[CI] Warm start CI before tests

[xyao-nv]

Summary

Short description of the change (max 50 chars)

Detailed description

• What was the reason for the change?
• What has been changed?
• What is the impact of this change?

[isaaclab-review-bot]

Nice idea — pre-warming the shader cache so every subsequent SimulationApp in the CI run starts fast. A few observations:

Overall

• The approach is sound: one throwaway SimulationApp launch compiles all RTX shaders into the on-disk cache, so the 4-5 test steps that follow skip the compilation tax. Clean and self-documenting.

scripts/warm_shader_cache.py

1. argparse.Namespace construction — Manually assembling a Namespace(headless=True, enable_cameras=True, visualizer=[]) is fragile; if AppLauncher ever adds a required arg or changes its defaults, this will silently break. Consider using AppLauncher.add_app_launcher_args() to build the parser, then parse known args, or at minimum document why the hand-rolled namespace is sufficient.
2. enable_cameras=True — Good call: this forces the rendering/RTX path so the shader compilation actually happens. Worth a short inline comment explaining why cameras are enabled (it's not obvious since this script doesn't render anything).
3. Error handling — If AppLauncher or app.close() throws (e.g., no GPU, missing Kit runtime), the CI step will still fail with a non-zero exit, which is fine. But a try/except with a descriptive message could save debugging time on new runner configurations.
4. Minor: The visualizer kwarg is passed as [] — is this needed? If AppLauncher defaults visualizer to None or [] already, this can be dropped for clarity.

.github/workflows/ci.yml

• The warm step is correctly placed before the first test step in both test and test_policy jobs. Since the two jobs run in separate containers (different images), each needs its own warm-up — good.
• The warm step doesn't set a timeout-minutes at the step level. Given shader compilation can take 60-180s (per the docstring), consider adding a step-level timeout (~5-10 min) as a safety net against hangs.

No blocking issues — this is a clean, well-scoped CI optimization. 👍

[isaaclab-review-bot]

Nit: Hand-rolling argparse.Namespace couples this to AppLauncher's internal expectations. If the launcher ever adds validation or required fields, this will break without a clear error.

Consider:

parser = argparse.ArgumentParser()
AppLauncher.add_app_launcher_args(parser)
args = parser.parse_args(["--headless", "--enable_cameras"])

This also self-documents the available flags.

[isaaclab-review-bot]

Worth a brief comment explaining why enable_cameras=True — the point is to trigger the RTX/rendering shader compilation path, not to actually use cameras. Future readers might otherwise wonder.

[isaaclab-review-bot]

Consider adding a step-level timeout-minutes: 10 here. The docstring says shader compilation can take 60-180s, but if Kit hangs (no GPU, license issues, etc.) you'd want the step to fail fast rather than burn the full job timeout.