Gives llm_db (the model catalog req_llm resolves models through) a snapshot that works on a phone.

llm_db reads its packaged priv/llm_db/snapshot.json (10 MB, 230 providers) at runtime, but deps' priv/ isn't shipped to a mob device and Application.app_dir/2 doesn't resolve there. compile_embed: true works but costs a 12.9 MB beam and a 2.4-20 s first lookup on a 1-scheduler phone BEAM. rho only talks to Anthropic, OpenAI (req_llm resolves openai_codex: models through openai's) and Google, so this module trims the packaged snapshot to those providers at compile time (its own snapshot_id, so llm_db's strict integrity check still holds), and install!/0 writes it to the data dir and points config :llm_db, :snapshot_path at it before the first lookup (Rho.Supervisor does this when it starts).