[CI] Cross-platform — Part 3: Windows workflow

[hujc7]

Summary

Part 3 of the cross-platform CI series. Adds .github/workflows/windows-ci.yaml — CI pipeline for Windows GPU self-hosted runners. Same shape as arm-ci.yaml (Part 2, #5698) but native install path instead of Docker.

• Tier 1 (smoke + install): general-windows, install-windows (uv install + wheel build + reinstall), kit-launch-windows.
• Tier 2 (meaningful, marker-driven): path-io-windows, perception-windows, ...
• All jobs continue-on-error: true while runners stabilize.
• All pytest invocations use --timeout=N + --timeout-method=signal so hung tests fail fast (fixes the previous perception-windows pattern where Vulkan failures hung the job instead of erroring).
• Marker-driven discovery: pytest <root> -m windows_ci --continue-on-collection-errors. Adding a new Windows-safe test requires only tagging it windows_ci, no yaml edit.

Series

PRs prefixed with [CI] Cross-platform —. Current siblings:

• Part 1 (base) — [CI] Cross-platform — Part 1: pytest markers and shared scaffolding (base) #5695 — Markers and shared scaffolding
• Part 2 — [CI] Cross-platform — Part 2: ARM/Spark workflow #5698 — ARM/Spark workflow
• Part 3 (this PR) — Windows workflow

Depends on Part 1 (#5695). Independent of Part 2 (#5698).

Test plan

• ./isaaclab.sh -f (pre-commit) passes.
• windows-ci.yaml triggers on PR push; all jobs run on Windows runners.
• perception-windows fails fast on Vulkan/runtime errors instead of hanging.

[isaaclab-review-bot]

Review Summary

This PR adds experimental Windows CI jobs for perception smoke testing, building on the uv-based installation path. The approach is well-structured, with incremental jobs that progressively test more functionality.

✅ Strengths

1. Incremental approach: The workflow builds from simple torch/scipy tests → IsaacLab core install → full perception smoke
2. Cross-platform fix in AssetConverterBase: Moving from hardcoded /tmp/IsaacLab to tempfile.gettempdir() is the correct POSIX/Windows-agnostic solution
3. Proper marker recognition in AppLauncher: The windows_ci and arm_ci markers are now correctly filtered from sys.argv
4. Good use of continue-on-error: true: Appropriate for experimental jobs that should not gate merges

🔍 Findings

1. Path handling in GITHUB_PATH append (Low Priority)

echo "$HOME\\.local\\bin" | Out-File -FilePath $env:GITHUB_PATH -Append -Encoding utf8

The $HOME variable in PowerShell uses forward slashes internally, but .local\\bin uses backslashes. While PowerShell is generally tolerant, consider using Join-Path for consistency:

Join-Path $HOME ".local\\bin" | Out-File -FilePath $env:GITHUB_PATH -Append -Encoding utf8

2. Potential silent failure in perception smoke (Medium Priority)

The inline Python smoke script does not capture or report specific failure modes:

python perception_smoke.py

If the script fails during Kit initialization (e.g., GPU detection issues), the error may be buried in the output. Consider adding explicit error handling:

import sys
try:
    # ... existing code ...
    print("perception smoke: PASS")
except Exception as e:
    print(f"perception smoke: FAIL - {type(e).__name__}: {e}", file=sys.stderr)
    sys.exit(1)

This would make CI failure reasons more immediately visible.

3. Artifact retention note

7-day retention for artifacts is appropriate for experimental CI. Once this stabilizes, consider if longer retention is needed for debugging flaky Windows-specific failures.

📝 Documentation

The PR description and commit messages are excellent - each iteration is well-documented with clear rationale for changes. The .skip changelog file appropriately notes this is CI-only.

✅ Verdict

LGTM - The changes are well-structured for an experimental CI addition. The continue-on-error: true at both job and step level appropriately prevents this from blocking the PR workflow while still providing visibility into Windows perception test stability.

The AssetConverterBase fix for cross-platform temp directories is a clean, correct change that should be merged regardless of the CI experiment outcome.

---

Update (6811028): PR rebased and squashed to two clean commits:

1. 9242498 — Adds Windows + ARM pytest markers, AppLauncher argv filtering, and AssetConverterBase cross-platform temp dir fix
2. 6811028 — Full Windows CI workflow with the 5-job pipeline

Content is unchanged from prior review. All previous findings remain valid optional recommendations. Approval stands.