Fix KeyError on category-5 (adversarial) QA: read ground truth from adversarial_answer

[kamaalg]

Problem

A clean checkout cannot evaluate any model on the category-5 (adversarial) questions. data/locomo10.json stores cat-5 ground truth under the key adversarial_answer, but the code reads answer. 444 of 446 cat-5 questions have no answer key, so evaluation crashes with KeyError: 'answer'. Category 5 is the 2nd-largest category, so the adversarial half of the benchmark is unrunnable as shipped.

Two crash sites, both reproduced on live HEAD (3eb6f2c) with no API key:

• Scoring: task_eval/evaluation.py:200,202 — answer = str(line['answer']) runs for every QA line before the cat-5 branch → KeyError.
• Generation: the cat-5 distractor blocks in gpt_utils.py, claude_utils.py, gemini_utils.py, hf_llm_utils.py build the multiple-choice distractor from qa['answer'] → KeyError.

README.MD documents only an answer key, confirming the data/code/doc mismatch. The adversarial_answer is exactly the "tempting wrong" option the (a) Not mentioned / (b) <distractor> MC is designed to present.

Fix

Read cat-5 ground truth defensively: gold = line.get('answer', line.get('adversarial_answer')) in the scorer, and adv_answer = qa.get('adversarial_answer', qa.get('answer')) for the distractor in the four model utils. Zero behavioural change for categories 1–4. (+34 / −18 across 5 files.)

Test

Adds task_eval/test_cat5_eval.py (no network, no keys). On unmodified evaluation.py → KeyError: 'answer' (fails); with the fix → passes. Also verified: the fixed scorer runs all 1986 QA (incl. all 446 cat-5) and the fixed generation builds all 446 cat-5 prompts with 0 crashes.