Self-Hardening VectorDB Memory

[luigi-agosti]

• When any scanner (IDPI or future ones) flags something, embed the bad input (use pure-Go embeddings like github.com/knights-analytics/berty or a tiny ONNX model).
• Store in a lightweight Go vector store (bleve, pgvector, or even an in-memory map for starters).
• On new input → cosine similarity check against known attacks.
• → Makes idpishield learn from real attacks in production.