REFACTOR: De-dup scorer output_key defaults onto Scorer base as ClassVars

[Roman Lutz (romanlutz)]

REFACTOR: De-dup scorer output_key defaults onto Scorer base as ClassVars

Scope

Part of the constants-audit cleanup (see PRs #1964 and #1965 for the established pattern). PR B-2 of the staged plan.

What changed

The same five string defaults for the LLM scoring JSON-response key names were duplicated across four sibling classes:

• score_value_output_key = "score_value"
• rationale_output_key = "rationale"
• description_output_key = "description"
• metadata_output_key = "metadata"
• category_output_key = "category"

They are now declared once on the Scorer base as ClassVar[str]:

class Scorer(Identifiable, abc.ABC):
    DEFAULT_SCORE_VALUE_OUTPUT_KEY: ClassVar[str] = "score_value"
    DEFAULT_RATIONALE_OUTPUT_KEY: ClassVar[str] = "rationale"
    DEFAULT_DESCRIPTION_OUTPUT_KEY: ClassVar[str] = "description"
    DEFAULT_METADATA_OUTPUT_KEY: ClassVar[str] = "metadata"
    DEFAULT_CATEGORY_OUTPUT_KEY: ClassVar[str] = "category"

Touched files:

• pyrit/score/scorer.py — promote constants; _score_value_with_llm_async now uses the sentinel pattern (str | None = None, resolved to self.DEFAULT_*_OUTPUT_KEY).
• pyrit/score/float_scale/float_scale_scorer.py — override signature widened to str | None = None; the override forwards values through to super() which resolves the sentinel.
• pyrit/score/float_scale/self_ask_general_float_scale_scorer.py — constructor now uses the sentinel pattern; self._*_output_key is resolved from the class default when no value is provided.
• pyrit/score/true_false/self_ask_general_true_false_scorer.py — same constructor change.

Behavioural compatibility

• SelfAskGeneralTrueFalseScorer() (no kwargs) still produces self._score_value_output_key == "score_value", etc.
• SelfAskGeneralTrueFalseScorer(score_value_output_key="custom") still produces self._score_value_output_key == "custom".
• Other internal callers of _score_value_with_llm_async that omit these args (e.g. SelfAskTrueFalseScorer, SelfAskRefusalScorer, SelfAskQuestionAnswerScorer, SelfAskCategoryScorer, SelfAskScaleScorer, SelfAskLikertScorer, InsecureCodeScorer) get the same resolved defaults as before.

No value changes. No behaviour changes.

Verified

• uv run --link-mode=copy ruff check pyrit/score/ — clean.
• uv run --link-mode=copy ty check pyrit/score/ — no new diagnostics (one pre-existing cv2 optional-import warning in video_scorer.py).
• uv run --link-mode=copy pytest tests/unit/score/ -x --no-header -q — 1184 passed, 16 skipped.

[Roman Lutz (romanlutz)]

Closing per maintainer feedback on the constants-audit work.

This PR introduced the sentinel-default pattern (changing x: T = <literal> into x: T | None = None + if x is None: x = self.DEFAULT_X) which widens the public API signature. The preference is to keep constants inside their owning class (Track A — see #1964, #1965, #1972, #1973, #1974, #1975) but not to promote hard-coded single-use defaults via a None-sentinel.

Branch left intact in case any portion is worth cherry-picking later.