Engine and solver starting points for a new scene

Describe contact scale, mass/load, sticking/sliding, contact-switch frequency, geometric complexity, compliance and acceptable cost. A recommendation identifies core / solver / runtime path / version / effective configuration. Matching solver names or asset files does not establish equivalence.

Scene features and difficulty	Evidenced starting candidate	First checks in the new scene
Small-cube normal compliance and loading/unloading	Normal response: native MuJoCo 3.11 and SuperDex 1.0 have separate frozen response profiles; historical PhysX retains only its original scope.	Effective stiffness, mass/inertia, contact count and separating loads; synthetic targets are not material properties.
Specified transients and computation cost	Transients and cost: low-impedance MuJoCo 3.14 is a candidate for this synthetic target, with full settings/timing in the report.	Oscillation, no tension, load epochs and native/observation cost; transfer failures require new mass/size checks.
Incline support and low-friction sliding	Six-engine matrix: select valid observations near the target conditions, such as the native PhysX 5.9 PGS development baseline.	Static-friction threshold, frictionless controls, momentum, acceleration and rotation; initial pass rates are not equal-budget tuned rankings.
Finite-pad pinch with load/drive limits	Genesis force-cap cases and MuJoCo cube-pinch records.	Per-side drive/normal force, torque, slip, low-force negatives and release; inertia mutation separately needs #145 admission.
Robot grasping with complex collision geometry	MuJoCo/SuperDex 14 s SDF scene, including its full assets, versions and control.	Collision preparation, native contact readback and complete lift/hold/release; one historical scene does not establish robustness for new assets.
Asset migration across frameworks or CPU/GPU paths	Matched-core comparison: use default/aligned attribution protocols first.	Source assets, intermediate models, effective parameters, clock, contact path/capacity and reset; unmatched versions cannot isolate a framework-only cause.
Cloth compliance/self-contact, pushing or in-hand rotation	Full scope and future tasks; existing cloth failures also guide diagnosis.	Task-specific positives, negatives and physical checks; not run is not unsupported, and folded drop is not active folding.

Retain candidate settings, parameter sensitivity, physical error, stability, cost, tested transfer range and failures for each selection. Prefer offline analysis of existing logs; add only the smallest experiment required by a concrete evidence gap. Without a matching scene, label the result a starting candidate and freeze development and unseen-validation conditions separately.

Use latest stable native versions for new admission and official compatible combinations for frameworks. Older versions still support historical mechanism knowledge; report matched-core attribution separately from versions used for selection. Resource limits, installation and visual replay cannot replace physical acceptance.

LIBERO starting point: preserve the native version/configuration and audit observation timing first. No new physics configuration qualifies in this cohort; 1 ms is not a general fix.