vLLM 0.30.0 for the production embeddings role on WSL (pin, receipt, docs)

[seathatflowsinourveins]

Why

The WSL embeddings route ran vLLM 0.25.0. That version is in range of two advisories:

• GHSA-25q3-v2hm-8vpf (high): a negative token id in /v1/embeddings or /pooling kills the engine. Fixed in 0.28.0.
• GHSA-5fj9-pfhr-6j48 (medium): an unauthenticated metrics DoS. Fixed in 0.30.0.

0.29.0 had failed WSL startup with UVA is not available (evidence/receipts/vllm-compatibility.json). 0.30.0 carries the WSL pinned-memory fallback (vllm PR #56908).

What was done on the NativeStack host (RTX 4090, WSL2)

1. Qualified on an owned instance. 0.30.0 was installed in a new tools/vllm-0.30.0 prefix (uv pip install vllm==0.30.0 --exclude-newer 2026-09-25T23:59:59Z, wheel sha256 ef52ee58…). It served the same Nemotron-3-Embed-1B-BF16 files, hash-identical before and after, on 127.0.0.1:38231 with production's flags. Compared with production 0.25.0:

   • gate-set min cosine 0.999999999999998;
   • identical top-1 on NVIDIA's example;
   • 10/10 top-10 agreement on a 10×50 corpus;
   • batch invariance at least 0.99997;
   • p50 latency 23.8 ms vs 21.4 ms.

   A separate reviewer agent re-ran the checks and rated it switch-ready.

2. Switched in production. A systemd --user drop-in changes only the ExecStart binary path. The unit was healthy after 48 s with 0 restarts. On the live unit, before vs after:

   • gate embeddings: cosine 1.0 on all 8 texts;
   • 5/5 live code-index queries (67,492 points) returned identical top-10 order;
   • GPU memory went from 13,879 to 13,032 MiB;
   • llama-server and qdrant stayed healthy.

   0.25.0 stays installed. Rollback is to delete the drop-in, then daemon-reload and restart.

Repository changes

• manifests/stack.json: vllm is now 0.30.0 @ ced6857a (the v0.30.0 tag), with the new evidence id.
• Audit and snapshot rows: blueprints/token-native-focus/saturation-audit.json row updated. The catalogs/landscape/upstream-snapshot.json row comes from a fresh run of its four gh api endpoints, the same method as Bump the application-delivery lock to Next.js 16.3.6 for GHSA-vcvr-r3jv-pc5j #252. The selected version now matches the latest stable release.
• evidence/receipts/vllm-030-switch-20260925.json (native_model_e2e): the claim, limitations, numbers and artifact hashes. Raw machine output is retained privately.
• Docs: dated updates to adoption/platforms/linux-wsl2.md, adoption/lifecycle.md, docs/token-native-saturation.md and the recipes/README.md vllm row.

Checks

• Full test suite: 5307 tests OK, 632 skipped. The skips are host-dependency skips: numpy and alpaca-py are absent from this uv test environment.
• Validators: validate.py, landscape.py, validate_catalogs.py and validate_foundation.py pass. component_matrix.py --check and new_host_grand_list.py --check pass.

Limitations

• Scope: one host, one model, and the /v1/embeddings route. NVIDIA's model card still names vllm==0.25.0 for BF16 /v2/embed, a route that is not used here.
• Not a benchmark: retrieval equality covers 8 gate texts and 5 live queries, not an aggregate retrieval-quality benchmark.
• Host receipt pending: a 0.30.0 vllm use host receipt will be recorded from this commit on main after merge. The other session's Workstation lane wave A: systemd re-record, SPY/LEAN parity host replay, Nemotron qualification on the RTX 4090, ai-memory 2.4.0 review #256 receipt stays a point-in-time record of 0.25.0.

Lane: lane:foundation

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.