feat: add EAGLE-3 speculative decoding support for Ornith-9B

[oniwakaa]

Description

EAGLE-3 speculative decoding merged into llama.cpp in June 2026 (PR #18039). It achieves 2-3× speedup using a tiny draft head (1 transformer layer) trained on the target model's hidden states.

LORE previously skipped speculative decoding because Falcon-H1 and Ornith have incompatible vocabs. EAGLE-3 doesn't use a standalone draft model — it shares the target's tokenizer.

Task

1. Check if AngelSlim/Qwen3-8B_eagle3 (or similar Qwen3 EAGLE-3 checkpoint) works on Ornith's Qwen3.5 architecture
2. Convert the checkpoint to GGUF using convert_hf_to_gguf.py --target-model-dir
3. Test with: llama-server -m ornith.gguf -md eagle3.gguf --spec-type draft-eagle3 --spec-draft-n-max 8
4. Measure: speedup vs baseline, acceptance rate, any quality regression

Expected Impact

If it works: 2-3× faster primary model inference. This would be transformative for LORE's orchestration latency.

References

• llama.cpp EAGLE-3 docs: https://github.com/ggml-org/llama.cpp/blob/master/docs/speculative.md
• Supported EAGLE-3 models: [Speculative decoding] feat: add EAGLE3 speculative decoding support ggml-org/llama.cpp#18039
• Ornith-1.0-9B: https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B