BUG: Crescendo passes the sent prompt to the refusal scorer as the attack objective #3561

Description

CrescendoAttack._perform_backtrack_if_refused_async passes prompt_sent, the full adversarial prompt that was just sent to the target, as the objective argument of _check_refusal_async. That parameter is documented as the attack objective and is rendered into the refusal judge's prompt as the conversation objective (conversation_objective: {objective} in SelfAskRefusalScorer).

With the default SelfAskRefusalScorer in an OBJECTIVE mode, the judge therefore evaluates deflection and redirection against the adversarial prompt itself instead of against the attack objective. The refusal score's persisted expectation also records the prompt as the objective, so expectation fingerprints and replay identity carry the wrong reference text.

Expected Behavior

The refusal check should receive context.objective, matching the parameter's contract and the other scoring calls in the same class, which all pass context.objective.

Reproduction Steps

1. Run a Crescendo attack with a SelfAskRefusalScorer in OBJECTIVE_STRICT mode.
2. Inspect the refusal judge's rendered request or the persisted scored_expectation.
3. The conversation objective shown to the judge is the prompt that was sent, not the attack objective.

Minimal Reproduction

pyrit/executor/attack/multi_turn/crescendo.py line 817: self._check_refusal_async(context, prompt_sent) where the method signature names the second parameter objective.

Environment

PyRIT main. Python 3.11.