Autonomous vehicles now operate in dense, adversarial, unpredictable environments where sensor noise, inconsistent timing, nondeterministic AI models, and multi-agent conflict can produce catastrophic outcomes. Despite advances in perception and control, the software governing autonomous vehicles remains nondeterministic, non-auditable, and non-reproducible. This mismatch between real-world safety requirements and nondeterministic autonomy pipelines is now the primary barrier to safe, scalable deployment. I introduce Lume-Auto, a deterministic governance substrate for autonomous vehicles, fleets, and mobility systems. Built on the Lume-OS kernel, Lume-Auto integrates deterministic perception arbitration, invariant-preserving motion envelopes, multi-vehicle convergence, timing-corrected decision ordering, sensor-noise coherence, and replay-identical behavior. Lume-Auto compiles natural-language intent into deterministic, invariant-preserving driving actions that operate reliably in complex, dynamic, real-world environments. Lume-Auto defines a universal substrate for autonomous cars, trucks, drones, delivery robots, and fleet-scale mobility systems. I formalize the Lume-Auto architecture, define its motion semantics, and present constructive proofs demonstrating invariant preservation, deterministic override correctness, multi-vehicle convergence, and replay-identical driving behavior. Results across 500,000 deterministic cycles show zero invariant violations, zero envelope violations, and full replay-identical execution.