chore: bump llama.cpp to b10481; 1.0.10450:0 → 1.0.10481:0

[helix-nine]

Summary

Bumps the pinned llama.cpp server images from b10450 to b10481, and the package version from 1.0.10450:0 to 1.0.10481:0.

b10481 is the newest build GHCR has actually published server images for — upstream cuts a GitHub release per merged commit but publishes images for only a small subset, so the pin comes from the GHCR tag list, not the release list. All four variants were confirmed to resolve before pinning:

Variant	Tag	Status
generic	ghcr.io/ggml-org/llama.cpp:server-b10481	OK
nvidia	ghcr.io/ggml-org/llama.cpp:server-cuda-b10481	OK
rocm	ghcr.io/ggml-org/llama.cpp:server-rocm-b10481	OK
vulkan	ghcr.io/ggml-org/llama.cpp:server-vulkan-b10481	OK

Upstream range

31 builds, b10450...b10481. Highlights that reach a StartOS user:

• BailingMoE3 architecture (#26608) — runs the Ling 3.0 Flash models.
• AMD memory reporting (#27083) — HIP builds now read available VRAM via hipMemGetInfo instead of inheriting the DGX Spark unified-memory override, which over-promised memory on APUs with a small carveout. Relevant to the rocm variant.
• Multimodal: the server stores processed mtmd chunks as placeholders (#27278), cutting prompt memory; plus hardened Granite preprocessing and an LFM2 thumbnail fix.
• Robustness: LoRA tensor data is bounds-checked against the file (#27056).
• Built-in chat UI: built-in tools regrouped into Server/Browser, new browser get_info tool.

No llama-server flag changes in the range — the diff of common/arg.cpp touches no option literals, so the -hf/-hff/-c/-ngl arguments built by Set Model and the presets in startos/actions/presets.ts are unaffected. No migration needed.

No SDK or dependency changes

@start9labs/start-sdk is already pinned at 2.0.9, which is the latest published version, and this package has no *-startos git dependencies — so package.json and package-lock.json are untouched.

Release notes

Rewritten for this bump in all five locales (en_US, es_ES, de_DE, pl_PL, fr_FR). They describe only what b10481 adds; the Set Model prefill work announced in 1.0.10450:0 is not repeated.

Test plan

• Confirmed 1.0.10481:0 is not already published — the registry's best for llama-cpp is 1.0.10450:0.
• All four b10481 image variants resolve on GHCR (docker manifest inspect).
• npm run check (tsc) green.
• make generic — deferred to review, per the monitor cycle's typecheck-only gate.

---

Merge with a merge commit — do not squash. next is long-lived: a merge commit leaves it a true ancestor of master, so it fast-forwards cleanly afterwards. A squash re-lands the same content under a new commit, so the branch is left carrying history master will never contain.