score_metadata
score.score_metadata = {"likert_value": ...}           # throws it away

So a judge that reports its own metadata loses it, and only likert_value survives on the persisted score:

judge reply: {"score_value": "3", ..., "metadata": {"verdict_confidence": 0.9, "raw_judge_output": "level 3"}}

SelfAskLikertScorer     -> 0.75  {'likert_value': 3}
SelfAskScaleScorer      -> 0.75  {'verdict_confidence': 0.9, 'raw_judge_output': 'level 3'}

Both rows are the same payload through two scorers in the same family. The other two siblings, SelfAskGeneralFloatScaleScorer and InsecureCodeScorer, also keep it.

response_handler is a documented argument of this scorer, so a caller-supplied handler that reports metadata could never have it survive. The fix extends the dict instead of replacing it, which is the merge idiom already used in true_false/regex/agent_threat_rules_scorer.py:207 and true_false/wildguard_scorer.py:360. The loss also propagates: FloatScaleThresholdScorer forwards aggregate_score.metadata into the score it builds, so the keys are missing there too.

Tests and Documentation

Test: pytest tests/unit/score/test_self_ask_likert.py -q — the new case fails on ea9d0b43 with assert {'likert_value': 3} == {'verdict_confidence': 0.9, 'raw_judge_output': 'level 3', 'likert_value': 3} and passes on this branch, where the file reads 44 passed. The two existing metadata assertions at :102 and :129 use replies with no metadata key, so they pass either way — with nothing to merge, the result is still exactly {"likert_value": N}. pytest tests/unit/score tests/unit/scenario -q gives 4199 passed, 43 skipped. ruff check and ruff format --check are clean on both changed files. Not run: ty (its pre-commit hook needs the project's --extra all venv), and JupyText (no notebooks or documentation pages changed).

Details

unvalidated.score_metadata is typed dict | None, so the merge guards for None the same way the sibling call sites do (**(x.score_metadata or {})).

This also applies to the judgment-replay path, which routes through the same _convert_score.