surface comm overlap rank errors

[sudhakarsingh27]

Summary

• Surface rank-local comm/GEMM overlap failures with @record and bounded subprocess stdout/stderr so the real per-rank exception is visible.
• Apply deterministic and backend overrides through a copied child environment so one parametrized subprocess cannot affect later tests.
• Keep FA4 out of the L1 PyTorch distributed suite because it validates distributed behavior; dedicated attention-backend suites own FA4 coverage.

[greptile-apps]

Greptile Summary

The PR improves failure reporting for distributed communication-overlap tests and isolates subprocess-specific environment overrides.

• Records rank-level exceptions in the distributed layer runner.
• Centralizes subprocess result validation with bounded stdout and stderr diagnostics.
• Prevents test-specific environment settings from leaking between parametrized cases.
• Disables FlashAttention v4 at the distributed QA suite boundary.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains within the eligible follow-up-review scope.

No blocking failure remains.

Important Files Changed

Filename	Overview
qa/L1_pytorch_distributed_unittest/test.sh	Moves the FlashAttention v4 disablement to a suite-wide export and removes redundant per-command assignments.
tests/pytorch/distributed/run_layer_with_overlap.py	Adds Torch Elastic exception recording so distributed rank failures provide structured diagnostics.
tests/pytorch/distributed/test_comm_gemm_overlap.py	Adds centralized subprocess validation, text-mode output capture, bounded diagnostics, and isolated child-process environments.

_{Reviews (2): Last reviewed commit: "Keep FA4 out of L1 distributed tests" | Re-trigger Greptile}

[sudhakarsingh27]

/te-ci pytorch L1