[Diag][DO NOT MERGE] Negative arm — revert #5633 torch-defer to validate #5656 numpy!=2.3.5 fix

[hujc7]

1. Purpose

Negative arm of the validation for PR #5656 (the numpy!=2.3.5 exclusion).

Develop currently carries two independent mitigations for the OpenBLAS atfork SIGSEGV:

1. PR Fixes OmniHub startup in Docker tests #5633 (Antoine Richard, merge a5eb9add4c3) which carried PR Prevents early numpy imports to avoid Kit crash #5620's content — defers import torch in AppLauncher until after SimulationApp starts, so numpy/OpenBLAS isn't loaded in pytest's master process before the Kit fork.
2. PR [Fix] Exclude numpy 2.3.5 from every IsaacLab install path #5656 — excludes numpy 2.3.5 from every install path, so even an early numpy import resolves to a non-buggy version.

If either mitigation alone is sufficient, the other is redundant. This PR reverts (1) so we can see whether (2) holds on its own.

2. Setup

• Branch base: develop (latest).
• Commit 1: surgical revert of source/isaaclab/isaaclab/app/app_launcher.py to its state just before a5eb9add4c3 (Fixes OmniHub startup in Docker tests #5633). Other files from Fixes OmniHub startup in Docker tests #5633 left alone — the defer-torch logic in AppLauncher.__init__ is the only piece relevant to the atfork bug.
• Commit 2: source/isaaclab/test/conftest.py that imports numpy and prints version + bundled OpenBLAS hash at pytest session start.

The conftest is deliberately placed at source/isaaclab/test/ rather than source/ to avoid the namespace-package shadowing that broke PR #5655 — anchoring pytest's rootdir at source/ prepends it to sys.path, making every source/<pkg>/ (no __init__.py) shadow the real pip-installed package and turn from isaaclab_teleop import IsaacTeleopCfg into an ImportError before any test runs.

3. Expected result

• Install resolves to numpy 2.3.4 (highest 2.3.x that satisfies !=2.3.5 under cmeel-boost's cap from [Fix] Exclude numpy 2.3.5 from every IsaacLab install path #5656).
• [dep-manifest] log line shows numpy 2.3.4 + libscipy_openblas64_-8fb3d286.so.
• import torch happens in AppLauncher.__init__ BEFORE SimulationApp fork, but resolves to non-buggy numpy 2.3.4, so canary jobs (isaaclab_physx, isaaclab_newton, isaaclab_rl) pass.

If any canary still SIGSEGVs in libomni.platforminfo fork, the numpy!=2.3.5 exclusion alone is not sufficient and #5633's torch-defer is also necessary — keep both.

If all canaries pass, the numpy!=2.3.5 exclusion alone is sufficient and #5633's torch-defer is redundant — could be reverted in a follow-up cleanup.

4. Lifecycle

Diagnostic-only. Do not merge. Closes once evidence is captured.

5. Related

• PR [Fix] Exclude numpy 2.3.5 from every IsaacLab install path #5656 — the numpy!=2.3.5 exclusion being validated
• PR Fixes OmniHub startup in Docker tests #5633 — Antoine's merge of Piotr's Prevents early numpy imports to avoid Kit crash #5620 torch-defer mitigation (the one we're temporarily reverting)
• PR Prevents early numpy imports to avoid Kit crash #5620 — Piotr's original (closed) version of the same mitigation
• PR [Diag][DO NOT MERGE] Negative arm — numpy 2.3.5 atfork crash (no 5656) #5655 — earlier diagnostic attempt; superseded by this PR (the source/conftest.py placement there broke test discovery before the bug could surface)
• BUG: Binary Builds Deadlock due to OpenBLAS threading issue with fork numpy/numpy#30092
• Deadlock after fork when calling dgetrf_ OpenMathLib/OpenBLAS#5520

[isaaclab-review-bot]

🔬 Diagnostic PR Review

Summary

This diagnostic PR reverts the torch-defer mitigation from PR #5633 to validate whether the numpy!=2.3.5 exclusion (PR #5656) alone is sufficient to prevent the OpenBLAS pthread_atfork SIGSEGV.

Changes Analysis

1. source/isaaclab/isaaclab/app/app_launcher.py (-43/+6 lines)

The revert correctly undoes the deferred torch import pattern:

• Removes _deferred_cuda_device_id tracking
• Removes _set_deferred_cuda_device() method
• Removes apply_rtx_determinism_settings() static method
• Removes --deterministic CLI argument and related handling
• Moves torch.cuda.set_device() back into _resolve_device_settings() (pre-SimulationApp fork)

This restores the original behavior where import torch (and transitively numpy/OpenBLAS) occurs before Kit's fork() call, which would trigger the SIGSEGV if numpy 2.3.5's buggy OpenBLAS atfork handler is registered.

2. source/isaaclab/test/conftest.py (+29 lines, new file)

Well-designed diagnostic fixture that:

• Prints numpy version at pytest session start
• Identifies the bundled OpenBLAS .so filename (contains commit hash)
• Correctly placed under source/isaaclab/test/ to avoid the namespace-package shadowing issue documented in PR #5655

The docstring clearly explains the rootdir anchoring rationale.

CI Status Notes

• Check changelog fragments failing is expected for diagnostic PRs
• Docker/canary jobs still pending — these will provide the actual validation signal

Validation Criteria

Per the PR description:

• ✅ Pass scenario: numpy resolves to 2.3.4, canaries pass → numpy!=2.3.5 exclusion is sufficient
• ❌ Fail scenario: SIGSEGV in libomni.platforminfo fork → both mitigations needed

Recommendations

1. No merge — correctly labeled as diagnostic-only
2. Monitor canary jobs (isaaclab_physx, isaaclab_newton, isaaclab_rl) for the actual SIGSEGV validation
3. Document outcome in a comment once CI completes for future reference

Clean diagnostic implementation. The conftest placement fix from #5655 learnings is a nice detail.