[FenjuFu]

The RTX 3090 case installs vLLM 0.22.0, which is affected by published security advisories. Upgrade new runs to vLLM 0.31.0, including the fixes for GHSA-h3rc-6mm3-gc2m and GHSA-823j-m4cj-hmjf. Earlier intermediate versions remain affected by these recently published advisories.

Resolve Transformers 5.10.4 and the existing CUDA 13.0 compiler pins together with vLLM, remove the conflicting second Transformers installation, and update startup version checks. Record actual installed versions and use that metadata in generated reports and previews. Both renderers reject runtime/environment version mismatches and smoke, evaluation, or latency results that predate the current run, so interrupted upgrades cannot relabel historical results. The committed 0.22.0 benchmark results and evidence retain their original versions; the README and runbook explain that new results require another GPU evaluation.

Validation:

• Linux x86_64 / Python 3.11 dependency resolution succeeds with uv pip compile ... --python-platform x86_64-manylinux_2_35 --only-binary :all: (vLLM 0.31.0, Transformers 5.10.4, Torch 2.13.0, CUDA compiler 13.0.88).
• Verified the Transformers 5.10.4 wheel contains the gemma4 model configuration.
• All 15 offline case-tool tests pass, including historical/new-run compatibility and negative checks for mixed runtime metadata, stale environment/evaluation/latency, and missing timestamps. This suite now also runs in CI.
• python scripts/validate_cases.py --self-test, python scripts/validate_cases.py, Python compilation, Bash syntax checking, and git diff --check pass.

GPU serving and the 4,057-sample evaluation were not rerun: this workstation has no Linux RTX 3090 test environment. The runbook retains the smoke, correctness, and latency gates for GPU validation. The 0.31.0 release still provides CUDA 13.0 PyPI wheels.