FAISS: apply score_threshold once on normalized relevance; clamp to [0,1]

[lokyaan]

Problem
score_threshold could be applied on raw FAISS scores (pre-normalization) and/or in multiple layers, leading to inconsistent behavior across distance strategies.

Change

• Remove raw-score thresholding in similarity_search_with_score_by_vector.
• Filter once in _similarity_search_with_relevance_scores (and async) after mapping to normalized relevance.
• Clamp normalized relevance to [0,1].

Why
Predictable threshold semantics across MAX_INNER_PRODUCT / EUCLIDEAN_DISTANCE / COSINE and with/without L2 normalization.

Testing

• Added a minimal integration test locally (FAISS inner-product case) confirming only items ≥ threshold are returned and all scores are in [0,1].

Notes
Happy to relocate/add tests to this repo’s preferred test dir if maintainers want.