chore: bump vLLM nightly to 68ee830; 0.23.1-rc.0:8 → 0.23.1-rc.0:9

[helix-nine]

Summary

Advances the bundled vLLM nightly build from commit 3f5a1e1 to 68ee830 (both nvidia and rocm prebuilt images plus the cpu source-build submodule). Still on the upstream 0.23.1rc0 line, so this is a StartOS revision bump: 0.23.1-rc.0:8 → 0.23.1-rc.0:9.

Synced all four version anchors per UPDATING.md:

• NIGHTLY_SHA → 68ee8300a047db78fb52bac477daaaac7be11216 (drives both vllm/vllm-openai and vllm/vllm-openai-rocm nightly tags — both confirmed present on Docker Hub).
• vllm/ submodule checked out at the matching commit.
• UPSTREAM_VLLM_VERSION unchanged (0.23.1rc0) — the nearest reachable upstream tag for the new commit is still v0.23.1rc0 (git describe → v0.23.1rc0-515-g68ee8300a).
• startos/versions/current.ts version + release notes (all locales).

No SDK bump — @start9labs/start-sdk is already at the latest npm release (1.5.3). npm update refreshed a transitive dep in the lockfile.

Confirmed the AS base-common anchor still present in the new vllm/docker/Dockerfile.cpu, so scripts/patch-dockerfiles.sh injection remains valid.

Upstream highlights (3f5a1e1…68ee830, ~91 commits)

• New GLM-5/DSV3.2 model support; DiffusionGemma tensor parallelism; Voxtral migrated to the mistral-common 1.11.5 audio API.
• Rust frontend: TLS backend switched rustls → native-tls/OpenSSL; gemma4 migrated to the unified parser; tool-call event ordering preserved; rejects min_tokens above max_tokens.
• Render endpoints gain token offsets; new VLLM_GPU_SYNC_CHECK debug env var.
• Bugfixes: relative allowed-local-media paths, NVFP4+MTP crash, macOS/Apple Silicon CPU build hang (OpenMP now enabled).
• Removed legacy Grok and Aquila model architectures.

Full changes: vllm-project/vllm@3f5a1e1...68ee830

Test plan

• npm run check (tsc --noEmit) — green.
• make / s9pk pack verification in review.