[thxCode]

ROCm 10 Pack run 41952538686 was blocked by the newly added end-of-build inspection. Remove that ROCm 10-specific inspection from image production and analyze the published image separately on a Linux amd64 Docker host.

• Remove the recipe's final inspection invocation and delete pack/rocm/check_vllm.py.
• Delete the package/SDK/native-loader tests that only exercised that script.
• Keep the compiler-target tests, runtime matrix coverage, and Ray/CuPy dependency regression.
• Replace the script-specific documentation with digest-based post-publication inspection guidance.

The existing LMCache ABI check and Pack's digest-bound dependency collection remain in place.

Validation: full pytest with the pinned auto-sync tools passed on macOS (1334 passed, 17 skipped). The retained compiler-target, Ray/CuPy resolver, and recipe/matrix selection tests passed (72 focused tests). All-file hooks, Python distribution build, and strict MkDocs passed.

After merge, start a new Pack invocation from the updated source. Published-image analysis and GPU validation remain pending.

Relates #413 and #414.

[gemini-code-assist]

Code Review

This pull request shifts the ROCm vLLM image verification gate from a build-time step inside the Dockerfile to an optional post-publish inspection. It updates the documentation with instructions on how to run the check_vllm.py script against a published image, adjusts the script to import onnxruntime before loading its ROCm provider to ensure proper initialization, and introduces a new test suite (test_rocm_native_loading.py) to verify native library loading and symbol resolution. Obsolete build-gate tests have been removed. I have no feedback to provide as there are no review comments.