strategy 2 compute a overall confidence score

investigation: #578606 (comment 2870291617)

Compute a confidence bucket based on score distribution:

• high: strong top candidate + steep drop-off
• medium: reasonable cluster / partial relevance
• low: flat distribution / ambiguous meaning

This helps the LLM decide:

• When to answer directly (high)
• When to hedge / propose alternatives (medium)
• When to ask clarifying questions (low)