A native-sim eval hangs silently on a headless box instead of naming `MUJOCO_GL`

[vertix]

On a headless Linux box, a native-sim eval hangs instead of running, and says nothing about why:

positronic eval run --eval=.sim.positronic.stack_cubes --eval.trial_count=3 \
  --policy=.remote --policy.url=localhost:8010 --output_dir=~/out

It sat for 7 minutes at ~2% CPU with no episode written and no error. The same command with MUJOCO_GL=egl
in front of it finished three episodes in 49 s. The box has no display (DISPLAY empty); MujocoSim constructs
its mj.Renderer with whatever backend mujoco picks by default, which without a display is GLFW.

The backend is pinned everywhere else it is needed, and only the native sim run through the CLI is left out:

• simulator/libero/launcher.py sets env.setdefault('MUJOCO_GL', 'egl') for the env server it spawns, with a
  comment on exactly this failure — GLFW "wants a display no headless host has";
• docker/docker-compose.yml sets MUJOCO_GL: egl on the positronic-inference service;
• the test suites set it themselves (test_inference_integration.py, env_server/tests/test_remote_env.py),
  and the libero-e2e workflow sets osmesa.

So the operator running a native sim eval outside Docker is the one case with nothing pinning it, and the
failure mode is a silent hang rather than a message naming the display.

Worth deciding which of these it should be:

• have the native sim path pin a headless default the way the LIBERO launcher does (setdefault, so an
  operator's own backend still wins), or
• leave the backend to the operator but fail fast with a message that names MUJOCO_GL when there is no
  display, instead of blocking in GL setup.

The hang is the part worth fixing either way — the fix is cheap and the symptom costs an operator a long time
to attribute, since nothing in the output points at rendering.

I did not capture a stack from the blocked process, so "blocks in GL context creation" is inferred from the
symptom disappearing under MUJOCO_GL=egl rather than observed directly.

Found while running a live-policy eval for #591; unrelated to that change.