Indic-first text embeddings and retrieval — domain-specialist models with auditable, refuse-by-default data provenance.