"""Narrator — the ONLY component that calls a model. Built in Phase 3.

Translates the engine's deterministic output into plain English via a LOCAL model
(llama.cpp llama-server on port 12781). Suggests, never asserts (Non-negotiable #7):
proven value-flow -> causal language; proximity -> sequence language only; clean turn
-> say it's clean. Proven numbers come from the engine and MUST NOT move when the
model is swapped — only the prose changes.
"""