v3.2: fp32 matrix at load (was recasting 186MB/query); add scope to split personal vs public layer

README: document real API routes

OmniBrain v3.1: real /search /stats /generate_context over exported brain

initial deploy

initial commit