Build a robotics pack #196

Context

Cyber and trading are both "agent talks to an external surface"
domains. Robotics is structurally different — the agent talks to a
simulator that has internal state, the action space is continuous,
"task success" is geometric / temporal rather than evidence-based.

OpenRange's RuntimeBacking protocol allows non-HTTP runtimes — but
nobody's actually built one. A robotics pack would be the
forcing function.

What this issue tracks

Sketch a robotics.tabletop pack that connects OpenRange's lifecycle
to a robotics simulator (MuJoCo, Isaac, similar):

• A simple tabletop scenario (block-stacking, pick-and-place, etc.)
• Runtime backing that wraps the sim — probably gRPC or shared memory
  rather than HTTP
• Verifier that grades task success geometrically (cube on top of
  cube within tolerance)
• Episode reset = sim reset; episode terminate = task success or
  step budget exhausted

Why it matters

Tests OpenRange's domain-agnostic claims at the extreme. If the
abstractions don't hold for robotics, that's structural drift worth
knowing about before we accumulate more cyber-shaped assumptions.

Where to start

• docs/start_here.md — read first.
• src/openrange/core/runtime_backing.py — the RuntimeBacking
  protocol. Robotics needs a non-HTTP backing; this is where the
  contract lives.
• src/openrange/packs/cyber_webapp_offense_v1/ — the only
  reference pack today. Robotics will diverge from it heavily.

Acceptance

This is exploratory. If you want to drive it:

1. Open a design doc PR proposing the ontology, entrypoint kind, and
   runtime backing shape.
2. Discussion on whether the existing abstractions work or need to
   flex.
3. Implementation only after the contract is agreed.

Notes

Pick a sim with reasonable Python bindings; the cost of a brittle
sim integration overwhelms the value of the pack itself. MuJoCo
Python bindings + a simple table scene is enough.