#7 llama: score Liquid d1-2B and d1-omni-2B

Summary

• Merges ggml-org/llama.cpp master into ra/main, which brings LFM3 and LFM3-Omni.
• Adds d1-2B and d1-omni-2B on that LFM3 graph, including /v1/systemone.

Test plan

• ./build/bin/test-lfm3-d1 on d1-2B-Q4_K_M.gguf plus the Q8 mmproj. noul P(true) was 0.991 for "The capital of France is Paris."
• llama-cli on LFM3-Omni-2B-Q4_K_M.gguf with --reasoning off answers text prompts.
• llama-server /v1/systemone with a picture on the mmproj.

Made with Cursor

---

Summary by cubic

Brings in LFM3 and LFM3-Omni support by merging llama.cpp master, then adds d1-2B and d1-omni-2B on that graph, including the /v1/systemone decision endpoint.

Testing

• Verified d1-2B text reasoning (P(true) 0.991) and LFM3-Omni-2B text prompts with --reasoning off.
• The /v1/systemone image path on the mmproj is not yet tested.