[DEMO] WrenchComposer zero-range regression (perf-gate demo)

[Neil4561]

Demonstration PR — not for merge

This PR exists to demonstrate the Performance Regression Gate catching a
real, documented performance regression end-to-end.

What it does

It deliberately reproduces the per-step regression that upstream
isaac-sim/IsaacLab#5265
(the WrenchComposer dual-buffer port) caused on locomotion tasks. When an
apply_external_force_torque event is configured with all-zero
force_range/torque_range (the default for velocity tasks such as H1, G1,
Anymal-C), the zero wrenches were still routed through the dual-buffer composer
every step, paying the full compose-and-apply cost in write_data_to_sim for a
semantic no-op (~47% throughput hit). The upstream fix added an early return
when both ranges are exactly zero; this PR removes that guard to recreate the
regression.

Expected gate verdict

• Isaac-Velocity-Flat-G1-v0 → RED (regressed; exercises the zero-range event)
• Isaac-Cartpole-Direct → GREEN (unaffected control)

The gate builds the image from this PR's own source (production-faithful path),
selects baselines by git merge-base ancestry, and posts a sticky verdict table.

The matrix is scoped to these two tasks for a fast, clean demo run.

[github-actions]

Performance Regression Gate

Overall result

🚫 One or more blocking-level performance regressions were detected

✅ 2 passed · ⚠️ 1 warning · 🚫 1 blocking signal · ❌ 0 benchmark failures

Advisory: results are reported for review but do not fail the PR.

Why this result is meaningful

The blocked G1/Newton benchmark was 15.30% slower than its baseline. Its compatible historical runs normally varied by about 1.70% (MAD ÷ median), so this slowdown was roughly 9× larger than normal run-to-run variation.

The gate's four-MAD blocking band was approximately 6.79%. The observed slowdown was more than twice that band, making it a clear regression rather than ordinary benchmark noise.

Run context

• GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
• Runtime: cuda=12.8, driver=595.71.05, warp=1.13.0
• Baseline snapshot: 9c115d3e3f5b

How to read this

Start with BLOCK and HARD FAILURE, then review any WARN rows.

• ✅ PASS: no meaningful slowdown was detected.
• ⚠️ WARN: the result is uncertain or performance is in the warning band.
• 🚫 BLOCK: performance crossed a blocking threshold.
• ❌ HARD FAILURE: the benchmark did not produce usable FPS.
• FPS: current throughput; higher is better. Baseline: median of compatible historical runs. Change: + is faster and - is slower.

Task	Backend	Result	FPS	Baseline	Change	Samples	What it means
Isaac-Cartpole-Direct	newton	⚠️ WARN	983196.0	1010818.0	-2.73%	8	Possible slowdown; review recommended
Isaac-Cartpole-Direct	physx	✅ PASS	286334.7	274577.2	+4.28%	8	No meaningful slowdown
Isaac-Velocity-Flat-G1-v0	newton	🚫 BLOCK	28001.9	33058.7	-15.30%	6	About 9× larger than normal 1.70% variation
Isaac-Velocity-Flat-G1-v0	physx	✅ PASS	11686.9	10962.8	+6.61%	6	No meaningful slowdown

Technical details

Threshold values, failure phases, retries, hardware, runtime versions, and diagnostic notes:

Task	Backend	Verdict	FPS	Baseline	Samples	Regression%	Floor	Threshold	Phase	Retry	GPU	Runtime	Note
Isaac-Cartpole-Direct	newton	WARN	983196.0	1010818.0	8	-2.73	N/A	rolling_window		no	NVIDIA RTX PRO 6000 Blackwell Server Edition	cuda=12.8, driver=595.71.05, warp=1.13.0	p99/med=1.193; outliers=0
Isaac-Cartpole-Direct	physx	PASS	286334.7	274577.2	8	4.28	N/A	rolling_window		no	NVIDIA RTX PRO 6000 Blackwell Server Edition	cuda=12.8, driver=595.71.05, warp=1.13.0	p99/med=1.242; outliers=0
Isaac-Velocity-Flat-G1-v0	newton	BLOCK	28001.9	33058.7	6	-15.30	N/A	rolling_window		no	NVIDIA RTX PRO 6000 Blackwell Server Edition	cuda=12.8, driver=595.71.05, warp=1.13.0	baseline MAD=561.0 FPS (1.70%); change=9.0× MAD; four-MAD band=6.79%; p99/med=1.132; outliers=0
Isaac-Velocity-Flat-G1-v0	physx	PASS	11686.9	10962.8	6	6.61	N/A	rolling_window		no	NVIDIA RTX PRO 6000 Blackwell Server Edition	cuda=12.8, driver=595.71.05, warp=1.13.0	p99/med=1.188; outliers=0

Updated for 38197210 · workflow run