Say what was substituted, and drive the joints the assets specify

[ZheningHuang]

Stacked on #45. Found running that branch end to end on Office-Elliott from a raw scan — ingest through to a MuJoCo scene, nothing reused.

#45 holds up. The seed export gave 99/99 colliders from compiled sidecars — no fallbacks, no rejections — and both seed and authored scenes pass the stability harness: 0 contacts and 0 mm penetration at t=0, no BADQACC/CONTACTFULL, nothing escaped over 30 s × 8 randomised resets. The articulation is real and authored: a door leaf on a 0–90° hinge, a sit-stand desk on a 0–0.28 m prismatic lift, three window sashes. The sim/ collider-overwrite data-loss bug is genuinely fixed — the two chair clusters' collider files are name-prefixed and hash differently.

What this PR fixes is what happened around that.

The pipeline reported success it had not earned

I hit this three times in one run.

Reference generation. My first run produced error_fallback for every object — "OpenAI image generation needs $OPENAI_API_KEY" — and carried on building assets from raw evidence sheets. The only signal was objects: 2 (0 references ok), which reads as a count rather than an error.

The export under-reported what it invented. _object_physics returned None before recording anything when an object had no source_glb — which is every fixture the authoring agent writes into Room.py. So the authored room reported an empty no_sidecar while most of its physics had been derived at export, and the simulate stage's warning loop, keyed on exactly that field, never fired. Re-exporting the authored Office-Elliott room:

colliders: 229   sidecar_colliders: 99   derived_colliders: 130   sidecar_coverage: 0.432
authored_no_package: 18  [AirCon0, Detector0, Luminaire0, ...]

The stage now says, whether or not anything failed:

130 of 229 colliders (57%) were derived at export time, not read from a compiled sidecar — mass, friction and pivots for those came from a category table

A truncated authoring run looked finished. The step budget lands gracefully and exits 0 through the same path as a room the model considered done. It now reports the truncation, and --author-steps N exposes the budget.

Config that was documented but not wired

settings.as_environment() forwarded OPENAI_API_KEY and never GEMINI_API_KEY/LR_IMAGE_PROVIDER, so the README's Gemini path could not work from .env — I only got the run going by exporting both into the shell. sanity.py separately hard-failed on a missing OpenAI key whatever the provider, and would have passed an OpenAI key sitting alongside LR_IMAGE_PROVIDER=gemini before failing at the first object.

The compiled physics was not reaching the simulator

No actuators on any object joint. nu was 3, all room shake drives. The sidecars carry effort and velocity; grep effort in the exporter hit only a dataclass default. Now every sidecar joint gets a motor at its stated effort:

act_Door0_Door_Leaf               ctrlrange=[-45,45]     vel=2
act_Table1_Desk_Lift              ctrlrange=[-800,800]   vel=0.04
act_Window0_Sash_Casement_Left    ctrlrange=[-40,40]     vel=2

A motor, not a position servo, deliberately — a servo holds a setpoint, so it would clamp every door shut at ctrl=0 and a hung door would stop swinging under the shake. Passive drift is unchanged (1.3 mm vs 1.0 mm, and that difference is the heavier chairs); driven at full effort every joint reaches 100% of its range. Joints recovered from the raw extras get no actuator, because nothing stated their effort.

drive_room_yaw was never emitted. mujoco_shake.py:221 asks for it by name; the exporter only ever made x/y/z. The twist the joint exists for — the thing the code documents as making a far corner move further than the middle — silently never happened, and the 50 t room was left free to rotate under contact torque. Under a shake that previously did nothing, the door now swings 16.1° and the sashes 3–11°, with initial_overlaps: 0.

Chairs were half their weight. OCCUPANCY_DENSITY["chair"] = 14.0 put a task chair at 7.2 kg, and after the room's fit rescaled it, 4.9 kg on the floor against a real 12–15. This is the direct cause of a scene where things fly on contact. Measured over their own bounding box, real chairs cluster at 23–27 kg/m³ regardless of type:

task chair, castors + gas strut   0.49 m³   13 kg -> 26.8
cantilever meeting chair          0.23 m³  5.5 kg -> 23.5
wooden dining chair               0.21 m³  5.0 kg -> 23.4
bar stool                         0.12 m³  5.0 kg -> 41.7

So chair: 25.0, and stool: 40.0 because a stool is small enough that its frame dominates its enclosed volume. A 20 N shove now moves a chair 1 mm rather than 27–52 mm, and 50 N still topples it — correct at 0.66 g against a ~0.55 g tipping threshold.

Also

--polish no longer includes the model-driven QC pass. It was the longest agent pass on the run for the smallest marginal change and nothing downstream reads its output; --quality-pass still asks for it.

scripts/simready/view_scene.py opens an exported scene interactively. MuJoCo's managed viewer fails constructing its window on macOS 26 — it dies the same way on a three-line test model — while the passive viewer is fine, so this drives that and owns the loop. Frozen by default, so a scene can be inspected without settling out from under you.

Caveats

One scan, and a small tidy one — it began with zero interpenetration, which #45's own table shows is not typical. The reporting and chair-density findings generalise; the clean stability result may not.

613 tests pass, ruff clean — and none of them would have caught any of this. grep "drive_room\|actuator" tests/ is empty: the suite tests the sidecar format thoroughly and never asserts on the emitted MJCF's dynamics. That gap is worth closing next, and I have deliberately not done it here.