Fix realistic answer parsing for text answers

[Artemis-ignis]

Summary

• Fix criteriaoutput in the realistic evaluator so \text{answer: } parsing reads the current reply instead of always indexing generatedtext[0].
• Add a bounds check while scanning answer digits so answers at the end of a reply do not raise IndexError.
• Add focused regression coverage for multi-reply parsing and end-of-reply answers.

Verification

• python -m unittest discover -s tests -v
• python -m compileall -q gsm-infinite tests