feat: fuzzy/embedding-based prefix similarity via FAISS

[manishklach]

Research Question

What changes if approximate prefix reuse moves beyond structural token similarity and experiments with embedding-based neighborhood lookup?

Scope

• keep the current MinHash path as the structural baseline
• prototype an optional FAISS-backed similarity experiment
• compare match quality, reuse rate, and false-positive risk
• clearly separate this from the current metadata-aware exact/structural matching model

Note

This should remain an exploratory simulation feature, not a claim of production retrieval quality.