fix: parse embedded ranking JSON

[alibababye]

Problem

RankingService._parse_llm_response could parse fenced JSON, but the fallback for prose-wrapped responses only matched objects without nested braces. Ranking responses are expected to contain nested criterion objects such as novelty.score, so a response like Here is the result: {...} could parse the first inner object instead of the full payload and leave the criteria at default values.

Solution

This updates the fallback JSON extraction to capture the full object and adds a regression test for nested ranking JSON embedded in explanatory text.

Testing

Ran in the sandbox:

python3 tests/test_ranking_variance.py