fix: apply similarity threshold on raw cosine before normalization

[yugalkaushik]

• The similarity threshold is now checked on the raw cosine score before normalization, so normalization can no longer stretch a weak pair to 1.0 and let it pass. Floor, normalize, and tag/link/temporal bonuses are kept, but they only rank pairs that already passed.
• Replaces absolute similarity threshold with a relative cutoff
• Keeps SEMANTIC_FLOOR (0.3) as absolute guard for tiny vaults

[personalizedrefrigerator]

Note: According to its model card, multilingual-e5-small (the model Joplin uses for word embedding) distributes cosine similarity scores in the range [0.7, 1] (related Joplin issue).

[malekhavasi]

Good point, thanks for flagging this. The absolute threshold doesn’t really separate related and unrelated pairs for this model. We should use a relative cutoff, with a small absolute floor for tiny vaults