BUG: Inferred objective is empty when tool calls come between the request and the answer #3515

Describe the bug

extract_objective_from_previous_turn_async (used when a scorer infers the objective from the request) only reads the turn at sequence - 1. When a target runs a tool loop, memory stores the user request, the function_call, the function_call_output and then the answer, so the turn before the answer is the tool output. It has no user text, and the scorer ends up with an empty objective.

#3410 fixed the same problem in MemoryInterface.get_request_from_response_async, and WildGuardScorer already looks back to the latest earlier user turn.

Steps to Reproduce

Store a conversation of user question (seq 2), assistant function_call (seq 3), tool function_call_output (seq 4) and assistant answer (seq 5), then call extract_objective_from_previous_turn_async on the answer.

Expected Results

The objective is the user question from seq 2.

Actual Results

An empty string.

Versions

Linux, Python 3.13.16, PyRIT 1.2.0.dev0 from main (96de757)