perf(rbd): skip inactive joint-limit constraint rows

[haixuanTao]

What

emit_limit_constraint always emits a kind = 1 row and computes its O(ndofs) compute_constraint_column LU back-solve — even when the joint sits strictly inside its bounds. In that state the row has zero positional bias and both impulse clamps at 0 (max_neg_impulse = max_pos_impulse = 0), so it can never apply any impulse: it is dead weight in every PGS sweep, re-paid every substep.

This PR skips emission when !min_enabled && !max_enabled. The slot stays kind = 0 (pre-zeroed by the discovery walk), which the PGS sweeps already skip (if cons.kind == 0 { continue }). Constraints are rebuilt from current coordinates each substep, so a joint reaching its bound re-emits on the next substep exactly as before — result-identical by construction (the skipped rows could only ever contribute a zero impulse).

Why it matters

For articulated robots with limits on every joint, essentially all limit rows are inactive on every substep of normal operation (a standing/walking robot lives well inside its joint ranges). The joint-constraint work then scales as #joints × O(ndofs) back-solves per substep for nothing.

Measured on a downstream biped-RL training stack (RTX 5090, 2,048–8,192 parallel envs, limits on all joints, profiled with Nsight Systems):

• the joint-constraint kernel was the feat: update deps + release v0.2.0 #1 GPU kernel for a 25-joint humanoid (~28 % of total GPU time);
• with the skip: +16 % end-to-end training throughput on a full-body Unitree G1, +5 % on a 12-DOF biped;
• training statistics bit-identical to the unpatched build.

The win grows with joint count, so it compounds for exactly the robots where the multibody solver is most loaded.

🤖 Generated with Claude Code

https://claude.ai/code/session_016Dbpu5NhqXK2Xt3YzUbJRM