fix(showcase): seat G1 on terrain surface so the Curriculum-Showcase stops NaN-ing

[KraHsu]

Problem

genelab play GeneLab-Curriculum-Showcase-v0 --agent zero --steps 5 crashed with:

genesis.GenesisException: Invalid constraint forces causing 'nan'. Please decrease Rigid simulation timestep.

Reproduced on RTX 5060 Ti (local) and H200 (server) — not hardware-specific.

Root cause (dt was a red herring)

A no-step diagnostic showed the G1 spawned ~0.77 m buried in the RandomRough heightfield. terrain_levels_vel (which runs on every reset) and the runner's teleport both wrote the robot base to the terrain cell origin at z=0 (= terrain.pos[2]), ignoring the surface elevation and the robot's 0.76 m standing height. Feet landed at z=−0.72 under a +0.05 m surface. That interpenetration exploded the rigid contact solver:

• dt=0.005 → instant NaN at step 1 (the reported symptom).
• dt=0.002 → an unconverging multi-minute solver grind (_kernel_solve_graph).

The Terrain-Showcase shares the identical buried-spawn bug (out of scope here; see follow-ups).

Fix (one concept: seat the G1 on the terrain, not in it)

• terrains/importer.py: add surface_height_at(xy) — world-frame heightfield surface lookup.
• mdp/curriculums.py: terrain_levels_vel reseats the base at surface + init_pos.z + spawn_clearance instead of the bare cell origin (new optional spawn_clearance=0.1).
• showcase runner.py: teleport reseats at standing height too, so the full 400-step demo survives its teleports.
• showcase env_cfg.py: dt 0.005→0.002, decimation 4→10. The G1 is independently NaN-unstable at dt=0.005 even from a clean stance; this matches the known-good Velocity/Tracking-Flat configs. Control rate stays 50 Hz; the 2 s episode (100 control steps) is unchanged.
• rl/runner.py: play_task caps scripted (zero/random) playback at simulation.steps, so play TASK --agent zero --steps N stops after N steps instead of looping forever headless (--steps was silently ignored on the RL play path; max_steps defaulted to None). Trained playback stays unbounded.

Verification

Check	Result
play … --agent zero --steps 5 (exact repro)	exit 0, no NaN (~12 s)
play … --steps 5 (showcase task.play() path)	exit 0, no NaN
Full play … --steps 400 (all teleports)	exit 0, no NaN; curriculum promotion fires (levels 0→4, histogram=[8,0,0,0,8])
No-step spawn diagnostic	feet 0.14 m above surface (was 0.77 m buried)
uv run pytest -q	507 passed

Reference audit (CLAUDE.md)

grep across src/ examples/ tests/:

• terrain_levels_vel — only non-test caller is the curriculum showcase cfg; signature change is additive (new optional spawn_clearance). No breakage. Behavior tests in tests/test_curriculum_terrain.py updated to the corrected (seated) spawn.
• surface_height_at — new method; callers are terrain_levels_vel and the showcase runner (both in this PR).
• reset_root_state_on_terrain — a reset-event approach explored mid-development was fully removed (the curriculum overwrites root pose after events, so the fix had to live in terrain_levels_vel); grep confirms zero references remain.

No plans/ files touched.