Re-indexed 4.2M vector embeddings using the new Titan-v2 model.
Experimental noise reduction filters applied to historical logs.
Stable release with legacy tokenizer support.