[richardoo-707]

Fixes #6509. Related to the Layer 2 infrastructure failures on #6506.

GPU containers could lose device permissions after a cgroup/resource update while /dev/nvidia* remained visible. Resolve allocated UUIDs to actual device minors from NVIDIA's XML inventory and explicitly register those GPU nodes and the CUDA control nodes with Docker. Every visible GPU must then complete an allocation, computation and synchronization probe before tests.

Move disposable source, HOME and compiler caches to a private /tmp tmpfs to avoid the observed overlay2 removal failure. Record the container exit code and cleanup error separately; cleanup failures still fail the job and preserve the original execution exception.

Numerical acceptance requires real Git provenance. Fetch the pinned commit/tree metadata, reconstruct file objects from the compressed archive, and verify the complete Git tree with lazy downloads disabled. The standalone .git survives the container's private source copy and supports offline commit and dirty-tree checks.

Validation on head 9382593b42c5ab0e70ccb8e1147a45b554ced74a, tested merge 05978230a94352c85ad45e67790851188c033687:

Check	Result
Linux CI helper tests	78 passed, 0 skipped
Layer 1	Formatting passed; CPU 151 passed / 26 existing hardware/version skips
Fork-safe gate	Lint/policy/helpers passed; CPU 155 passed / 26 existing hardware/version skips
Layer 2, batch 2	23 passed / 2 existing HF5 reference skips, 1 GPU
Same run, batch 4a	22 passed, 0 skipped, 1 GPU
Same run, batch 5	69 passed, 0 skipped, 2 GPUs

All three GPU containers exited 0 and were removed successfully. Runtime: H20-5, Python 3.10.12 / Torch 2.13.0+cu130, existing approved image sha256:ae8ff34bf340a692803edca6e0f38a576fc1a2b206c01b4ef6f7c2a40dfd8c10. Regular Layer 2 uses FAST_MODE=1 / FAST_TEST=1; this final run needed no test reruns. The earlier Draft smoke also passed 22 batch-4a tests with fast mode disabled.

A controlled two-H20 probe on runtime-fix commit d31489ff134e99c9dcd267e1508b426f7d686d1e also computed successfully before and after updating only its own container's CPU quota. Source checkout regressions cover offline copied metadata, gitlinks, a missing SHA and a mismatched archive. Repository pre-commit and runner-policy checks passed.

The dependency image, GPU pool, test selection, numerical tolerances and version.txt are unchanged. tmpfs shares the existing container memory limit, capped at COLOSSAL_CI_MEMORY. Existing Docker records already stuck in removal still require administrator maintenance. Four/eight-GPU Layer 3 is outside this validation.