[DEMO] WrenchComposer zero-range regression on perf-smoke (RED)

[Neil4561]

Demo PR (not for merge)

RED case for the performance smoke-test demo.

Change: reintroduces the WrenchComposer regression from upstream isaac-sim#5265.

Explanation: some environments configure an external force event with force and torque set to zero. That means "do not actually push the robot." The normal code sees that everything is zero and skips the work. This PR removes that shortcut, so the simulator still goes through the WrenchComposer path to write zero forces and zero torques into the wrench buffers. The physical result is the same, but the simulator does extra bookkeeping work, which shows up as lower FPS on affected tasks.

Isaac-Velocity-Flat-G1-v0 is the important affected task in the demo because it exercises that external-wrench path. Cartpole and Factory act as controls: they stay near baseline, showing that the regression is not a global slowdown across every benchmark cell.

Matrix: Isaac-Cartpole-Direct + Isaac-Factory-GearMesh-Direct-v0 (physx) + Isaac-Velocity-Flat-G1-v0.

Trigger via mirror to pull-request/<N> or workflow_dispatch.

[github-actions]

Performance Smoke Results

Baseline read SHA: e87de2efea41

Run context

• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
• Runtime: cuda=12.8, driver=595.71.05, warp=1.13.0
• Threshold policy: rolling baseline per row. Each task/backend is compared against its own compatible baseline samples using median/MAD.

How to read this table

• PASS: no meaningful slowdown was detected.
• WARN: suspicious or uncertain result, such as insufficient baselines, retry-only success, or a slowdown in the WARN band.
• BLOCK: a blocking-level regression signal was detected. In this POC, BLOCK is advisory (it flags a regression) and only fails the check when the gate is explicitly configured as blocking.
• HARD_FAILURE: the benchmark did not produce usable FPS data, for example an import/init/runtime failure or config mismatch.
• Delta: change in FPS versus the baseline. + means faster than baseline (speedup); - means slower than baseline (slowdown).
• Threshold policy: explains which comparison rule produced the verdict. rolling_window means each row uses its own compatible baseline median/MAD window, not one shared FPS cutoff.
• Noise: baseline MAD as a percent of baseline FPS. Higher noise means the task/backend naturally varies more from run to run.
• Phase: the stage a HARD_FAILURE happened in (e.g. import, init, runtime); blank when the benchmark ran to completion.
• Retried: yes if the cell only passed after an automatic re-run.

Task	Backend	Verdict	FPS	Baseline	Delta (+ faster / - slower)	Noise	Samples	Phase	Notes	Retried
Isaac-Cartpole-Direct	newton	PASS	1020799.6	1010818.0	+0.99%	0.80%	8		p99/med=1.174; outliers=0	no
Isaac-Cartpole-Direct	physx	PASS	288676.8	274577.2	+5.14%	1.98%	8		p99/med=1.2; outliers=0	no
Isaac-Factory-GearMesh-Direct-v0	physx	PASS	1209.0	1199.0	+0.84%	1.48%	6		p99/med=1.03; outliers=1	no
Isaac-Velocity-Flat-G1-v0	newton	BLOCK	29527.6	33058.7	-10.68%	1.70%	6		p99/med=1.113; outliers=0	no
Isaac-Velocity-Flat-G1-v0	physx	BLOCK	8790.6	10962.8	-19.81%	2.71%	6		p99/med=1.097; outliers=0	no

Updated for 33bd4b2 · run