Jev Self-Driving Sim

# Jev Self-Driving Sim

A 2D, top-down autonomous car that runs in the browser and uses
[TypeSafe's Jev](https://docs.typesafe.ai/introduction) (a "System One" decision model)
as its driving classifier. Every ~200 ms the car turns what its sensors see into a JSON
state, sends it to Jev with four typed questions, and executes the answers:

| question           | type   | what Jev decides                                  |
|--------------------|--------|---------------------------------------------------|
| `lane_action`      | choice | `keep_lane` / `change_left` / `change_right`      |
| `speed_action`     | choice | `stop` / `slow_down` / `hold` / `speed_up`        |
| `hazard`           | score  | 0 (clear) … 3 (collision likely within seconds)   |
| `pedestrian_yield` | noul   | probability that ego must stop for a pedestrian   |

All four go in a single API call (speculative fan-out). Code then applies
confidence-gated routing: a low-confidence lane change is ignored, a strong
`pedestrian_yield` overrides speed, a severe `hazard` forces at least `slow_down`,
and a `stop` is softened to `slow_down` when nothing is within 1.5× the stopping
distance (so far-away pedestrians or obstacles cause a gentle slowdown, not a halt).
Speed steps are scaled by the time since the previous answer, so a fast Jev does not
brake harder than a slow one.
A small "reflex" in code (emergency brake, blind-spot abort) exists only for
imminent impacts and can be switched off in the UI.

## Jev vs LLM, side by side

With an OpenRouter key the page offers a **Compare** mode: two tracks on the same
course, the left car driven by Jev and the right one by an LLM (DeepSeek V4.1 Flash by
default) that receives the very same state and questions and must answer in the same
JSON shape. The course (`pu