[nicoloboschi]

Summary

• Replace rank-based normalization with passthrough for reranker scores already in [0, 1]. Calibrated rerankers (Cohere, Jina, llama.cpp/Qwen) return meaningful absolute confidence — rank normalization was inflating weak candidates (e.g. 0.007 → 1.0) simply for being top-ranked among bad results.
• Sigmoid normalization for logit-based local models is unchanged.
• Downstream multiplicative boosts (recency, temporal, proof count) are scale-invariant and require no adjustment.

Closes #1823

Test plan

• Updated test_reranker_score_normalization.py — passthrough tests replace rank normalization tests
• New test_low_confidence_scores_preserved validates the core issue scenario (all-weak candidates stay low)
• test_combined_scoring.py passes unchanged (boosts are scale-invariant)
• Verify with a calibrated external reranker (Cohere/llama.cpp) that recall scores reflect actual confidence