Background

Megatron-LM PR #6054 detects flaky failures live and cancels its NeMo Run Docker job. Merge-queue validation showed cancellation removed the container without a terminal status. Fork PRs also could not pass the required release rehearsal because GitHub withholds BOT_KEY.

What changed

• Persist failed terminal status before deleting a cancelled Docker container.
• Cover cancellation with and without a stored job request.
• Skip release rehearsal and its summary only for fork-origin PRs.

Details

Trusted branches and same-repository PRs retain release validation. Fork PRs cannot access the required GitHub App private key, so their required summary no longer fails on unavailable secrets.

Companion downstream: NVIDIA/Megatron-LM#6054.

Tested

• Workflow YAML parsed successfully
• Ruff format clean
• Both focused Docker cancellation tests passed
• Prior full product CI and Codecov green